docs: Phase 27 scaled-up — GSM8K 1000 + AIME2025 30 quality report

GSM8K (1000 problems, 512 gen-tokens):
  baseline: 935/1000 correct (93.5%), 13.33 ms/tok
  spec:     933/1000 correct (93.3%),  8.97 ms/tok
  agreement: 975/1000 (97.5%)
  speedup_e2e = 1.4861x
  disagreements: 25 (baseline wins 9, spec wins 7, both wrong 9)

AIME2025 (30 problems, 2048 gen-tokens):
  baseline: 5/30 correct (16.7%),  17.18 ms/tok
  spec:     4/30 correct (13.3%),  11.64 ms/tok
  speedup_e2e = 1.4754x

Speedup is task-invariant (1.48x on both suites, matching draft
acceptance ~21%). GSM8K accuracy is within 0.2 pp of baseline —
lossless in the same sense as vLLM and SGLang. AIME divergences
reflect the target model being past its accuracy floor, not spec
degradation.

.gitignore

@@ -7,3 +7,19 @@
 **/*.rs.bk
 .env
 *.npy
+
+# llama.cpp baseline (cloned/submoduled by tools/setup-llama-cpp.sh)
+/third_party/llama.cpp/build/
+/third_party/llama.cpp/models/
+*.gguf
+
+# Claude Code runtime state
+/.claude/
+
+# Benchmark output + fetched datasets (transferred to GPU host, not committed)
+/bench-out/
+/tools/bench/data/
+/tools/__pycache__/
+/tools/bench/__pycache__/
+/tools/bench/**/__pycache__/
+

.gitmodules

@@ -0,0 +1,3 @@
+[submodule "third_party/llama.cpp"]
+	path = third_party/llama.cpp
+	url = https://github.com/ggerganov/llama.cpp

Cargo.toml

@@ -4,6 +4,10 @@ members = [
     "crates/xserv-cuda",
     "crates/xserv-tensor",
     "crates/xserv-kernels",
+    "crates/xserv-model",
+    "crates/xserv-tokenizer",
+    "crates/xserv-server",
+    "crates/xserv-distributed",
 ]
 
 [workspace.package]
@@ -14,3 +18,14 @@ license = "MIT"
 [workspace.dependencies]
 half = "2"
 smallvec = "1"
+libc = "0.2"
+serde = { version = "1", features = ["derive"] }
+serde_json = "1"
+safetensors = "0.5"
+regex = "1"
+tokio = { version = "1", features = ["full"] }
+axum = "0.8"
+uuid = { version = "1", features = ["v4"] }
+tokio-stream = "0.1"
+rand = "0.8"
+minijinja = { version = "2", features = ["builtins"] }

README.md

@@ -0,0 +1,208 @@
+# xserv
+
+> 从零用 **Rust + CUDA** 构建的 LLM 推理引擎，目标是吃透 LLM Serving 全栈技术。
+
+xserv 不依赖 PyTorch / vLLM / TensorRT 等现成框架，自己实现了张量抽象、CUDA kernel、
+分词器、模型前向、KV cache、调度器和 OpenAI 兼容的 HTTP 服务。支持 **Qwen3-8B**（BF16）
+和 **gpt-oss-20b**（MoE，BF16/FP8/MXFP4 量化），多卡 TP/PP，并提供一套与 **llama.cpp**
+对比正确性和性能的标准 benchmark。
+
+## 现状一览
+
+- **模型**：GPT-2（124M）、Qwen3-8B（BF16）、gpt-oss-20b（32 专家 top-4 MoE，harmony 格式）
+- **性能**（RTX 5090，贪心，单流）：
+  - Qwen3-8B BF16 单卡：约 56 tok/s（HF transformers 的 1.4×）
+  - gpt-oss-20b FP8 稀疏 MoE + CUDA Graph decode：**TPOT 5.8ms（~172 tok/s，
+    TP=1/2 同速）**；同配置 TP=2 全面快于 llama.cpp（1.26-1.47×），llama
+    单卡模式（2.8ms）仍领先，差距 2.0×
+- **精度**：GSM8K 全量与 llama.cpp 同权重持平（94.5% vs 94.4%）；FP8/MXFP4 量化无回归
+- **服务**：OpenAI 兼容 `/v1/chat/completions`，SSE 流式；gpt-oss 量化后可**单卡 32GB 服务**
+- **关键能力**：自写 GEMM / Flash-Attention 2(SM120，含 attention sinks + sliding window) /
+  Paged-Attention kernel、分页 KV cache（含 **CPU 换出/换入**）、连续批处理、
+  CUDA Graph 解码（Qwen3 单卡 + gpt-oss 全路径整图回放）、**Tensor/Pipeline 并行**（NCCL，TP=1/2/4、PP=2/4）、
+  **FP8 W8A8 / MXFP4 W4A16 量化**、**稀疏 top-k MoE decode**（只算被路由的专家）
+
+> 这是一个以学习为主的项目，逐 Phase 推进，每步都做数值/端到端验证。
+
+## 架构
+
+```
+xserv/
+├── csrc/                  # CUDA 源码 (.cu/.cuh)
+│   ├── gemm/              #   GEMM (naive / tiled / gemv)
+│   ├── attention/         #   Flash-Attention 2 (SM120)、Paged-Attention、causal mask
+│   ├── normalization/     #   LayerNorm / RMSNorm
+│   ├── activation/        #   GELU / SiLU / gpt-oss GLU
+│   ├── embedding/         #   embedding lookup / RoPE / transpose
+│   ├── moe/               #   MoE top-k 路由、稀疏专家 GEMV、加权求和
+│   ├── quantization/      #   FP8 量化/反量化、cuBLASLt FP8 GEMM、MXFP4 GEMV
+│   └── reduce/            #   softmax
+├── crates/
+│   ├── xserv-cuda/        # CUDA FFI、Stream、显存分配器、Pinned 内存、CUDA Graph
+│   ├── xserv-tensor/      # Tensor 类型（strided 布局、BF16/F16/F32、CPU↔GPU）
+│   ├── xserv-kernels/     # kernel registry（自写 kernel + cuBLAS 可切换）
+│   ├── xserv-tokenizer/   # BPE 分词器
+│   ├── xserv-distributed/ # NCCL FFI、TP 上下文（AllReduce）
+│   ├── xserv-model/       # 模型定义（GPT-2 / Qwen3 / gpt-oss MoE）、权重加载、KV cache、采样
+│   └── xserv-server/      # tokio + axum HTTP 服务、调度器、TP/PP 引擎
+├── tools/                 # 辅助脚本 + benchmark 套件（见下）
+└── docs/                  # 每个 Phase 的设计文档 + benchmark 报告
+```
+
+## 环境要求
+
+- **GPU**：NVIDIA，计算能力 SM120（RTX 5090 / Blackwell）。其它架构需调整 `CUDA_ARCH`。
+- **CUDA Toolkit**：12.9（`nvcc` 需在 `PATH`，构建 `.cu` 依赖它）
+- **Rust**：edition 2024（建议较新的 stable 工具链）
+- **模型**：HuggingFace 目录格式（含 `config.json`、`tokenizer.json`、`*.safetensors`）
+
+## 构建
+
+```bash
+export CUDA_HOME=/usr/local/cuda-12.9
+export PATH=$CUDA_HOME/bin:$PATH
+cargo build --release
+```
+
+如果本地没有 GPU/CUDA，可用远端构建脚本把代码同步到带卡的机器上构建/运行/测试：
+
+```bash
+./tools/sync-and-build.sh build      # 远端 cargo build --release
+./tools/sync-and-build.sh test       # 远端 cargo test
+```
+
+（远端主机、目录、模型路径在 `tools/sync-and-build.sh` 顶部配置。）
+
+## 基本用法
+
+### 1. 启动 HTTP 服务（OpenAI 兼容）
+
+```bash
+./target/release/xserv-server /path/to/qwen3-8b \
+    --port 8080 \
+    --max-batch 4 \
+    --max-seq-len 8192 \
+    --swap-space-gb 8
+```
+
+参数说明：
+
+| 参数 | 含义 | 默认 |
+|------|------|------|
+| `--port` | 监听端口 | 8080 |
+| `--max-batch` | 解码批大小（并发上限） | 4 |
+| `--max-seq-len` | 单序列最大长度 | 2048 |
+| `--swap-space-gb` | KV 换出到 CPU 的 pinned 内存大小（0 关闭） | 8 |
+
+请求示例（流式）：
+
+```bash
+curl http://localhost:8080/v1/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "qwen3-8b",
+    "messages": [{"role": "user", "content": "用一句话解释什么是注意力机制"}],
+    "max_tokens": 256,
+    "temperature": 0,
+    "stream": true
+  }'
+```
+
+其它端点：`GET /health`、`GET /v1/models`。
+
+### 2. 命令行推理
+
+```bash
+# 单轮生成
+cargo run --release --bin xserv-cli -- /path/to/qwen3-8b --max-tokens 256
+
+# 交互式多轮对话
+cargo run --release --bin xserv-chat -- /path/to/qwen3-8b
+```
+
+### 3. 单机性能基准
+
+```bash
+# 输出每个 prompt 的 TTFT / TBT / TPOT（JSON）
+cargo run --release --bin bench-qwen3 -- /path/to/qwen3-8b --gen-tokens 64 [--cuda-graph]
+```
+
+## 与 llama.cpp 对比 benchmark
+
+`tools/bench/` 提供一套一键对比套件，把 xserv 和 **llama.cpp**（同一份 BF16 权重）放在
+相同负载下，黑盒通过 OpenAI API 对比：
+
+- **性能**：TTFT、TPOT、吞吐（单流 + 不同并发）
+- **精度**：AIME 2025、GSM8K（标准数据集，exact-match 评分）
+
+```bash
+# 一次性准备（需联网的机器）：拉取 llama.cpp 子模块 + 下载数据集
+git submodule update --init third_party/llama.cpp        # 固定在 tag b9371
+HF_ENDPOINT=https://hf-mirror.com python3 -m tools.bench.fetch_datasets
+
+# 一键对比（构建 llama.cpp + 转 GGUF + 构建 xserv + 跑两套 + 出报告）
+./tools/sync-and-build.sh bench -- --max-seq-len 8192 --quality-limit 50
+./tools/sync-and-build.sh fetch-bench-out
+# 报告产物：bench-out/comparison-<时间戳>.{md,json}
+```
+
+设计细节见 `docs/16-llama-cpp-comparison.md`，结果报告见 `docs/benchmarks/llama-cpp-comparison.md`。
+
+## 文档
+
+- `docs/00-roadmap.md`：总体路线图与各 Phase 设计
+- `docs/01..15-*.md`：CUDA FFI / Tensor / GEMM / Attention / KV cache / 性能优化等每个 Phase 的设计文档
+- `docs/16-llama-cpp-comparison.md`：llama.cpp 对比基准的设计
+- `docs/17-tensor-parallelism.md`：张量并行（TP）设计
+- `docs/18-pipeline-parallelism.md`：流水线并行（PP）设计
+- `docs/benchmarks/`：各阶段的 benchmark 报告（含 `pp-sweep.md`）
+
+## 多卡并行（TP / PP）
+
+单机多卡，复用 NCCL（crate `xserv-distributed`）。两种切法正交、二选一：
+
+- **张量并行 `--tp N`**：按 head / 中间维切每一层，层内用 AllReduce 聚合（每 token `2·层数` 次）。
+- **流水线并行 `--pp N`**：按层切成 N 段，相邻段间用 NCCL **P2P** 传 hidden state（每 token 仅 `N-1` 次），
+  通信量远小于 AllReduce，对无 NVLink 的 PCIe 更友好。
+
+```bash
+# 组内 GPU 0-3：4 卡张量并行 / 4 卡流水线并行
+CUDA_VISIBLE_DEVICES=0,1,2,3 ./target/release/xserv-server /path/to/qwen3-8b --tp 4
+CUDA_VISIBLE_DEVICES=0,1,2,3 ./target/release/xserv-server /path/to/qwen3-8b --pp 4
+```
+
+**PP 实测**（dash5，Qwen3-8B BF16，单流贪心；每卡显存为权重+最小 KV 池）：
+
+| 配置 | TTFT | TPOT | tok/s | 每卡显存 |
+|------|------|------|-------|----------|
+| 单卡 | 33ms | 17.4ms | 57.5 | 24.0 GB |
+| PP=2 | 36ms | 18.1ms | 55.3 | 11.6 / 13.6 GB |
+| PP=4 | 36ms | 17.9ms | 55.8 | 7.3 / 5.3 / 5.3 / 9.4 GB |
+
+**质量对比**（AIME 2025 30 题 + GSM8K 30 题，贪心，xserv 在 GPU 0-3、llama.cpp 在 GPU 4-7 并行）：
+
+| 引擎 | PP | AIME | GSM8K |
+|------|----|------|-------|
+| xserv | 1/2/4 | 8 / 7 / 7 (/30) | 29/30 (96.7%) 全部一致 |
+| llama | 1/2/4 | 7 / 7 / 7 (/30) | 29/30 (96.7%) 全部一致 |
+
+正确性：hidden state 跨段是 **bit-exact BF16 P2P 拷贝**，PP=4 输出与单卡逐字节一致（用「单卡×2 vs
+PP=4×2」对照确认——单卡自身因 cuBLAS 非确定性 run-to-run 会变，而 PP=4 可复现且落在某次单卡轨迹上）。
+GSM8K 12 个格子全是 29/30，xserv 与 llama.cpp 完全一致；AIME 的 ±1 是长生成下贪心对 GEMM 抖动的敏感，
+非 PP 或引擎效应。**收益在显存**（每卡权重+KV ≈ 1/N）；v1 为串行流水线，单流 TPOT 基本持平、不优于单卡，
+真正的吞吐提升需后续做 microbatch / 1F1B 重叠。完整数据见 `docs/benchmarks/pp-sweep.md`。
+
+## 路线图（节选）
+
+已完成 Phase 0–21：CUDA 基础设施 → Tensor → GEMM → Transformer kernels → Attention →
+模型加载 → 分词器 → GPT-2 → KV cache → Qwen3-8B → Paged Attention → 连续批处理 →
+HTTP API → Flash Attention 2 → 性能优化 → **张量并行（TP）** → **流水线并行（PP）** →
+**gpt-oss MoE + FP8/MXFP4 量化** → **稀疏 top-k MoE decode** → **decode CUDA Graph 整图回放**；
+并加入了 **llama.cpp 对比基准** 与 **KV CPU 换出** 等基础设施。
+
+后续方向：非专家权重量化（lm_head/qkv/o）、稀疏 prefill（grouped GEMM）、server 侧 harmony
+channel 分离、PP microbatch/1F1B、投机解码、多模态。详见 `docs/00-roadmap.md` 的实际进展记录。
+
+## 许可
+
+MIT

crates/xserv-cuda/src/allocator.rs

@@ -1,6 +1,7 @@
 use crate::error::Result;
 use crate::ffi;
 use crate::memory::GpuBuffer;
+use std::cell::RefCell;
 use std::collections::HashMap;
 
 /// Caching allocator that reuses freed GPU buffers instead of calling
@@ -84,6 +85,94 @@ impl Drop for CachingAllocator {
     }
 }
 
+thread_local! {
+    static ALLOCATOR: RefCell<CachingAllocator> = RefCell::new(CachingAllocator::new());
+}
+
+/// Allocate a GPU buffer through the caching allocator.
+/// The returned buffer has `pooled = true` so it will be returned
+/// to the pool on drop instead of calling cudaFree.
+pub fn cached_alloc(size: usize) -> Result<GpuBuffer> {
+    ALLOCATOR.with(|cell| {
+        let mut buf = cell.borrow_mut().alloc(size)?;
+        buf.set_pooled(true);
+        Ok(buf)
+    })
+}
+
+/// Free all cached (unused) GPU buffers back to the driver.
+pub fn cached_trim() {
+    ALLOCATOR.with(|cell| {
+        cell.borrow_mut().trim();
+    });
+}
+
+/// Return a raw GPU pointer to the caching allocator's free list.
+/// Called from `GpuBuffer::Drop` for pooled buffers. Takes raw pointer
+/// and size to avoid re-triggering Drop.
+pub fn return_to_pool(ptr: *mut u8, len: usize) {
+    // During CUDA graph capture, buffers freed by the captured code are
+    // quarantined instead of pooled: the instantiated graph references their
+    // addresses on every replay, so they must never be handed to another
+    // consumer for as long as the graph lives.
+    let quarantined = RETAINED.with(|cell| {
+        let mut r = cell.borrow_mut();
+        if let Some(list) = r.as_mut() {
+            list.push((ptr, len));
+            true
+        } else {
+            false
+        }
+    });
+    if quarantined {
+        return;
+    }
+    ALLOCATOR.with(|cell| {
+        let mut alloc = cell.borrow_mut();
+        let bucket = bucket_size(len);
+        alloc.stats.current_allocated = alloc.stats.current_allocated.saturating_sub(len);
+        alloc.free_lists.entry(bucket).or_default().push((ptr, len));
+    });
+}
+
+thread_local! {
+    static RETAINED: RefCell<Option<Vec<(*mut u8, usize)>>> = const { RefCell::new(None) };
+}
+
+/// Buffers freed while a retain window was active. Holding this keeps their
+/// memory out of the pool; dropping it returns the blocks (on the owning
+/// thread) for reuse.
+pub struct RetainedBlocks(Vec<(*mut u8, usize)>);
+
+impl Drop for RetainedBlocks {
+    fn drop(&mut self) {
+        for (ptr, len) in self.0.drain(..) {
+            return_to_pool(ptr, len);
+        }
+    }
+}
+
+/// Start quarantining buffers freed on this thread (see `return_to_pool`).
+/// Must be paired with `end_retain` on the same thread; nesting unsupported.
+pub fn begin_retain() {
+    RETAINED.with(|cell| {
+        let mut r = cell.borrow_mut();
+        assert!(r.is_none(), "begin_retain: retain window already active");
+        *r = Some(Vec::new());
+    });
+}
+
+/// Stop quarantining and hand the quarantined blocks to the caller.
+pub fn end_retain() -> RetainedBlocks {
+    RETAINED.with(|cell| {
+        let list = cell
+            .borrow_mut()
+            .take()
+            .expect("end_retain without begin_retain");
+        RetainedBlocks(list)
+    })
+}
+
 /// Round up to next power-of-2, minimum 512 bytes.
 fn bucket_size(size: usize) -> usize {
     let min = 512;

crates/xserv-cuda/src/device.rs

@@ -1,6 +1,7 @@
 use crate::error::{self, Result};
 use crate::ffi;
 use std::ffi::CStr;
+use std::os::raw::c_char;
 
 #[derive(Debug, Clone)]
 pub struct DeviceInfo {
@@ -44,10 +45,12 @@ pub fn current_device() -> Result<u32> {
 }
 
 pub fn device_info(device: u32) -> Result<DeviceInfo> {
-    // Get device name from cudaGetDeviceProperties (only use the name field).
-    let mut prop = unsafe { std::mem::zeroed::<ffi::CudaDeviceProp>() };
-    error::check(unsafe { ffi::cudaGetDeviceProperties(&mut prop, device as i32) })?;
-    let name = unsafe { CStr::from_ptr(prop.name.as_ptr()) }
+    // Heap-allocate oversized buffer for cudaDeviceProp (layout varies by CUDA version).
+    // CUDA 12.x struct is ~5-6 KB; use 32 KB to guard against future growth.
+    let mut prop_buf = vec![0u8; 32768];
+    error::check(unsafe { ffi::cudaGetDeviceProperties(prop_buf.as_mut_ptr(), device as i32) })?;
+    // Name is always the first field: char[256].
+    let name = unsafe { CStr::from_ptr(prop_buf.as_ptr() as *const c_char) }
         .to_string_lossy()
         .into_owned();
 

crates/xserv-cuda/src/ffi.rs

@@ -3,6 +3,8 @@ use std::os::raw::c_char;
 
 pub type CudaStream = *mut c_void;
 pub type CudaEvent = *mut c_void;
+pub type CudaGraph = *mut c_void;
+pub type CudaGraphExec = *mut c_void;
 
 pub const CUDA_MEMCPY_H2D: i32 = 1;
 pub const CUDA_MEMCPY_D2H: i32 = 2;
@@ -11,31 +13,17 @@ pub const CUDA_MEMCPY_D2D: i32 = 3;
 pub const CUDA_SUCCESS: i32 = 0;
 pub const CUDA_ERROR_OUT_OF_MEMORY: i32 = 2;
 
-#[repr(C)]
-pub struct CudaDeviceProp {
-    pub name: [c_char; 256],
-    pub total_global_mem: usize,
-    pub shared_mem_per_block: usize,
-    pub regs_per_block: i32,
-    pub warp_size: i32,
-    pub max_threads_per_block: i32,
-    pub max_threads_dim: [i32; 3],
-    pub max_grid_size: [i32; 3],
-    pub clock_rate: i32,
-    pub total_const_mem: usize,
-    pub major: i32,
-    pub minor: i32,
-    // There are many more fields; we only read up to what we need.
-    // cudaDeviceProp is a large struct (~1KB). We pad the rest.
-    _pad: [u8; 4096],
-}
+/// cudaStreamCaptureMode::cudaStreamCaptureModeGlobal
+pub const CUDA_STREAM_CAPTURE_MODE_GLOBAL: i32 = 0;
+pub const CUDA_STREAM_CAPTURE_MODE_THREAD_LOCAL: i32 = 1;
 
 unsafe extern "C" {
     // --- Device ---
     pub fn cudaGetDeviceCount(count: *mut i32) -> i32;
     pub fn cudaSetDevice(device: i32) -> i32;
     pub fn cudaGetDevice(device: *mut i32) -> i32;
-    pub fn cudaGetDeviceProperties(prop: *mut CudaDeviceProp, device: i32) -> i32;
+    /// Takes a raw pointer; caller provides a heap buffer large enough for any CUDA version.
+    pub fn cudaGetDeviceProperties(prop: *mut u8, device: i32) -> i32;
     pub fn cudaDeviceSynchronize() -> i32;
 
     // --- Memory ---
@@ -52,6 +40,7 @@ unsafe extern "C" {
         stream: CudaStream,
     ) -> i32;
     pub fn cudaMemset(devptr: *mut u8, value: i32, count: usize) -> i32;
+    pub fn cudaMemsetAsync(devptr: *mut u8, value: i32, count: usize, stream: CudaStream) -> i32;
 
     // --- Stream ---
     pub fn cudaStreamCreate(stream: *mut CudaStream) -> i32;
@@ -62,12 +51,18 @@ unsafe extern "C" {
     pub fn cudaGetLastError() -> i32;
     pub fn cudaGetErrorString(error: i32) -> *const c_char;
 
+    // --- CUDA Graphs ---
+    pub fn cudaStreamBeginCapture(stream: CudaStream, mode: i32) -> i32;
+    pub fn cudaStreamEndCapture(stream: CudaStream, graph: *mut CudaGraph) -> i32;
+    pub fn cudaGraphInstantiate(
+        graph_exec: *mut CudaGraphExec,
+        graph: CudaGraph,
+        flags: u64,
+    ) -> i32;
+    pub fn cudaGraphLaunch(graph_exec: CudaGraphExec, stream: CudaStream) -> i32;
+    pub fn cudaGraphDestroy(graph: CudaGraph) -> i32;
+    pub fn cudaGraphExecDestroy(graph_exec: CudaGraphExec) -> i32;
+
     // --- Our test kernel ---
-    pub fn launch_vecadd_f32(
-        a: *const f32,
-        b: *const f32,
-        c: *mut f32,
-        n: i32,
-        stream: CudaStream,
-    );
+    pub fn launch_vecadd_f32(a: *const f32, b: *const f32, c: *mut f32, n: i32, stream: CudaStream);
 }

crates/xserv-cuda/src/graph.rs

@@ -0,0 +1,92 @@
+//! CUDA Graphs: capture a sequence of kernel launches and replay them with
+//! near-zero host-side overhead (~3-5 us per launch eliminated).
+//!
+//! Usage:
+//! ```ignore
+//! let stream = CudaStream::new()?;
+//! let mut graph = CudaGraph::new();
+//!
+//! // First call: capture
+//! graph.begin_capture(&stream)?;
+//! // ... launch kernels on `stream` ...
+//! graph.end_capture(&stream)?;
+//!
+//! // Subsequent calls: replay
+//! graph.launch(&stream)?;
+//! ```
+//!
+//! Requirements for captured kernels:
+//! - All tensor shapes must be identical between capture and replay.
+//! - No host-side branching during the captured section.
+//! - Memory addresses used during capture must remain valid during replay.
+
+use crate::error::{self, Result};
+use crate::ffi;
+use crate::stream::CudaStream;
+
+/// RAII wrapper around a captured CUDA graph and its executable instance.
+pub struct CudaGraph {
+    graph: ffi::CudaGraph,
+    exec: ffi::CudaGraphExec,
+}
+
+impl CudaGraph {
+    /// Create an empty graph handle (not yet captured).
+    pub fn new() -> Self {
+        Self {
+            graph: std::ptr::null_mut(),
+            exec: std::ptr::null_mut(),
+        }
+    }
+
+    /// Returns true if a graph has been captured and instantiated.
+    pub fn is_ready(&self) -> bool {
+        !self.exec.is_null()
+    }
+
+    /// Begin capturing kernel launches on `stream`.
+    /// All subsequent kernel launches on this stream are recorded into the
+    /// graph instead of being executed.
+    pub fn begin_capture(&mut self, stream: &CudaStream) -> Result<()> {
+        // If we have an old graph, destroy it first
+        self.destroy_inner();
+        // THREAD_LOCAL: only "potentially unsafe" CUDA calls (cudaMalloc etc.)
+        // made by THIS thread invalidate the capture. With GLOBAL mode, TP rank
+        // threads capturing concurrently would poison each other's captures.
+        error::check(unsafe {
+            ffi::cudaStreamBeginCapture(stream.as_raw(), ffi::CUDA_STREAM_CAPTURE_MODE_THREAD_LOCAL)
+        })
+    }
+
+    /// End capture and instantiate the executable graph.
+    pub fn end_capture(&mut self, stream: &CudaStream) -> Result<()> {
+        error::check(unsafe { ffi::cudaStreamEndCapture(stream.as_raw(), &mut self.graph) })?;
+        error::check(unsafe { ffi::cudaGraphInstantiate(&mut self.exec, self.graph, 0) })
+    }
+
+    /// Replay the captured graph on `stream`.
+    /// Panics if no graph has been captured yet.
+    pub fn launch(&self, stream: &CudaStream) -> Result<()> {
+        assert!(self.is_ready(), "CudaGraph::launch called before capture");
+        error::check(unsafe { ffi::cudaGraphLaunch(self.exec, stream.as_raw()) })
+    }
+
+    fn destroy_inner(&mut self) {
+        if !self.exec.is_null() {
+            unsafe { ffi::cudaGraphExecDestroy(self.exec) };
+            self.exec = std::ptr::null_mut();
+        }
+        if !self.graph.is_null() {
+            unsafe { ffi::cudaGraphDestroy(self.graph) };
+            self.graph = std::ptr::null_mut();
+        }
+    }
+}
+
+impl Drop for CudaGraph {
+    fn drop(&mut self) {
+        self.destroy_inner();
+    }
+}
+
+unsafe impl Send for CudaGraph {}

crates/xserv-cuda/src/lib.rs

@@ -2,11 +2,13 @@ pub mod allocator;
 pub mod device;
 pub mod error;
 pub mod ffi;
+pub mod graph;
 pub mod memory;
 pub mod stream;
 
 pub use allocator::CachingAllocator;
 pub use device::DeviceInfo;
 pub use error::{CudaError, Result};
+pub use graph::CudaGraph;
 pub use memory::{GpuBuffer, PinnedBuffer};
-pub use stream::CudaStream;
+pub use stream::{CudaStream, StreamGuard, current_stream_raw, push_stream};

crates/xserv-cuda/src/memory.rs

@@ -3,9 +3,18 @@ use crate::ffi;
 use crate::stream::CudaStream;
 
 /// RAII wrapper around a GPU memory allocation.
+///
+/// When `owned` is true (the default), dropping frees the GPU memory.
+/// A borrowed buffer (`owned = false`) does NOT free on drop — the
+/// caller must ensure the backing allocation outlives all borrows.
+///
+/// When `pooled` is true, dropping returns the buffer to the caching
+/// allocator's free list instead of calling cudaFree.
 pub struct GpuBuffer {
     ptr: *mut u8,
     len: usize,
+    owned: bool,
+    pooled: bool,
 }
 
 impl GpuBuffer {
@@ -13,7 +22,18 @@ impl GpuBuffer {
         assert!(len > 0, "cannot allocate 0 bytes on GPU");
         let mut ptr = std::ptr::null_mut();
         error::check(unsafe { ffi::cudaMalloc(&mut ptr, len) })?;
-        Ok(Self { ptr, len })
+        Ok(Self {
+            ptr,
+            len,
+            owned: true,
+            pooled: false,
+        })
+    }
+
+    /// Mark this buffer as pooled (returned to caching allocator on drop)
+    /// or not. Called by `cached_alloc` after obtaining a buffer.
+    pub fn set_pooled(&mut self, pooled: bool) {
+        self.pooled = pooled;
     }
 
     pub fn len(&self) -> usize {
@@ -77,9 +97,7 @@ impl GpuBuffer {
     /// Copy from another GPU buffer (D2D).
     pub fn copy_from_device(&mut self, src: &GpuBuffer) -> Result<()> {
         let n = src.len.min(self.len);
-        error::check(unsafe {
-            ffi::cudaMemcpy(self.ptr, src.ptr, n, ffi::CUDA_MEMCPY_D2D)
-        })
+        error::check(unsafe { ffi::cudaMemcpy(self.ptr, src.ptr, n, ffi::CUDA_MEMCPY_D2D) })
     }
 
     /// Fill buffer with zeros.
@@ -87,6 +105,81 @@ impl GpuBuffer {
         error::check(unsafe { ffi::cudaMemset(self.ptr, 0, self.len) })
     }
 
+    /// Copy `count` bytes from `src` buffer at `src_offset` to this buffer at `dst_offset`.
+    pub fn copy_from_device_at(
+        &mut self,
+        src: &GpuBuffer,
+        src_offset: usize,
+        dst_offset: usize,
+        count: usize,
+    ) -> Result<()> {
+        assert!(src_offset + count <= src.len);
+        assert!(dst_offset + count <= self.len);
+        error::check(unsafe {
+            ffi::cudaMemcpy(
+                self.ptr.add(dst_offset),
+                src.ptr.add(src_offset),
+                count,
+                ffi::CUDA_MEMCPY_D2D,
+            )
+        })
+    }
+
+    /// Async copy `count` bytes from `src` at `src_offset` to `self` at `dst_offset` on `stream`.
+    pub fn copy_from_device_at_async(
+        &mut self,
+        src: &GpuBuffer,
+        src_offset: usize,
+        dst_offset: usize,
+        count: usize,
+        stream: &CudaStream,
+    ) -> Result<()> {
+        assert!(src_offset + count <= src.len);
+        assert!(dst_offset + count <= self.len);
+        error::check(unsafe {
+            ffi::cudaMemcpyAsync(
+                self.ptr.add(dst_offset),
+                src.ptr.add(src_offset),
+                count,
+                ffi::CUDA_MEMCPY_D2D,
+                stream.as_raw(),
+            )
+        })
+    }
+
+    /// Copy `count` bytes from this GPU buffer at `src_offset` to a host slice (D2H).
+    pub fn copy_to_host_at(&self, dst: &mut [u8], src_offset: usize, count: usize) -> Result<()> {
+        assert!(src_offset + count <= self.len, "src range out of bounds");
+        assert!(count <= dst.len(), "host dst too small");
+        error::check(unsafe {
+            ffi::cudaMemcpy(
+                dst.as_mut_ptr(),
+                self.ptr.add(src_offset),
+                count,
+                ffi::CUDA_MEMCPY_D2H,
+            )
+        })
+    }
+
+    /// Copy `count` bytes from a host slice to this GPU buffer at `dst_offset` (H2D).
+    pub fn copy_from_host_at(&mut self, src: &[u8], dst_offset: usize, count: usize) -> Result<()> {
+        assert!(dst_offset + count <= self.len, "dst range out of bounds");
+        assert!(count <= src.len(), "host src too small");
+        error::check(unsafe {
+            ffi::cudaMemcpy(
+                self.ptr.add(dst_offset),
+                src.as_ptr(),
+                count,
+                ffi::CUDA_MEMCPY_H2D,
+            )
+        })
+    }
+
+    /// Async zero fill on stream.
+    pub fn zero_async(&mut self, stream: &CudaStream) -> Result<()> {
+        error::check(unsafe { ffi::cudaMemsetAsync(self.ptr, 0, self.len, stream.as_raw()) })
+    }
+
     /// Consume the buffer without freeing GPU memory. Returns the raw pointer and length.
     /// Caller is responsible for eventually calling cudaFree.
     pub fn into_raw(self) -> (*mut u8, usize) {
@@ -99,14 +192,39 @@ impl GpuBuffer {
     /// Reconstruct a GpuBuffer from a raw pointer + length.
     /// Safety: ptr must have been allocated with cudaMalloc, len must be correct.
     pub unsafe fn from_raw(ptr: *mut u8, len: usize) -> Self {
-        Self { ptr, len }
+        Self {
+            ptr,
+            len,
+            owned: true,
+            pooled: false,
+        }
+    }
+
+    /// Create a non-owning view of GPU memory. Dropping this buffer does NOT
+    /// call `cudaFree`. The caller must ensure the underlying allocation
+    /// outlives this borrow.
+    ///
+    /// # Safety
+    /// `ptr` must point to a valid GPU allocation of at least `len` bytes that
+    /// will remain live for the lifetime of the returned `GpuBuffer`.
+    pub unsafe fn borrow_raw(ptr: *mut u8, len: usize) -> Self {
+        Self {
+            ptr,
+            len,
+            owned: false,
+            pooled: false,
+        }
     }
 }
 
 impl Drop for GpuBuffer {
     fn drop(&mut self) {
-        if !self.ptr.is_null() {
-            unsafe { ffi::cudaFree(self.ptr) };
+        if self.owned && !self.ptr.is_null() {
+            if self.pooled {
+                crate::allocator::return_to_pool(self.ptr, self.len);
+            } else {
+                unsafe { ffi::cudaFree(self.ptr) };
+            }
         }
     }
 }

crates/xserv-cuda/src/stream.rs

@@ -31,3 +31,39 @@ impl Drop for CudaStream {
 
 // Can move across threads, but not shared without synchronization
 unsafe impl Send for CudaStream {}
+
+// --- Thread-local launch stream -------------------------------------------
+//
+// Every kernel wrapper in xserv-kernels launches on `current_stream_raw()`,
+// which defaults to the legacy null stream (the historical behavior). CUDA
+// graph capture requires work to be issued on an explicit stream, so capture
+// code installs its stream here for the duration of the captured region via
+// `push_stream` / `StreamGuard`.
+
+use std::cell::Cell;
+
+thread_local! {
+    static CURRENT_STREAM: Cell<ffi::CudaStream> = const { Cell::new(std::ptr::null_mut()) };
+}
+
+/// The stream kernel launches on this thread should use (null = legacy default).
+pub fn current_stream_raw() -> ffi::CudaStream {
+    CURRENT_STREAM.with(|c| c.get())
+}
+
+/// RAII guard that installs a launch stream for the current thread and
+/// restores the previous one on drop.
+pub struct StreamGuard {
+    prev: ffi::CudaStream,
+}
+
+pub fn push_stream(stream: &CudaStream) -> StreamGuard {
+    let prev = CURRENT_STREAM.with(|c| c.replace(stream.as_raw()));
+    StreamGuard { prev }
+}
+
+impl Drop for StreamGuard {
+    fn drop(&mut self) {
+        CURRENT_STREAM.with(|c| c.set(self.prev));
+    }
+}

crates/xserv-cuda/tests/integration.rs

@@ -14,7 +14,10 @@ fn test_device_info() {
         info.compute_major, info.compute_minor
     );
     println!("  SM Count: {}", info.sm_count);
-    println!("  Shared Mem/Block: {} KB", info.shared_mem_per_block / 1024);
+    println!(
+        "  Shared Mem/Block: {} KB",
+        info.shared_mem_per_block / 1024
+    );
     println!("  Warp Size: {}", info.warp_size);
     println!("  Max Threads/Block: {}", info.max_threads_per_block);
 
@@ -145,7 +148,11 @@ fn test_caching_allocator() {
 
     // Second allocation of same size: should hit cache
     let _buf2 = alloc.alloc(1024).unwrap();
-    assert_eq!(alloc.stats().cuda_malloc_count, 1, "should reuse cached buffer");
+    assert_eq!(
+        alloc.stats().cuda_malloc_count,
+        1,
+        "should reuse cached buffer"
+    );
     assert_eq!(alloc.stats().cache_hit_count, 1);
 }
 
@@ -198,11 +205,17 @@ fn test_async_copy() {
     }
 
     let mut gpu = GpuBuffer::alloc(4096).unwrap();
-    unsafe { gpu.copy_from_host_async(pinned.as_slice(), &stream).unwrap() };
+    unsafe {
+        gpu.copy_from_host_async(pinned.as_slice(), &stream)
+            .unwrap()
+    };
     stream.synchronize().unwrap();
 
     let mut out_pinned = PinnedBuffer::alloc(4096).unwrap();
-    unsafe { gpu.copy_to_host_async(out_pinned.as_mut_slice(), &stream).unwrap() };
+    unsafe {
+        gpu.copy_to_host_async(out_pinned.as_mut_slice(), &stream)
+            .unwrap()
+    };
     stream.synchronize().unwrap();
 
     assert_eq!(pinned.as_slice(), out_pinned.as_slice());

crates/xserv-distributed/Cargo.toml

@@ -0,0 +1,8 @@
+[package]
+name = "xserv-distributed"
+version.workspace = true
+edition.workspace = true
+
+[dependencies]
+xserv-cuda = { path = "../xserv-cuda" }
+half.workspace = true

crates/xserv-distributed/build.rs

@@ -0,0 +1,13 @@
+use std::env;
+
+fn main() {
+    let cuda_path = env::var("CUDA_HOME")
+        .or_else(|_| env::var("CUDA_PATH"))
+        .unwrap_or_else(|_| "/usr/local/cuda".to_string());
+
+    println!("cargo:rustc-link-search=native={cuda_path}/lib64");
+    // NCCL is typically installed as a system library.
+    println!("cargo:rustc-link-search=native=/usr/lib/x86_64-linux-gnu");
+    println!("cargo:rustc-link-lib=dylib=nccl");
+    println!("cargo:rustc-link-lib=dylib=cudart");
+}

crates/xserv-distributed/src/ffi.rs

@@ -0,0 +1,92 @@
+//! Minimal NCCL FFI bindings (hand-written, like the CUDA bindings).
+//! Only the collectives we need for tensor parallelism.
+
+use std::ffi::c_void;
+use std::os::raw::c_char;
+use xserv_cuda::ffi::CudaStream;
+
+/// Opaque NCCL communicator handle (`ncclComm_t`).
+pub type NcclComm = *mut c_void;
+
+/// `ncclUniqueId` is a 128-byte opaque blob shared from rank 0 to all ranks.
+#[repr(C)]
+#[derive(Clone, Copy)]
+pub struct NcclUniqueId {
+    pub internal: [c_char; 128],
+}
+
+impl Default for NcclUniqueId {
+    fn default() -> Self {
+        Self { internal: [0; 128] }
+    }
+}
+
+// ncclDataType_t (subset)
+pub const NCCL_FLOAT32: i32 = 7;
+pub const NCCL_BF16: i32 = 9;
+
+// ncclRedOp_t
+pub const NCCL_SUM: i32 = 0;
+
+// ncclResult_t
+pub const NCCL_SUCCESS: i32 = 0;
+
+unsafe extern "C" {
+    pub fn ncclGetUniqueId(uid: *mut NcclUniqueId) -> i32;
+    // ncclUniqueId is passed BY VALUE (a 128-byte struct) per the NCCL ABI.
+    pub fn ncclCommInitRank(
+        comm: *mut NcclComm,
+        nranks: i32,
+        commid: NcclUniqueId,
+        rank: i32,
+    ) -> i32;
+    pub fn ncclCommDestroy(comm: NcclComm) -> i32;
+    pub fn ncclAllReduce(
+        sendbuff: *const c_void,
+        recvbuff: *mut c_void,
+        count: usize,
+        datatype: i32,
+        op: i32,
+        comm: NcclComm,
+        stream: CudaStream,
+    ) -> i32;
+    // Point-to-point primitives for pipeline parallelism (Phase 18).
+    pub fn ncclSend(
+        sendbuff: *const c_void,
+        count: usize,
+        datatype: i32,
+        peer: i32,
+        comm: NcclComm,
+        stream: CudaStream,
+    ) -> i32;
+    pub fn ncclRecv(
+        recvbuff: *mut c_void,
+        count: usize,
+        datatype: i32,
+        peer: i32,
+        comm: NcclComm,
+        stream: CudaStream,
+    ) -> i32;
+    pub fn ncclGroupStart() -> i32;
+    pub fn ncclGroupEnd() -> i32;
+    pub fn ncclGetErrorString(result: i32) -> *const c_char;
+}
+
+pub fn err_string(result: i32) -> String {
+    unsafe {
+        let p = ncclGetErrorString(result);
+        if p.is_null() {
+            return format!("nccl error {result}");
+        }
+        std::ffi::CStr::from_ptr(p).to_string_lossy().into_owned()
+    }
+}
+
+pub fn check(result: i32, what: &str) {
+    assert_eq!(
+        result,
+        NCCL_SUCCESS,
+        "{what} failed: {}",
+        err_string(result)
+    );
+}

crates/xserv-distributed/src/lib.rs

@@ -0,0 +1,192 @@
+//! Tensor-parallel primitives for xserv.
+//!
+//! Process model: one OS thread per TP rank, each bound to one GPU. NCCL is
+//! used for the collective (AllReduce); a hand-rolled P2P AllReduce may replace
+//! it later as a learning exercise (see docs/17-tensor-parallelism.md).
+
+pub mod ffi;
+
+use std::ffi::c_void;
+
+use ffi::{NcclComm, NcclUniqueId};
+use xserv_cuda::GpuBuffer;
+use xserv_cuda::device;
+
+pub use ffi::NcclUniqueId as UniqueId;
+
+/// NCCL is issued on the thread's current launch stream (legacy null stream
+/// by default, the capture stream during CUDA graph capture). The model's
+/// kernels run on the same stream, so AllReduce stays correctly ordered after
+/// the producing matmul and before the consuming kernel — no extra sync.
+fn launch_stream() -> xserv_cuda::ffi::CudaStream {
+    xserv_cuda::stream::current_stream_raw()
+}
+
+/// Generate a unique id on one rank (typically rank 0) and broadcast the bytes
+/// to all ranks out-of-band (e.g. via a shared variable across threads).
+pub fn get_unique_id() -> NcclUniqueId {
+    let mut id = NcclUniqueId::default();
+    ffi::check(unsafe { ffi::ncclGetUniqueId(&mut id) }, "ncclGetUniqueId");
+    id
+}
+
+/// Per-rank tensor-parallel context: NCCL communicator + a dedicated stream.
+pub struct TpContext {
+    pub rank: usize,
+    pub world: usize,
+    pub device: u32,
+    comm: NcclComm,
+}
+
+// The NCCL communicator is owned by exactly one rank thread.
+unsafe impl Send for TpContext {}
+
+impl TpContext {
+    /// Initialize this rank. Must be called from the thread that will own this
+    /// rank's GPU work; binds the thread to `device` first. All ranks must call
+    /// this concurrently with the same `id` and `world`.
+    pub fn init(rank: usize, world: usize, id: NcclUniqueId, device: u32) -> Self {
+        device::set_device(device).expect("set_device");
+        let mut comm: NcclComm = std::ptr::null_mut();
+        // Wrap the concurrent inits in a group so they rendezvous without deadlock.
+        ffi::check(unsafe { ffi::ncclGroupStart() }, "ncclGroupStart(init)");
+        ffi::check(
+            unsafe { ffi::ncclCommInitRank(&mut comm, world as i32, id, rank as i32) },
+            "ncclCommInitRank",
+        );
+        ffi::check(unsafe { ffi::ncclGroupEnd() }, "ncclGroupEnd(init)");
+        Self {
+            rank,
+            world,
+            device,
+            comm,
+        }
+    }
+
+    /// In-place AllReduce(sum) over `count` BF16 elements in `buf`.
+    pub fn all_reduce_sum_bf16(&self, buf: &mut GpuBuffer, count: usize) {
+        self.all_reduce_sum_bf16_ptr(buf.as_mut_ptr() as *mut c_void, count);
+    }
+
+    /// In-place AllReduce(sum) directly on a device pointer (`count` BF16 elems),
+    /// issued on the null stream so it is ordered with the model's kernels.
+    /// Asynchronous: a later sync (e.g. the D2H logits copy) completes it.
+    ///
+    /// # Safety
+    /// `ptr` must point to at least `count` BF16 elements of valid device memory
+    /// on this rank's device. The reduction is in-place (send == recv).
+    pub fn all_reduce_sum_bf16_ptr(&self, ptr: *mut c_void, count: usize) {
+        if self.world == 1 {
+            return; // nothing to reduce
+        }
+        ffi::check(
+            unsafe {
+                ffi::ncclAllReduce(
+                    ptr as *const c_void,
+                    ptr,
+                    count,
+                    ffi::NCCL_BF16,
+                    ffi::NCCL_SUM,
+                    self.comm,
+                    launch_stream(),
+                )
+            },
+            "ncclAllReduce",
+        );
+    }
+}
+
+impl Drop for TpContext {
+    fn drop(&mut self) {
+        if !self.comm.is_null() {
+            unsafe { ffi::ncclCommDestroy(self.comm) };
+        }
+    }
+}
+
+/// Per-stage pipeline-parallel context: a NCCL communicator spanning all `P`
+/// stages plus point-to-point send/recv of the hidden state to the neighbour
+/// stages. Init is identical to `TpContext` (one comm across `world` ranks);
+/// only the collective differs — PP hands off `[tokens, hidden]` between
+/// consecutive stages instead of AllReducing within a layer.
+pub struct PpContext {
+    pub stage: usize,
+    pub world: usize,
+    pub device: u32,
+    comm: NcclComm,
+}
+
+// The NCCL communicator is owned by exactly one stage thread.
+unsafe impl Send for PpContext {}
+
+impl PpContext {
+    /// Initialize this stage. Must be called from the thread that owns this
+    /// stage's GPU; binds the thread to `device` first. All stages call this
+    /// concurrently with the same `id` and `world`.
+    pub fn init(stage: usize, world: usize, id: NcclUniqueId, device: u32) -> Self {
+        device::set_device(device).expect("set_device");
+        let mut comm: NcclComm = std::ptr::null_mut();
+        ffi::check(unsafe { ffi::ncclGroupStart() }, "ncclGroupStart(init)");
+        ffi::check(
+            unsafe { ffi::ncclCommInitRank(&mut comm, world as i32, id, stage as i32) },
+            "ncclCommInitRank",
+        );
+        ffi::check(unsafe { ffi::ncclGroupEnd() }, "ncclGroupEnd(init)");
+        Self {
+            stage,
+            world,
+            device,
+            comm,
+        }
+    }
+
+    /// Send `count` BF16 elements at `ptr` to `peer`, on the null stream so it is
+    /// ordered after the producing matmul. Asynchronous — a later `synchronize`
+    /// (the caller must do one before reusing/freeing the buffer) completes it.
+    ///
+    /// # Safety
+    /// `ptr` must point to at least `count` BF16 elements of valid device memory.
+    pub fn send_bf16_ptr(&self, ptr: *const c_void, count: usize, peer: usize) {
+        ffi::check(
+            unsafe {
+                ffi::ncclSend(
+                    ptr,
+                    count,
+                    ffi::NCCL_BF16,
+                    peer as i32,
+                    self.comm,
+                    launch_stream(),
+                )
+            },
+            "ncclSend",
+        );
+    }
+
+    /// Receive `count` BF16 elements from `peer` into `ptr`, on the null stream.
+    ///
+    /// # Safety
+    /// `ptr` must point to at least `count` BF16 elements of valid device memory.
+    pub fn recv_bf16_ptr(&self, ptr: *mut c_void, count: usize, peer: usize) {
+        ffi::check(
+            unsafe {
+                ffi::ncclRecv(
+                    ptr,
+                    count,
+                    ffi::NCCL_BF16,
+                    peer as i32,
+                    self.comm,
+                    launch_stream(),
+                )
+            },
+            "ncclRecv",
+        );
+    }
+}
+
+impl Drop for PpContext {
+    fn drop(&mut self) {
+        if !self.comm.is_null() {
+            unsafe { ffi::ncclCommDestroy(self.comm) };
+        }
+    }
+}

crates/xserv-distributed/tests/allreduce.rs

@@ -0,0 +1,48 @@
+//! 2-GPU AllReduce smoke test. Skips if fewer than 2 GPUs are present.
+
+use half::bf16;
+use std::thread;
+use xserv_cuda::{GpuBuffer, device};
+use xserv_distributed::{TpContext, get_unique_id};
+
+#[test]
+fn allreduce_two_gpu_sum() {
+    let world = 2usize;
+    if device::device_count().unwrap_or(0) < world as i32 {
+        eprintln!("skip: need >= {world} GPUs");
+        return;
+    }
+
+    let id = get_unique_id();
+    let n = 4096usize;
+
+    let handles: Vec<_> = (0..world)
+        .map(|rank| {
+            let id = id;
+            thread::spawn(move || {
+                let ctx = TpContext::init(rank, world, id, rank as u32);
+
+                // Rank r fills its buffer with (r + 1).
+                let val = bf16::from_f32((rank + 1) as f32);
+                let host = vec![val; n];
+                let src = unsafe { std::slice::from_raw_parts(host.as_ptr() as *const u8, n * 2) };
+                let mut buf = GpuBuffer::alloc(n * 2).unwrap();
+                buf.copy_from_host(src).unwrap();
+
+                ctx.all_reduce_sum_bf16(&mut buf, n);
+
+                let mut out = vec![0u8; n * 2];
+                buf.copy_to_host(&mut out).unwrap();
+                let res = unsafe { std::slice::from_raw_parts(out.as_ptr() as *const bf16, n) };
+                (res[0].to_f32(), res[n - 1].to_f32())
+            })
+        })
+        .collect();
+
+    // sum over ranks of (r+1) = 1 + 2 = 3
+    for h in handles {
+        let (first, last) = h.join().unwrap();
+        assert_eq!(first, 3.0, "AllReduce(sum) first element");
+        assert_eq!(last, 3.0, "AllReduce(sum) last element");
+    }
+}

crates/xserv-distributed/tests/sendrecv.rs

@@ -0,0 +1,63 @@
+//! 2-GPU NCCL P2P send/recv smoke test for pipeline parallelism.
+//! Stage 0 sends a known vector to stage 1, which verifies it. Skips if fewer
+//! than 2 GPUs are present. Mirrors `allreduce.rs` (GpuBuffer + half only —
+//! this crate does not depend on xserv-tensor).
+
+use half::bf16;
+use std::ffi::c_void;
+use std::thread;
+use xserv_cuda::{GpuBuffer, device};
+use xserv_distributed::{PpContext, get_unique_id};
+
+#[test]
+fn pp_send_recv_two_stages() {
+    let world = 2usize;
+    if device::device_count().unwrap_or(0) < world as i32 {
+        eprintln!("skip: need >= {world} GPUs");
+        return;
+    }
+
+    let id = get_unique_id();
+    let n = 4096usize; // one [1, hidden]-sized hand-off
+
+    let handles: Vec<_> = (0..world)
+        .map(|stage| {
+            let id = id;
+            thread::spawn(move || {
+                let pp = PpContext::init(stage, world, id, stage as u32);
+                let mut buf = GpuBuffer::alloc(n * 2).unwrap();
+
+                if stage == 0 {
+                    // Fill with a known pattern and send to stage 1.
+                    let host: Vec<bf16> = (0..n).map(|i| bf16::from_f32((i % 97) as f32)).collect();
+                    let src =
+                        unsafe { std::slice::from_raw_parts(host.as_ptr() as *const u8, n * 2) };
+                    buf.copy_from_host(src).unwrap();
+                    pp.send_bf16_ptr(buf.as_mut_ptr() as *const c_void, n, 1);
+                    device::synchronize().unwrap();
+                    None
+                } else {
+                    // Receive into a zeroed buffer and read it back.
+                    buf.copy_from_host(&vec![0u8; n * 2]).unwrap();
+                    pp.recv_bf16_ptr(buf.as_mut_ptr() as *mut c_void, n, 0);
+                    device::synchronize().unwrap();
+                    let mut out = vec![0u8; n * 2];
+                    buf.copy_to_host(&mut out).unwrap();
+                    let res = unsafe { std::slice::from_raw_parts(out.as_ptr() as *const bf16, n) };
+                    Some((res[0].to_f32(), res[1].to_f32(), res[n - 1].to_f32()))
+                }
+            })
+        })
+        .collect();
+
+    let mut checked = false;
+    for h in handles {
+        if let Some((first, second, last)) = h.join().unwrap() {
+            assert_eq!(first, 0.0, "recv[0]");
+            assert_eq!(second, 1.0, "recv[1]");
+            assert_eq!(last, ((n - 1) % 97) as f32, "recv[last]");
+            checked = true;
+        }
+    }
+    assert!(checked, "stage 1 never verified the received buffer");
+}

crates/xserv-kernels/build.rs

@@ -8,6 +8,7 @@ fn main() {
     println!("cargo:rustc-link-search=native={cuda_path}/lib64");
     println!("cargo:rustc-link-lib=dylib=cudart");
     println!("cargo:rustc-link-lib=dylib=cublas");
+    println!("cargo:rustc-link-lib=dylib=cublasLt");
 
     cc::Build::new()
         .cuda(true)
@@ -16,12 +17,24 @@ fn main() {
         .include("../../csrc")
         .file("../../csrc/gemm/naive.cu")
         .file("../../csrc/gemm/tiled.cu")
+        .file("../../csrc/gemm/gemv.cu")
         .file("../../csrc/normalization/rmsnorm.cu")
         .file("../../csrc/normalization/layernorm.cu")
         .file("../../csrc/activation/activations.cu")
         .file("../../csrc/reduce/softmax.cu")
+        .file("../../csrc/reduce/argmax.cu")
         .file("../../csrc/embedding/embedding.cu")
         .file("../../csrc/embedding/rope.cu")
+        .file("../../csrc/attention/causal_mask.cu")
+        .file("../../csrc/embedding/transpose.cu")
+        .file("../../csrc/attention/flash_attention.cu")
+        .file("../../csrc/attention/paged_attention.cu")
+        .file("../../csrc/attention/reshape_and_cache.cu")
+        .file("../../csrc/moe/moe_kernels.cu")
+        .file("../../csrc/moe/moe_sparse.cu")
+        .file("../../csrc/quantization/dequant_fp8.cu")
+        .file("../../csrc/quantization/quantize_fp8.cu")
+        .file("../../csrc/quantization/mxfp4_gemm.cu")
         .compile("xserv_kernels");
 
     println!("cargo:rerun-if-changed=../../csrc/");

crates/xserv-kernels/src/activation.rs

@@ -6,36 +6,271 @@ unsafe extern "C" {
     fn launch_gelu_bf16(x: *const c_void, out: *mut c_void, n: i32, stream: *mut c_void);
     fn launch_silu_f32(x: *const c_void, out: *mut c_void, n: i32, stream: *mut c_void);
     fn launch_silu_bf16(x: *const c_void, out: *mut c_void, n: i32, stream: *mut c_void);
+    fn launch_scale_f32(
+        x: *const c_void,
+        out: *mut c_void,
+        scale: f32,
+        n: i32,
+        stream: *mut c_void,
+    );
+    fn launch_scale_bf16(
+        x: *const c_void,
+        out: *mut c_void,
+        scale: f32,
+        n: i32,
+        stream: *mut c_void,
+    );
+    fn launch_add_f32(
+        a: *const c_void,
+        b: *const c_void,
+        out: *mut c_void,
+        n: i32,
+        stream: *mut c_void,
+    );
+    fn launch_add_bf16(
+        a: *const c_void,
+        b: *const c_void,
+        out: *mut c_void,
+        n: i32,
+        stream: *mut c_void,
+    );
+    fn launch_mul_f32(
+        a: *const c_void,
+        b: *const c_void,
+        out: *mut c_void,
+        n: i32,
+        stream: *mut c_void,
+    );
+    fn launch_mul_bf16(
+        a: *const c_void,
+        b: *const c_void,
+        out: *mut c_void,
+        n: i32,
+        stream: *mut c_void,
+    );
+    fn launch_silu_mul_bf16(
+        gate: *const c_void,
+        up: *const c_void,
+        out: *mut c_void,
+        n: i32,
+        stream: *mut c_void,
+    );
+    fn launch_gpt_oss_glu_bf16(
+        gate_up: *const c_void,
+        out: *mut c_void,
+        n_elements: i32,
+        alpha: f32,
+        limit: f32,
+        stream: *mut c_void,
+    );
+    fn launch_bias_add_2d_bf16(
+        x: *const c_void,
+        bias: *const c_void,
+        out: *mut c_void,
+        rows: i32,
+        cols: i32,
+        stream: *mut c_void,
+    );
+}
+
+fn dispatch_unary(
+    x: &Tensor,
+    f32_fn: unsafe extern "C" fn(*const c_void, *mut c_void, i32, *mut c_void),
+    bf16_fn: unsafe extern "C" fn(*const c_void, *mut c_void, i32, *mut c_void),
+) -> Tensor {
+    assert!(x.is_contiguous() && matches!(x.device(), Device::Cuda(_)));
+    let out = Tensor::empty(x.shape(), x.dtype(), x.device());
+    let n = x.numel();
+    assert!(
+        n <= i32::MAX as usize,
+        "tensor too large for i32 kernel param ({n} elements)"
+    );
+    let n = n as i32;
+    unsafe {
+        match x.dtype() {
+            DType::F32 => f32_fn(
+                x.data_ptr() as _,
+                out.data_ptr() as *mut c_void,
+                n,
+                xserv_cuda::current_stream_raw(),
+            ),
+            DType::BF16 => bf16_fn(
+                x.data_ptr() as _,
+                out.data_ptr() as *mut c_void,
+                n,
+                xserv_cuda::current_stream_raw(),
+            ),
+            _ => panic!("unsupported dtype"),
+        }
+    }
+    out
+}
+
+fn dispatch_binary(
+    a: &Tensor,
+    b: &Tensor,
+    f32_fn: unsafe extern "C" fn(*const c_void, *const c_void, *mut c_void, i32, *mut c_void),
+    bf16_fn: unsafe extern "C" fn(*const c_void, *const c_void, *mut c_void, i32, *mut c_void),
+) -> Tensor {
+    assert_eq!(a.shape(), b.shape());
+    assert!(a.is_contiguous() && b.is_contiguous());
+    assert!(matches!(a.device(), Device::Cuda(_)));
+    assert_eq!(a.dtype(), b.dtype());
+    let out = Tensor::empty(a.shape(), a.dtype(), a.device());
+    let n = a.numel();
+    assert!(
+        n <= i32::MAX as usize,
+        "tensor too large for i32 kernel param ({n} elements)"
+    );
+    let n = n as i32;
+    unsafe {
+        match a.dtype() {
+            DType::F32 => f32_fn(
+                a.data_ptr() as _,
+                b.data_ptr() as _,
+                out.data_ptr() as *mut c_void,
+                n,
+                xserv_cuda::current_stream_raw(),
+            ),
+            DType::BF16 => bf16_fn(
+                a.data_ptr() as _,
+                b.data_ptr() as _,
+                out.data_ptr() as *mut c_void,
+                n,
+                xserv_cuda::current_stream_raw(),
+            ),
+            _ => panic!("unsupported dtype"),
+        }
+    }
+    out
 }
 
 pub fn gelu(x: &Tensor) -> Tensor {
-    assert!(x.is_contiguous());
-    assert!(matches!(x.device(), Device::Cuda(_)));
-    let out = Tensor::zeros(x.shape(), x.dtype(), x.device());
-    let n = x.numel() as i32;
+    dispatch_unary(x, launch_gelu_f32, launch_gelu_bf16)
+}
+pub fn silu(x: &Tensor) -> Tensor {
+    dispatch_unary(x, launch_silu_f32, launch_silu_bf16)
+}
+
+pub fn scale(x: &Tensor, scale_val: f32) -> Tensor {
+    assert!(x.is_contiguous() && matches!(x.device(), Device::Cuda(_)));
+    let out = Tensor::empty(x.shape(), x.dtype(), x.device());
+    let n = x.numel();
+    assert!(
+        n <= i32::MAX as usize,
+        "tensor too large for i32 kernel param ({n} elements)"
+    );
+    let n = n as i32;
     unsafe {
         match x.dtype() {
-            DType::F32 => launch_gelu_f32(x.data_ptr() as _, out.data_ptr() as *mut c_void, n, std::ptr::null_mut()),
-            DType::BF16 => launch_gelu_bf16(x.data_ptr() as _, out.data_ptr() as *mut c_void, n, std::ptr::null_mut()),
-            _ => panic!("unsupported dtype for gelu"),
+            DType::F32 => launch_scale_f32(
+                x.data_ptr() as _,
+                out.data_ptr() as *mut c_void,
+                scale_val,
+                n,
+                xserv_cuda::current_stream_raw(),
+            ),
+            DType::BF16 => launch_scale_bf16(
+                x.data_ptr() as _,
+                out.data_ptr() as *mut c_void,
+                scale_val,
+                n,
+                xserv_cuda::current_stream_raw(),
+            ),
+            _ => panic!("unsupported dtype for scale"),
         }
     }
-    xserv_cuda::device::synchronize().unwrap();
     out
 }
 
-pub fn silu(x: &Tensor) -> Tensor {
-    assert!(x.is_contiguous());
+pub fn add(a: &Tensor, b: &Tensor) -> Tensor {
+    dispatch_binary(a, b, launch_add_f32, launch_add_bf16)
+}
+pub fn mul(a: &Tensor, b: &Tensor) -> Tensor {
+    dispatch_binary(a, b, launch_mul_f32, launch_mul_bf16)
+}
+
+/// Row-broadcast bias add: out[r, c] = x[r, c] + bias[c] (BF16 only).
+pub fn bias_add_2d(x: &Tensor, bias: &Tensor) -> Tensor {
+    assert_eq!(x.ndim(), 2);
+    assert_eq!(bias.ndim(), 1);
+    assert_eq!(x.dtype(), DType::BF16);
+    assert_eq!(bias.dtype(), DType::BF16);
+    assert!(x.is_contiguous() && bias.is_contiguous());
     assert!(matches!(x.device(), Device::Cuda(_)));
-    let out = Tensor::zeros(x.shape(), x.dtype(), x.device());
-    let n = x.numel() as i32;
+    let rows = x.shape()[0];
+    let cols = x.shape()[1];
+    assert_eq!(
+        bias.shape()[0],
+        cols,
+        "bias size {} != cols {cols}",
+        bias.shape()[0]
+    );
+    assert!(rows * cols <= i32::MAX as usize);
+    let out = Tensor::empty(&[rows, cols], DType::BF16, x.device());
     unsafe {
-        match x.dtype() {
-            DType::F32 => launch_silu_f32(x.data_ptr() as _, out.data_ptr() as *mut c_void, n, std::ptr::null_mut()),
-            DType::BF16 => launch_silu_bf16(x.data_ptr() as _, out.data_ptr() as *mut c_void, n, std::ptr::null_mut()),
-            _ => panic!("unsupported dtype for silu"),
-        }
+        launch_bias_add_2d_bf16(
+            x.data_ptr() as _,
+            bias.data_ptr() as _,
+            out.data_ptr() as *mut c_void,
+            rows as i32,
+            cols as i32,
+            xserv_cuda::current_stream_raw(),
+        );
+    }
+    out
+}
+
+/// Fused SiLU×Mul: out = silu(gate) * up (BF16 only)
+/// Saves one HBM read + one HBM write compared to separate silu + mul.
+pub fn silu_mul(gate: &Tensor, up: &Tensor) -> Tensor {
+    assert_eq!(gate.shape(), up.shape());
+    assert!(gate.is_contiguous() && up.is_contiguous());
+    assert!(matches!(gate.device(), Device::Cuda(_)));
+    assert_eq!(gate.dtype(), DType::BF16, "silu_mul requires BF16");
+    let out = Tensor::empty(gate.shape(), gate.dtype(), gate.device());
+    let n = gate.numel();
+    assert!(
+        n <= i32::MAX as usize,
+        "tensor too large for i32 kernel param ({n} elements)"
+    );
+    let n = n as i32;
+    unsafe {
+        launch_silu_mul_bf16(
+            gate.data_ptr() as *const c_void,
+            up.data_ptr() as *const c_void,
+            out.data_ptr() as *mut c_void,
+            n,
+            xserv_cuda::current_stream_raw(),
+        );
+    }
+    out
+}
+
+/// gpt-oss fused GLU activation (BF16 only).
+/// Input: gate_up [rows, 2*D] with interleaved columns (gate=even, up=odd).
+/// Output: [rows, D]
+/// Computes: gate.clamp(max=limit) * sigmoid(gate * alpha) * (up.clamp(-limit,limit) + 1)
+pub fn gpt_oss_glu(gate_up: &Tensor, alpha: f32, limit: f32) -> Tensor {
+    assert!(gate_up.is_contiguous());
+    assert!(matches!(gate_up.device(), Device::Cuda(_)));
+    assert_eq!(gate_up.dtype(), DType::BF16, "gpt_oss_glu requires BF16");
+    assert_eq!(gate_up.ndim(), 2);
+    let rows = gate_up.shape()[0];
+    let cols = gate_up.shape()[1];
+    assert_eq!(cols % 2, 0);
+    let d = cols / 2;
+    let out = Tensor::empty(&[rows, d], gate_up.dtype(), gate_up.device());
+    let n_elements = (rows * d) as i32;
+    unsafe {
+        launch_gpt_oss_glu_bf16(
+            gate_up.data_ptr() as *const c_void,
+            out.data_ptr() as *mut c_void,
+            n_elements,
+            alpha,
+            limit,
+            xserv_cuda::current_stream_raw(),
+        );
     }
-    xserv_cuda::device::synchronize().unwrap();
     out
 }

crates/xserv-kernels/src/argmax.rs

@@ -0,0 +1,72 @@
+use std::ffi::c_void;
+use xserv_tensor::{DType, Device, Tensor};
+
+unsafe extern "C" {
+    fn launch_argmax_bf16(
+        logits: *const c_void,
+        out_idx: *mut c_void,
+        rows: i32,
+        cols: i32,
+        stream: *mut c_void,
+    );
+}
+
+/// GPU argmax over the last dim of a [rows, cols] BF16 tensor.
+///
+/// Returns a host `Vec<u32>` of length `rows`. Internally:
+///   - launches one kernel that writes [rows] i32 indices on device
+///   - D2H copies just `rows * 4` bytes (vs `rows * cols * 2` for the
+///     "copy logits to CPU then argmax" path it replaces)
+///
+/// This is the greedy-decode hot path: avoids touching the full
+/// [B, vocab] logits buffer on the host every step.
+pub fn argmax_bf16_to_host(logits: &Tensor) -> Vec<u32> {
+    assert_eq!(logits.ndim(), 2, "argmax expects a 2D [rows, cols] tensor");
+    assert_eq!(logits.dtype(), DType::BF16, "argmax kernel is BF16-only");
+    assert!(logits.is_contiguous(), "argmax requires contiguous input");
+    assert!(
+        matches!(logits.device(), Device::Cuda(_)),
+        "argmax requires GPU input"
+    );
+
+    let rows = logits.shape()[0];
+    let cols = logits.shape()[1];
+    assert!(rows <= i32::MAX as usize);
+    assert!(cols <= i32::MAX as usize);
+
+    // Output buffer: rows * i32. Pooled allocator so this is essentially free
+    // after the first call.
+    let bytes = rows * std::mem::size_of::<i32>();
+    let mut out = xserv_cuda::allocator::cached_alloc(bytes).expect("argmax out alloc");
+
+    unsafe {
+        launch_argmax_bf16(
+            logits.data_ptr() as *const c_void,
+            out.as_mut_ptr() as *mut c_void,
+            rows as i32,
+            cols as i32,
+            xserv_cuda::current_stream_raw(),
+        );
+    }
+
+    let mut host_bytes = vec![0u8; bytes];
+    out.copy_to_host(&mut host_bytes).expect("argmax D2H");
+    drop(out); // returned to pool
+
+    let host_i32: &[i32] =
+        unsafe { std::slice::from_raw_parts(host_bytes.as_ptr() as *const i32, rows) };
+    host_i32.iter().map(|&v| v as u32).collect()
+}
+
+/// Convenience: argmax of a single row [1, cols] (or [cols] reshaped to [1, cols]).
+pub fn argmax_bf16_single(logits: &Tensor) -> u32 {
+    let cols = *logits.shape().last().unwrap();
+    let rows = logits.numel() / cols;
+    assert_eq!(rows, 1, "argmax_bf16_single requires a single row");
+    let view = if logits.ndim() == 2 {
+        logits.clone()
+    } else {
+        logits.reshape(&[1, cols])
+    };
+    argmax_bf16_to_host(&view)[0]
+}

crates/xserv-kernels/src/attention.rs

@@ -0,0 +1,682 @@
+use std::ffi::c_void;
+use xserv_tensor::{DType, Tensor};
+
+use crate::activation::scale;
+use crate::gemm::batched_matmul;
+use crate::softmax::softmax;
+
+unsafe extern "C" {
+    fn launch_causal_mask_f32(
+        scores: *mut c_void,
+        batch: i32,
+        rows: i32,
+        cols: i32,
+        offset: i32,
+        stream: *mut c_void,
+    );
+    fn launch_causal_mask_bf16(
+        scores: *mut c_void,
+        batch: i32,
+        rows: i32,
+        cols: i32,
+        offset: i32,
+        stream: *mut c_void,
+    );
+    fn launch_flash_attention_bf16(
+        q: *const c_void,
+        k: *const c_void,
+        v: *const c_void,
+        o: *mut c_void,
+        batch: i32,
+        num_q_heads: i32,
+        num_kv_heads: i32,
+        q_len: i32,
+        kv_len: i32,
+        head_dim: i32,
+        scale: f32,
+        causal: i32,
+        stream: *mut c_void,
+    );
+    fn launch_flash_attention_sinks_bf16(
+        q: *const c_void,
+        k: *const c_void,
+        v: *const c_void,
+        o: *mut c_void,
+        sinks: *const c_void,
+        batch: i32,
+        num_q_heads: i32,
+        num_kv_heads: i32,
+        q_len: i32,
+        kv_len: i32,
+        head_dim: i32,
+        scale: f32,
+        causal: i32,
+        window_size: i32,
+        stream: *mut c_void,
+    );
+    fn launch_decode_attention_bf16(
+        q: *const c_void,
+        k: *const c_void,
+        v: *const c_void,
+        o: *mut c_void,
+        batch: i32,
+        num_q_heads: i32,
+        num_kv_heads: i32,
+        kv_len: i32,
+        head_dim: i32,
+        scale: f32,
+        causal: i32,
+        stream: *mut c_void,
+    );
+    fn launch_paged_decode_attention_bf16(
+        q: *const c_void,
+        k_cache: *const c_void,
+        v_cache: *const c_void,
+        o: *mut c_void,
+        block_tables: *const i32,
+        context_lens: *const i32,
+        batch: i32,
+        num_q_heads: i32,
+        num_kv_heads: i32,
+        head_dim: i32,
+        max_blocks_per_seq: i32,
+        scale: f32,
+        stream: *mut c_void,
+    );
+    fn launch_paged_decode_attention_tree_bf16(
+        q: *const c_void,
+        k_cache: *const c_void,
+        v_cache: *const c_void,
+        o: *mut c_void,
+        block_tables: *const i32,
+        context_lens: *const i32,
+        tree_mask: *const i32,
+        batch: i32,
+        num_q_heads: i32,
+        num_kv_heads: i32,
+        head_dim: i32,
+        max_blocks_per_seq: i32,
+        tree_start: i32,
+        tree_len: i32,
+        scale: f32,
+        stream: *mut c_void,
+    );
+    fn launch_paged_decode_attention_sinks_bf16(
+        q: *const c_void,
+        k_cache: *const c_void,
+        v_cache: *const c_void,
+        o: *mut c_void,
+        block_tables: *const i32,
+        context_lens: *const i32,
+        sinks: *const c_void,
+        batch: i32,
+        num_q_heads: i32,
+        num_kv_heads: i32,
+        head_dim: i32,
+        max_blocks_per_seq: i32,
+        scale: f32,
+        window_size: i32,
+        stream: *mut c_void,
+    );
+    fn launch_reshape_and_cache_bf16(
+        k_src: *const c_void,
+        v_src: *const c_void,
+        k_pool: *mut c_void,
+        v_pool: *mut c_void,
+        block_ids: *const c_void,
+        num_tokens: i32,
+        num_heads: i32,
+        head_dim: i32,
+        start_pos: i32,
+        block_size: i32,
+        stream: *mut c_void,
+    );
+    fn launch_reshape_and_cache_batched_bf16(
+        k_src: *const c_void,
+        v_src: *const c_void,
+        k_pool: *mut c_void,
+        v_pool: *mut c_void,
+        block_tables: *const c_void,
+        kv_lens: *const c_void,
+        batch: i32,
+        num_heads: i32,
+        head_dim: i32,
+        block_size: i32,
+        max_blocks_per_seq: i32,
+        stream: *mut c_void,
+    );
+    fn launch_copy_kv_position(
+        k_pool: *mut c_void,
+        v_pool: *mut c_void,
+        block_ids: *const i32,
+        src_pos: i32,
+        dst_pos: i32,
+        num_kv_heads: i32,
+        head_dim: i32,
+        block_size: i32,
+        stream: *mut c_void,
+    );
+}
+
+/// Scatter `[num_kv_heads, num_tokens, head_dim]` BF16 K/V into a paged
+/// pool for a single sequence whose block table lives at `block_ids_gpu`
+/// (int32, on device).
+///
+/// `k_pool_ptr`/`v_pool_ptr` point to one layer's pool, of logical shape
+/// `[num_blocks_total, num_kv_heads, block_size, head_dim]`.
+///
+/// All pointers must be on the same GPU as the launching context.
+///
+/// # Safety
+/// Pointers must be valid GPU pointers with the documented layouts.
+/// `block_ids_gpu` must contain at least `(start_pos + num_tokens + block_size - 1) / block_size`
+/// valid physical block ids.
+pub unsafe fn reshape_and_cache_bf16(
+    k_src: *const c_void,
+    v_src: *const c_void,
+    k_pool_ptr: *mut c_void,
+    v_pool_ptr: *mut c_void,
+    block_ids_gpu: *const i32,
+    num_tokens: usize,
+    num_heads: usize,
+    head_dim: usize,
+    start_pos: usize,
+    block_size: usize,
+    stream: *mut c_void,
+) {
+    unsafe {
+        launch_reshape_and_cache_bf16(
+            k_src,
+            v_src,
+            k_pool_ptr,
+            v_pool_ptr,
+            block_ids_gpu as *const c_void,
+            num_tokens as i32,
+            num_heads as i32,
+            head_dim as i32,
+            start_pos as i32,
+            block_size as i32,
+            stream,
+        );
+    }
+}
+
+/// Batched scatter for the multi-sequence decode step. Reads
+/// `block_tables` (`[batch, max_blocks_per_seq]` int32 — same buffer the
+/// paged-attention kernel reads) and `kv_lens` (`[batch]` int32, current
+/// seq_len + 1 — i.e., the index of the just-written token + 1) so the
+/// caller doesn't need a separate per-step upload of block ids.
+///
+/// # Safety
+/// All pointers must be on the same GPU. `block_tables` and `kv_lens` must
+/// already be synced to the device for the active batch.
+pub unsafe fn reshape_and_cache_batched_bf16(
+    k_src: *const c_void,
+    v_src: *const c_void,
+    k_pool_ptr: *mut c_void,
+    v_pool_ptr: *mut c_void,
+    block_tables_gpu: *const i32,
+    kv_lens_gpu: *const i32,
+    batch: usize,
+    num_heads: usize,
+    head_dim: usize,
+    block_size: usize,
+    max_blocks_per_seq: usize,
+    stream: *mut c_void,
+) {
+    unsafe {
+        launch_reshape_and_cache_batched_bf16(
+            k_src,
+            v_src,
+            k_pool_ptr,
+            v_pool_ptr,
+            block_tables_gpu as *const c_void,
+            kv_lens_gpu as *const c_void,
+            batch as i32,
+            num_heads as i32,
+            head_dim as i32,
+            block_size as i32,
+            max_blocks_per_seq as i32,
+            stream,
+        );
+    }
+}
+
+/// Copy one token's K/V from `src_pos` to `dst_pos` within the same sequence's
+/// paged cache (one layer). Used by tree speculative decoding to remap
+/// accepted sibling K/V to canonical sequential positions after acceptance.
+///
+/// # Safety
+/// Pool and block_ids pointers must be valid GPU pointers for the given layer.
+pub unsafe fn copy_kv_position(
+    k_pool_ptr: *mut c_void,
+    v_pool_ptr: *mut c_void,
+    block_ids_gpu: *const i32,
+    src_pos: usize,
+    dst_pos: usize,
+    num_kv_heads: usize,
+    head_dim: usize,
+    block_size: usize,
+    stream: *mut c_void,
+) {
+    launch_copy_kv_position(
+        k_pool_ptr,
+        v_pool_ptr,
+        block_ids_gpu,
+        src_pos as i32,
+        dst_pos as i32,
+        num_kv_heads as i32,
+        head_dim as i32,
+        block_size as i32,
+        stream,
+    );
+}
+
+fn apply_causal_mask(scores: &Tensor, offset: usize) {
+    let ndim = scores.ndim();
+    let rows = scores.shape()[ndim - 2];
+    let cols = scores.shape()[ndim - 1];
+    let batch: usize = scores.shape()[..ndim - 2].iter().product();
+
+    unsafe {
+        match scores.dtype() {
+            DType::F32 => launch_causal_mask_f32(
+                scores.data_ptr() as *mut c_void,
+                batch as i32,
+                rows as i32,
+                cols as i32,
+                offset as i32,
+                xserv_cuda::current_stream_raw(),
+            ),
+            DType::BF16 => launch_causal_mask_bf16(
+                scores.data_ptr() as *mut c_void,
+                batch as i32,
+                rows as i32,
+                cols as i32,
+                offset as i32,
+                xserv_cuda::current_stream_raw(),
+            ),
+            _ => panic!("unsupported dtype for causal mask"),
+        }
+    }
+}
+
+/// Multi-head attention (naive, materializes S×S score matrix).
+///
+/// q, k, v: [batch, num_heads, seq_len, head_dim] — contiguous, on GPU
+/// Returns: [batch, num_heads, seq_len, head_dim]
+pub fn attention(q: &Tensor, k: &Tensor, v: &Tensor, causal: bool) -> Tensor {
+    assert_eq!(q.ndim(), 4);
+    assert_eq!(k.ndim(), 4);
+    assert_eq!(v.ndim(), 4);
+    assert!(q.is_contiguous() && k.is_contiguous() && v.is_contiguous());
+
+    let batch = q.shape()[0];
+    let num_heads = q.shape()[1];
+    let q_len = q.shape()[2];
+    let head_dim = q.shape()[3];
+    let kv_len = k.shape()[2];
+
+    assert_eq!(k.shape(), &[batch, num_heads, kv_len, head_dim]);
+    assert_eq!(v.shape(), &[batch, num_heads, kv_len, head_dim]);
+
+    // scores = Q @ K^T → [B, H, q_len, kv_len]
+    let k_t = k.transpose(2, 3).contiguous();
+    let scores = batched_matmul(q, &k_t);
+
+    // Scale by 1/sqrt(head_dim)
+    let scale_factor = 1.0 / (head_dim as f32).sqrt();
+    let scaled_scores = scale(&scores, scale_factor);
+
+    // Causal mask
+    if causal {
+        let offset = kv_len - q_len;
+        apply_causal_mask(&scaled_scores, offset);
+    }
+
+    // Softmax
+    let weights = softmax(&scaled_scores);
+
+    // output = weights @ V → [B, H, q_len, head_dim]
+    batched_matmul(&weights, v)
+}
+
+/// Decode Attention — optimized for single-token decode (q_len=1).
+///
+/// q: [batch, num_q_heads, 1, head_dim]       BF16, contiguous, GPU
+/// k: [batch, num_kv_heads, kv_len, head_dim] BF16, contiguous, GPU
+/// v: [batch, num_kv_heads, kv_len, head_dim] BF16, contiguous, GPU
+///
+/// Returns: [batch, num_q_heads, 1, head_dim] BF16
+pub fn decode_attention(q: &Tensor, k: &Tensor, v: &Tensor) -> Tensor {
+    assert_eq!(q.ndim(), 4);
+    assert_eq!(q.shape()[2], 1, "decode_attention requires q_len == 1");
+
+    let batch = q.shape()[0];
+    let num_q_heads = q.shape()[1];
+    let head_dim = q.shape()[3];
+    let num_kv_heads = k.shape()[1];
+    let kv_len = k.shape()[2];
+
+    let scale = 1.0 / (head_dim as f32).sqrt();
+    let output = Tensor::empty(&[batch, num_q_heads, 1, head_dim], DType::BF16, q.device());
+
+    unsafe {
+        launch_decode_attention_bf16(
+            q.data_ptr() as *const c_void,
+            k.data_ptr() as *const c_void,
+            v.data_ptr() as *const c_void,
+            output.data_ptr() as *mut c_void,
+            batch as i32,
+            num_q_heads as i32,
+            num_kv_heads as i32,
+            kv_len as i32,
+            head_dim as i32,
+            scale,
+            1, // causal (always 1 for decode)
+            xserv_cuda::current_stream_raw(),
+        );
+    }
+
+    output
+}
+
+/// Flash Attention 2 — O(1) extra memory, supports GQA natively.
+/// Auto-dispatches to decode_attention when q_len == 1.
+///
+/// q: [batch, num_q_heads, q_len, head_dim]   BF16, contiguous, GPU
+/// k: [batch, num_kv_heads, kv_len, head_dim] BF16, contiguous, GPU
+/// v: [batch, num_kv_heads, kv_len, head_dim] BF16, contiguous, GPU
+///
+/// Returns: [batch, num_q_heads, q_len, head_dim] BF16
+pub fn flash_attention(q: &Tensor, k: &Tensor, v: &Tensor, causal: bool) -> Tensor {
+    assert_eq!(q.ndim(), 4);
+    assert_eq!(k.ndim(), 4);
+    assert_eq!(v.ndim(), 4);
+    assert!(q.is_contiguous() && k.is_contiguous() && v.is_contiguous());
+    assert_eq!(q.dtype(), DType::BF16, "flash_attention requires BF16");
+    assert_eq!(k.dtype(), DType::BF16);
+    assert_eq!(v.dtype(), DType::BF16);
+
+    let batch = q.shape()[0];
+    let num_q_heads = q.shape()[1];
+    let q_len = q.shape()[2];
+    let head_dim = q.shape()[3];
+    let num_kv_heads = k.shape()[1];
+    let kv_len = k.shape()[2];
+
+    assert_eq!(k.shape(), &[batch, num_kv_heads, kv_len, head_dim]);
+    assert_eq!(v.shape(), &[batch, num_kv_heads, kv_len, head_dim]);
+    assert!(
+        num_q_heads % num_kv_heads == 0,
+        "num_q_heads must be divisible by num_kv_heads"
+    );
+    assert!(
+        head_dim <= 128,
+        "flash_attention supports head_dim up to 128"
+    );
+
+    // Dispatch to specialized decode kernel for single-token generation
+    if q_len == 1 {
+        return decode_attention(q, k, v);
+    }
+
+    let scale = 1.0 / (head_dim as f32).sqrt();
+    let output = Tensor::empty(
+        &[batch, num_q_heads, q_len, head_dim],
+        DType::BF16,
+        q.device(),
+    );
+
+    unsafe {
+        launch_flash_attention_bf16(
+            q.data_ptr() as *const c_void,
+            k.data_ptr() as *const c_void,
+            v.data_ptr() as *const c_void,
+            output.data_ptr() as *mut c_void,
+            batch as i32,
+            num_q_heads as i32,
+            num_kv_heads as i32,
+            q_len as i32,
+            kv_len as i32,
+            head_dim as i32,
+            scale,
+            if causal { 1 } else { 0 },
+            xserv_cuda::current_stream_raw(),
+        );
+    }
+
+    output
+}
+
+/// Flash attention for prefill with gpt-oss attention sinks + optional sliding window.
+///
+/// Same layout/contract as `flash_attention`, plus a per-head `sinks` tensor
+/// ([num_q_heads] BF16, GPU) folded into the softmax denominator, and a
+/// `window_size` (0 = full causal, >0 = sliding window). Always causal.
+pub fn flash_attention_sinks(
+    q: &Tensor,
+    k: &Tensor,
+    v: &Tensor,
+    sinks: &Tensor,
+    window_size: usize,
+) -> Tensor {
+    assert_eq!(q.ndim(), 4);
+    assert_eq!(k.ndim(), 4);
+    assert_eq!(v.ndim(), 4);
+    assert!(q.is_contiguous() && k.is_contiguous() && v.is_contiguous());
+    assert_eq!(q.dtype(), DType::BF16);
+    assert_eq!(k.dtype(), DType::BF16);
+    assert_eq!(v.dtype(), DType::BF16);
+
+    let batch = q.shape()[0];
+    let num_q_heads = q.shape()[1];
+    let q_len = q.shape()[2];
+    let head_dim = q.shape()[3];
+    let num_kv_heads = k.shape()[1];
+    let kv_len = k.shape()[2];
+
+    assert_eq!(k.shape(), &[batch, num_kv_heads, kv_len, head_dim]);
+    assert_eq!(v.shape(), &[batch, num_kv_heads, kv_len, head_dim]);
+    assert!(num_q_heads % num_kv_heads == 0);
+    assert!(head_dim <= 128);
+    assert_eq!(
+        sinks.shape()[0],
+        num_q_heads,
+        "sinks must have num_q_heads entries"
+    );
+
+    let scale = 1.0 / (head_dim as f32).sqrt();
+    let output = Tensor::empty(
+        &[batch, num_q_heads, q_len, head_dim],
+        DType::BF16,
+        q.device(),
+    );
+
+    unsafe {
+        launch_flash_attention_sinks_bf16(
+            q.data_ptr() as *const c_void,
+            k.data_ptr() as *const c_void,
+            v.data_ptr() as *const c_void,
+            output.data_ptr() as *mut c_void,
+            sinks.data_ptr() as *const c_void,
+            batch as i32,
+            num_q_heads as i32,
+            num_kv_heads as i32,
+            q_len as i32,
+            kv_len as i32,
+            head_dim as i32,
+            scale,
+            1, // always causal
+            window_size as i32,
+            xserv_cuda::current_stream_raw(),
+        );
+    }
+
+    output
+}
+
+/// Paged decode attention.
+///
+/// q: [batch, num_q_heads, 1, head_dim] BF16, contiguous, GPU
+/// k_cache_ptr / v_cache_ptr: pointers to [num_blocks, num_kv_heads, BLOCK_SIZE, head_dim] BF16 pools
+/// block_tables_ptr: i32 [batch, max_blocks_per_seq] (rows already arranged for this batch)
+/// context_lens_ptr: i32 [batch]
+///
+/// Returns: [batch, num_q_heads, 1, head_dim] BF16
+#[allow(clippy::too_many_arguments)]
+pub fn paged_decode_attention(
+    q: &Tensor,
+    k_cache_ptr: *const c_void,
+    v_cache_ptr: *const c_void,
+    block_tables_ptr: *const i32,
+    context_lens_ptr: *const i32,
+    batch: usize,
+    num_q_heads: usize,
+    num_kv_heads: usize,
+    head_dim: usize,
+    max_blocks_per_seq: usize,
+) -> Tensor {
+    assert_eq!(q.ndim(), 4);
+    assert_eq!(
+        q.shape()[2],
+        1,
+        "paged_decode_attention requires q_len == 1"
+    );
+    assert_eq!(q.dtype(), DType::BF16);
+    assert!(
+        num_q_heads % num_kv_heads == 0,
+        "GQA: num_q_heads must be divisible by num_kv_heads"
+    );
+    assert!(head_dim <= 128);
+
+    let scale = 1.0 / (head_dim as f32).sqrt();
+    let output = Tensor::empty(&[batch, num_q_heads, 1, head_dim], DType::BF16, q.device());
+
+    unsafe {
+        launch_paged_decode_attention_bf16(
+            q.data_ptr() as *const c_void,
+            k_cache_ptr,
+            v_cache_ptr,
+            output.data_ptr() as *mut c_void,
+            block_tables_ptr,
+            context_lens_ptr,
+            batch as i32,
+            num_q_heads as i32,
+            num_kv_heads as i32,
+            head_dim as i32,
+            max_blocks_per_seq as i32,
+            scale,
+            xserv_cuda::current_stream_raw(),
+        );
+    }
+
+    output
+}
+
+/// Tree-aware paged decode attention. Adds a per-query attention mask over
+/// the newly-written K/V region `[tree_start, tree_start+tree_len)`. Query i
+/// attends to position tree_start+j iff tree_mask[i, j] != 0. Positions <
+/// tree_start are always attended.
+///
+/// Used by speculative decoding with tree drafting to let sibling candidates
+/// share position slots without seeing each other's K/V.
+#[allow(clippy::too_many_arguments)]
+pub fn paged_decode_attention_tree(
+    q: &Tensor,
+    k_cache_ptr: *const c_void,
+    v_cache_ptr: *const c_void,
+    block_tables_ptr: *const i32,
+    context_lens_ptr: *const i32,
+    tree_mask_ptr: *const i32,
+    batch: usize,
+    num_q_heads: usize,
+    num_kv_heads: usize,
+    head_dim: usize,
+    max_blocks_per_seq: usize,
+    tree_start: usize,
+    tree_len: usize,
+) -> Tensor {
+    assert_eq!(q.ndim(), 4);
+    assert_eq!(q.shape()[2], 1);
+    assert_eq!(q.dtype(), DType::BF16);
+    assert!(num_q_heads % num_kv_heads == 0);
+    assert!(head_dim <= 128);
+
+    let scale = 1.0 / (head_dim as f32).sqrt();
+    let output = Tensor::empty(&[batch, num_q_heads, 1, head_dim], DType::BF16, q.device());
+
+    unsafe {
+        launch_paged_decode_attention_tree_bf16(
+            q.data_ptr() as *const c_void,
+            k_cache_ptr,
+            v_cache_ptr,
+            output.data_ptr() as *mut c_void,
+            block_tables_ptr,
+            context_lens_ptr,
+            tree_mask_ptr,
+            batch as i32,
+            num_q_heads as i32,
+            num_kv_heads as i32,
+            head_dim as i32,
+            max_blocks_per_seq as i32,
+            tree_start as i32,
+            tree_len as i32,
+            scale,
+            xserv_cuda::current_stream_raw(),
+        );
+    }
+
+    output
+}
+
+/// Paged decode attention with attention sinks and optional sliding window.
+///
+/// sinks_ptr: pointer to [num_q_heads] BF16 on GPU (or null for no sinks)
+/// window_size: 0 = full attention, >0 = sliding window
+#[allow(clippy::too_many_arguments)]
+pub fn paged_decode_attention_sinks(
+    q: &Tensor,
+    k_cache_ptr: *const c_void,
+    v_cache_ptr: *const c_void,
+    block_tables_ptr: *const i32,
+    context_lens_ptr: *const i32,
+    sinks_ptr: *const c_void,
+    batch: usize,
+    num_q_heads: usize,
+    num_kv_heads: usize,
+    head_dim: usize,
+    max_blocks_per_seq: usize,
+    window_size: usize,
+) -> Tensor {
+    assert_eq!(q.ndim(), 4);
+    assert_eq!(q.shape()[2], 1);
+    assert_eq!(q.dtype(), DType::BF16);
+    assert!(num_q_heads % num_kv_heads == 0);
+    assert!(head_dim <= 128);
+
+    let scale = 1.0 / (head_dim as f32).sqrt();
+    let output = Tensor::empty(&[batch, num_q_heads, 1, head_dim], DType::BF16, q.device());
+
+    unsafe {
+        launch_paged_decode_attention_sinks_bf16(
+            q.data_ptr() as *const c_void,
+            k_cache_ptr,
+            v_cache_ptr,
+            output.data_ptr() as *mut c_void,
+            block_tables_ptr,
+            context_lens_ptr,
+            sinks_ptr,
+            batch as i32,
+            num_q_heads as i32,
+            num_kv_heads as i32,
+            head_dim as i32,
+            max_blocks_per_seq as i32,
+            scale,
+            window_size as i32,
+            xserv_cuda::current_stream_raw(),
+        );
+    }
+
+    output
+}

crates/xserv-kernels/src/dispatch.rs

@@ -0,0 +1,316 @@
+//! Low-level kernel dispatchers for CUDA Graph capture.
+//! These functions write to pre-allocated output buffers and accept an explicit stream.
+
+use std::ffi::c_void;
+
+// Re-declare the extern functions we need (same as in the individual modules)
+unsafe extern "C" {
+    fn launch_rmsnorm_bf16(
+        x: *const c_void,
+        gamma: *const c_void,
+        out: *mut c_void,
+        rows: i32,
+        hidden_size: i32,
+        eps: f32,
+        stream: *mut c_void,
+    );
+    fn launch_add_rmsnorm_bf16(
+        x: *const c_void,
+        residual: *const c_void,
+        gamma: *const c_void,
+        normed_out: *mut c_void,
+        sum_out: *mut c_void,
+        rows: i32,
+        hidden_size: i32,
+        eps: f32,
+        stream: *mut c_void,
+    );
+    fn launch_silu_mul_bf16(
+        gate: *const c_void,
+        up: *const c_void,
+        out: *mut c_void,
+        n: i32,
+        stream: *mut c_void,
+    );
+    fn launch_add_bf16(
+        a: *const c_void,
+        b: *const c_void,
+        out: *mut c_void,
+        n: i32,
+        stream: *mut c_void,
+    );
+    fn launch_embedding_bf16(
+        table: *const c_void,
+        token_ids: *const c_void,
+        out: *mut c_void,
+        num_tokens: i32,
+        hidden_size: i32,
+        vocab_size: i32,
+        stream: *mut c_void,
+    );
+    fn launch_reshape_heads_bf16(
+        inp: *const c_void,
+        out: *mut c_void,
+        seq_len: i32,
+        num_heads: i32,
+        head_dim: i32,
+        stream: *mut c_void,
+    );
+    fn launch_merge_heads_bf16(
+        inp: *const c_void,
+        out: *mut c_void,
+        seq_len: i32,
+        num_heads: i32,
+        head_dim: i32,
+        stream: *mut c_void,
+    );
+    fn launch_transpose_hsd_to_shd_bf16(
+        inp: *const c_void,
+        out: *mut c_void,
+        seq_len: i32,
+        num_heads: i32,
+        head_dim: i32,
+        stream: *mut c_void,
+    );
+    fn launch_transpose_shd_to_hsd_bf16(
+        inp: *const c_void,
+        out: *mut c_void,
+        seq_len: i32,
+        num_heads: i32,
+        head_dim: i32,
+        stream: *mut c_void,
+    );
+    fn launch_rope_bf16(
+        x: *mut c_void,
+        cos_cache: *const c_void,
+        sin_cache: *const c_void,
+        positions: *const c_void,
+        num_tokens: i32,
+        num_heads: i32,
+        head_dim: i32,
+        stream: *mut c_void,
+    );
+    fn launch_gemv_bf16(
+        x: *const c_void,
+        w: *const c_void,
+        y_bf16: *mut c_void,
+        y_fp32_buf: *mut c_void,
+        k: i32,
+        n: i32,
+        stream: *mut c_void,
+    );
+    fn launch_decode_attention_bf16(
+        q: *const c_void,
+        k: *const c_void,
+        v: *const c_void,
+        o: *mut c_void,
+        batch: i32,
+        num_q_heads: i32,
+        num_kv_heads: i32,
+        kv_len: i32,
+        head_dim: i32,
+        scale: f32,
+        causal: i32,
+        stream: *mut c_void,
+    );
+}
+
+/// Raw rmsnorm dispatch: writes to pre-allocated `out`.
+pub unsafe fn rmsnorm_bf16(
+    x: *const c_void,
+    gamma: *const c_void,
+    out: *mut c_void,
+    rows: i32,
+    hidden_size: i32,
+    eps: f32,
+    stream: *mut c_void,
+) {
+    launch_rmsnorm_bf16(x, gamma, out, rows, hidden_size, eps, stream);
+}
+
+/// Raw add_rmsnorm dispatch.
+pub unsafe fn add_rmsnorm_bf16(
+    x: *const c_void,
+    residual: *const c_void,
+    gamma: *const c_void,
+    normed_out: *mut c_void,
+    sum_out: *mut c_void,
+    rows: i32,
+    hidden_size: i32,
+    eps: f32,
+    stream: *mut c_void,
+) {
+    launch_add_rmsnorm_bf16(
+        x,
+        residual,
+        gamma,
+        normed_out,
+        sum_out,
+        rows,
+        hidden_size,
+        eps,
+        stream,
+    );
+}
+
+/// Raw silu_mul dispatch.
+pub unsafe fn silu_mul_bf16(
+    gate: *const c_void,
+    up: *const c_void,
+    out: *mut c_void,
+    n: i32,
+    stream: *mut c_void,
+) {
+    launch_silu_mul_bf16(gate, up, out, n, stream);
+}
+
+/// Raw add dispatch.
+pub unsafe fn add_bf16(
+    a: *const c_void,
+    b: *const c_void,
+    out: *mut c_void,
+    n: i32,
+    stream: *mut c_void,
+) {
+    launch_add_bf16(a, b, out, n, stream);
+}
+
+/// Raw embedding dispatch.
+pub unsafe fn embedding_bf16(
+    table: *const c_void,
+    token_ids: *const c_void,
+    out: *mut c_void,
+    num_tokens: i32,
+    hidden_size: i32,
+    vocab_size: i32,
+    stream: *mut c_void,
+) {
+    launch_embedding_bf16(
+        table,
+        token_ids,
+        out,
+        num_tokens,
+        hidden_size,
+        vocab_size,
+        stream,
+    );
+}
+
+/// Raw reshape_heads dispatch.
+pub unsafe fn reshape_heads_bf16(
+    inp: *const c_void,
+    out: *mut c_void,
+    seq_len: i32,
+    num_heads: i32,
+    head_dim: i32,
+    stream: *mut c_void,
+) {
+    launch_reshape_heads_bf16(inp, out, seq_len, num_heads, head_dim, stream);
+}
+
+/// Raw merge_heads dispatch.
+pub unsafe fn merge_heads_bf16(
+    inp: *const c_void,
+    out: *mut c_void,
+    seq_len: i32,
+    num_heads: i32,
+    head_dim: i32,
+    stream: *mut c_void,
+) {
+    launch_merge_heads_bf16(inp, out, seq_len, num_heads, head_dim, stream);
+}
+
+/// Raw transpose HSD->SHD dispatch.
+pub unsafe fn transpose_hsd_to_shd_bf16(
+    inp: *const c_void,
+    out: *mut c_void,
+    seq_len: i32,
+    num_heads: i32,
+    head_dim: i32,
+    stream: *mut c_void,
+) {
+    launch_transpose_hsd_to_shd_bf16(inp, out, seq_len, num_heads, head_dim, stream);
+}
+
+/// Raw transpose SHD->HSD dispatch.
+pub unsafe fn transpose_shd_to_hsd_bf16(
+    inp: *const c_void,
+    out: *mut c_void,
+    seq_len: i32,
+    num_heads: i32,
+    head_dim: i32,
+    stream: *mut c_void,
+) {
+    launch_transpose_shd_to_hsd_bf16(inp, out, seq_len, num_heads, head_dim, stream);
+}
+
+/// Raw RoPE dispatch (in-place).
+pub unsafe fn rope_bf16(
+    x: *mut c_void,
+    cos_cache: *const c_void,
+    sin_cache: *const c_void,
+    positions: *const c_void,
+    num_tokens: i32,
+    num_heads: i32,
+    head_dim: i32,
+    stream: *mut c_void,
+) {
+    launch_rope_bf16(
+        x, cos_cache, sin_cache, positions, num_tokens, num_heads, head_dim, stream,
+    );
+}
+
+/// Raw GEMV dispatch (BF16, M=1). Caller must provide fp32 accumulator buffer.
+pub unsafe fn gemv_bf16(
+    x: *const c_void,
+    w: *const c_void,
+    y_bf16: *mut c_void,
+    y_fp32_buf: *mut c_void,
+    k: i32,
+    n: i32,
+    stream: *mut c_void,
+) {
+    launch_gemv_bf16(x, w, y_bf16, y_fp32_buf, k, n, stream);
+}
+
+/// Raw decode attention dispatch.
+pub unsafe fn decode_attention_bf16(
+    q: *const c_void,
+    k: *const c_void,
+    v: *const c_void,
+    o: *mut c_void,
+    batch: i32,
+    num_q_heads: i32,
+    num_kv_heads: i32,
+    kv_len: i32,
+    head_dim: i32,
+    scale: f32,
+    stream: *mut c_void,
+) {
+    launch_decode_attention_bf16(
+        q,
+        k,
+        v,
+        o,
+        batch,
+        num_q_heads,
+        num_kv_heads,
+        kv_len,
+        head_dim,
+        scale,
+        1,
+        stream,
+    );
+}
+
+// cuBLAS FFI
+pub type CublasHandle = *mut c_void;
+
+unsafe extern "C" {
+    fn cublasSetStream_v2(handle: CublasHandle, stream: *mut c_void) -> i32;
+}
+
+/// Set cuBLAS stream. Must be called before any cuBLAS operations during graph capture.
+pub unsafe fn set_cublas_stream(handle: CublasHandle, stream: *mut c_void) {
+    cublasSetStream_v2(handle, stream);
+}

crates/xserv-kernels/src/embedding.rs

@@ -1,12 +1,25 @@
 use std::ffi::c_void;
-use xserv_cuda::GpuBuffer;
 use xserv_tensor::{DType, Device, Tensor};
 
 unsafe extern "C" {
-    fn launch_embedding_f32(table: *const c_void, token_ids: *const c_void, out: *mut c_void,
-                            num_tokens: i32, hidden_size: i32, stream: *mut c_void);
-    fn launch_embedding_bf16(table: *const c_void, token_ids: *const c_void, out: *mut c_void,
-                             num_tokens: i32, hidden_size: i32, stream: *mut c_void);
+    fn launch_embedding_f32(
+        table: *const c_void,
+        token_ids: *const c_void,
+        out: *mut c_void,
+        num_tokens: i32,
+        hidden_size: i32,
+        vocab_size: i32,
+        stream: *mut c_void,
+    );
+    fn launch_embedding_bf16(
+        table: *const c_void,
+        token_ids: *const c_void,
+        out: *mut c_void,
+        num_tokens: i32,
+        hidden_size: i32,
+        vocab_size: i32,
+        stream: *mut c_void,
+    );
 }
 
 /// Embedding lookup: table[token_ids[i]] for each i.
@@ -18,6 +31,15 @@ pub fn embedding(table: &Tensor, token_ids: &[u32]) -> Tensor {
 
     let hidden_size = table.shape()[1];
     let num_tokens = token_ids.len();
+    let vocab_size = table.shape()[0];
+    assert!(
+        num_tokens <= i32::MAX as usize,
+        "too many tokens for i32 kernel param"
+    );
+    assert!(
+        hidden_size <= i32::MAX as usize,
+        "hidden_size too large for i32 kernel param"
+    );
 
     // Upload token_ids to GPU
     let ids_bytes = unsafe {
@@ -26,26 +48,54 @@ pub fn embedding(table: &Tensor, token_ids: &[u32]) -> Tensor {
             num_tokens * std::mem::size_of::<u32>(),
         )
     };
-    let mut ids_gpu = GpuBuffer::alloc(ids_bytes.len()).expect("alloc token_ids");
+    let mut ids_gpu =
+        xserv_cuda::allocator::cached_alloc(ids_bytes.len()).expect("alloc token_ids");
     ids_gpu.copy_from_host(ids_bytes).unwrap();
 
-    let out = Tensor::zeros(&[num_tokens, hidden_size], table.dtype(), table.device());
+    for &tid in token_ids {
+        assert!(
+            (tid as usize) < vocab_size,
+            "token_id {tid} out of bounds (vocab_size={vocab_size})"
+        );
+    }
+
+    embedding_device_ids(table, ids_gpu.as_ptr() as *const c_void, num_tokens)
+}
+
+/// Embedding lookup with token ids already on the GPU (u32, [num_tokens]).
+/// Used by the CUDA-graph decode path, where ids live in a persistent device
+/// buffer updated outside the captured region (no bounds check possible here).
+pub fn embedding_device_ids(table: &Tensor, ids_gpu: *const c_void, num_tokens: usize) -> Tensor {
+    assert_eq!(table.ndim(), 2);
+    assert!(table.is_contiguous());
+    assert!(matches!(table.device(), Device::Cuda(_)));
+    let hidden_size = table.shape()[1];
+    let vocab_size = table.shape()[0];
+
+    let out = Tensor::empty(&[num_tokens, hidden_size], table.dtype(), table.device());
 
     unsafe {
         match table.dtype() {
             DType::F32 => launch_embedding_f32(
-                table.data_ptr() as _, ids_gpu.as_ptr() as _,
+                table.data_ptr() as _,
+                ids_gpu,
                 out.data_ptr() as *mut c_void,
-                num_tokens as i32, hidden_size as i32, std::ptr::null_mut(),
+                num_tokens as i32,
+                hidden_size as i32,
+                vocab_size as i32,
+                xserv_cuda::current_stream_raw(),
             ),
             DType::BF16 => launch_embedding_bf16(
-                table.data_ptr() as _, ids_gpu.as_ptr() as _,
+                table.data_ptr() as _,
+                ids_gpu,
                 out.data_ptr() as *mut c_void,
-                num_tokens as i32, hidden_size as i32, std::ptr::null_mut(),
+                num_tokens as i32,
+                hidden_size as i32,
+                vocab_size as i32,
+                xserv_cuda::current_stream_raw(),
             ),
             _ => panic!("unsupported dtype for embedding"),
         }
     }
-    xserv_cuda::device::synchronize().unwrap();
     out
 }

crates/xserv-kernels/src/gemm.rs

@@ -1,7 +1,36 @@
+use std::cell::RefCell;
 use std::ffi::c_void;
+use xserv_cuda::GpuBuffer;
 use xserv_cuda::error::{self, Result};
 use xserv_tensor::{DType, Device, Tensor};
 
+const CUBLAS_WORKSPACE_BYTES: usize = 32 * 1024 * 1024;
+const GEMV_TILE_K: usize = 256;
+
+// GEMV: single-kernel, no FP32 temp buffer needed
+unsafe extern "C" {
+    fn launch_gemv_bf16(
+        x: *const c_void,
+        w: *const c_void,
+        y_bf16: *mut c_void,
+        y_fp32_buf: *mut c_void,
+        k: i32,
+        n: i32,
+        stream: *mut c_void,
+    );
+
+    fn launch_gemv_bf16_batched(
+        x: *const c_void,
+        w: *const c_void,
+        y_bf16: *mut c_void,
+        y_fp32_buf: *mut c_void,
+        m: i32,
+        k: i32,
+        n: i32,
+        stream: *mut c_void,
+    );
+}
+
 #[derive(Debug, Clone, Copy)]
 pub enum GemmBackend {
     Naive,
@@ -9,16 +38,101 @@ pub enum GemmBackend {
     CuBlas,
 }
 
+pub fn gemv_scratch_elems(k: usize, n: usize) -> usize {
+    n * k.div_ceil(GEMV_TILE_K)
+}
+
+/// Batched GEMV: [M, K] × [K, N] → [M, N], all BF16.
+/// Bit-exact with calling matmul on each row individually (same K-block partial
+/// + fixed-order reduction path), but in a single kernel launch per phase.
+pub fn matmul_batched_gemv(a: &Tensor, b: &Tensor) -> Tensor {
+    assert_eq!(a.ndim(), 2);
+    assert_eq!(b.ndim(), 2);
+    assert!(a.is_contiguous());
+    assert!(b.is_contiguous());
+    assert_eq!(a.dtype(), DType::BF16);
+    assert_eq!(b.dtype(), DType::BF16);
+    let m = a.shape()[0];
+    let k = a.shape()[1];
+    let n = b.shape()[1];
+    assert_eq!(b.shape()[0], k);
+
+    let out = Tensor::empty(&[m, n], DType::BF16, a.device());
+    let scratch_elems = m * gemv_scratch_elems(k, n);
+    let mut fp32_buf = xserv_cuda::allocator::cached_alloc(scratch_elems * 4).unwrap();
+
+    let null_stream = xserv_cuda::current_stream_raw();
+    if m == 1 {
+        unsafe {
+            launch_gemv_bf16(
+                a.data_ptr() as *const c_void,
+                b.data_ptr() as *const c_void,
+                out.data_ptr() as *mut c_void,
+                fp32_buf.as_mut_ptr() as *mut c_void,
+                k as i32,
+                n as i32,
+                null_stream,
+            );
+        }
+    } else {
+        unsafe {
+            launch_gemv_bf16_batched(
+                a.data_ptr() as *const c_void,
+                b.data_ptr() as *const c_void,
+                out.data_ptr() as *mut c_void,
+                fp32_buf.as_mut_ptr() as *mut c_void,
+                m as i32,
+                k as i32,
+                n as i32,
+                null_stream,
+            );
+        }
+    }
+    out
+}
+
 // --- FFI: custom CUDA kernels ---
 unsafe extern "C" {
-    fn launch_gemm_naive_f32(a: *const c_void, b: *const c_void, c: *mut c_void, m: i32, n: i32, k: i32, stream: *mut c_void);
-    fn launch_gemm_naive_bf16(a: *const c_void, b: *const c_void, c: *mut c_void, m: i32, n: i32, k: i32, stream: *mut c_void);
-    fn launch_gemm_tiled_f32(a: *const c_void, b: *const c_void, c: *mut c_void, m: i32, n: i32, k: i32, stream: *mut c_void);
-    fn launch_gemm_tiled_bf16(a: *const c_void, b: *const c_void, c: *mut c_void, m: i32, n: i32, k: i32, stream: *mut c_void);
+    fn launch_gemm_naive_f32(
+        a: *const c_void,
+        b: *const c_void,
+        c: *mut c_void,
+        m: i32,
+        n: i32,
+        k: i32,
+        stream: *mut c_void,
+    );
+    fn launch_gemm_naive_bf16(
+        a: *const c_void,
+        b: *const c_void,
+        c: *mut c_void,
+        m: i32,
+        n: i32,
+        k: i32,
+        stream: *mut c_void,
+    );
+    fn launch_gemm_tiled_f32(
+        a: *const c_void,
+        b: *const c_void,
+        c: *mut c_void,
+        m: i32,
+        n: i32,
+        k: i32,
+        stream: *mut c_void,
+    );
+    fn launch_gemm_tiled_bf16(
+        a: *const c_void,
+        b: *const c_void,
+        c: *mut c_void,
+        m: i32,
+        n: i32,
+        k: i32,
+        stream: *mut c_void,
+    );
 }
 
 // --- FFI: cuBLAS ---
-type CublasHandle = *mut c_void;
+pub type CublasHandle = *mut c_void;
 
 #[allow(non_upper_case_globals)]
 const CUBLAS_OP_N: i32 = 0;
@@ -34,15 +148,50 @@ unsafe extern "C" {
     fn cublasCreate_v2(handle: *mut CublasHandle) -> i32;
     fn cublasDestroy_v2(handle: CublasHandle) -> i32;
     fn cublasSetStream_v2(handle: CublasHandle, stream: *mut c_void) -> i32;
+    fn cublasSetWorkspace_v2(handle: CublasHandle, workspace: *mut c_void, size: usize) -> i32;
     fn cublasGemmEx(
         handle: CublasHandle,
-        transa: i32, transb: i32,
-        m: i32, n: i32, k: i32,
+        transa: i32,
+        transb: i32,
+        m: i32,
+        n: i32,
+        k: i32,
         alpha: *const c_void,
-        a: *const c_void, a_type: i32, lda: i32,
-        b: *const c_void, b_type: i32, ldb: i32,
+        a: *const c_void,
+        a_type: i32,
+        lda: i32,
+        b: *const c_void,
+        b_type: i32,
+        ldb: i32,
         beta: *const c_void,
-        c: *mut c_void, c_type: i32, ldc: i32,
+        c: *mut c_void,
+        c_type: i32,
+        ldc: i32,
+        compute_type: i32,
+        algo: i32,
+    ) -> i32;
+    fn cublasGemmStridedBatchedEx(
+        handle: CublasHandle,
+        transa: i32,
+        transb: i32,
+        m: i32,
+        n: i32,
+        k: i32,
+        alpha: *const c_void,
+        a: *const c_void,
+        a_type: i32,
+        lda: i32,
+        stride_a: i64,
+        b: *const c_void,
+        b_type: i32,
+        ldb: i32,
+        stride_b: i64,
+        beta: *const c_void,
+        c: *mut c_void,
+        c_type: i32,
+        ldc: i32,
+        stride_c: i64,
+        batch_count: i32,
         compute_type: i32,
         algo: i32,
     ) -> i32;
@@ -50,13 +199,32 @@ unsafe extern "C" {
 
 pub struct CublasContext {
     handle: CublasHandle,
+    /// Dedicated 32 MiB workspace owned by this handle. Held to keep the GPU
+    /// buffer alive for the lifetime of the handle; cuBLAS reads/writes into
+    /// it during GEMM. Dropped after `cublasDestroy_v2` so cuBLAS can't touch
+    /// freed memory.
+    _workspace: Option<GpuBuffer>,
 }
 
 impl CublasContext {
     pub fn new() -> Result<Self> {
         let mut handle = std::ptr::null_mut();
         error::check(unsafe { cublasCreate_v2(&mut handle) })?;
-        Ok(Self { handle })
+        // Attach a per-handle workspace. cublasSetWorkspace requires the
+        // pointer to remain valid until destroy or until a new workspace is
+        // set, so we keep the GpuBuffer in this struct.
+        let mut workspace = GpuBuffer::alloc(CUBLAS_WORKSPACE_BYTES)?;
+        error::check(unsafe {
+            cublasSetWorkspace_v2(
+                handle,
+                workspace.as_mut_ptr() as *mut c_void,
+                CUBLAS_WORKSPACE_BYTES,
+            )
+        })?;
+        Ok(Self {
+            handle,
+            _workspace: Some(workspace),
+        })
     }
 }
 
@@ -65,9 +233,32 @@ impl Drop for CublasContext {
         if !self.handle.is_null() {
             unsafe { cublasDestroy_v2(self.handle) };
         }
+        // _workspace drops here, after cublasDestroy_v2 has released the handle.
     }
 }
 
+thread_local! {
+    static CUBLAS_CTX: RefCell<CublasContext> = RefCell::new(
+        CublasContext::new().expect("failed to create thread-local cuBLAS handle")
+    );
+}
+
+/// Borrow the thread-local cuBLAS handle for the duration of a closure.
+fn with_cublas<F, R>(f: F) -> R
+where
+    F: FnOnce(CublasHandle) -> R,
+{
+    CUBLAS_CTX.with(|cell| {
+        let ctx = cell.borrow();
+        f(ctx.handle)
+    })
+}
+
+/// Get the thread-local cuBLAS handle for use with dispatch module.
+pub fn cublas_handle() -> CublasHandle {
+    CUBLAS_CTX.with(|cell| cell.borrow().handle)
+}
+
 /// Matrix multiplication: C = A @ B
 /// A: [M, K], B: [K, N], C: [M, N]
 /// All tensors must be contiguous and on the same GPU.
@@ -76,76 +267,206 @@ pub fn matmul(a: &Tensor, b: &Tensor, backend: GemmBackend) -> Tensor {
     assert_eq!(b.ndim(), 2);
     assert_eq!(a.shape()[1], b.shape()[0], "inner dimension mismatch");
     assert_eq!(a.dtype(), b.dtype(), "dtype mismatch");
-    assert!(a.is_contiguous() && b.is_contiguous(), "matmul requires contiguous tensors");
-    assert!(matches!(a.device(), Device::Cuda(_)), "matmul requires GPU tensors");
+    assert!(
+        a.is_contiguous() && b.is_contiguous(),
+        "matmul requires contiguous tensors"
+    );
+    assert!(
+        matches!(a.device(), Device::Cuda(_)),
+        "matmul requires GPU tensors"
+    );
 
     let m = a.shape()[0];
     let k = a.shape()[1];
     let n = b.shape()[1];
     let dtype = a.dtype();
 
-    let c = Tensor::zeros(&[m, n], dtype, a.device());
+    // All backends (naive, tiled, cuBLAS with beta=0, custom GEMV) fully
+    // overwrite every element of C, so we skip the cudaMemset.
+    let c = Tensor::empty(&[m, n], dtype, a.device());
 
     let a_ptr = a.data_ptr() as *const c_void;
     let b_ptr = b.data_ptr() as *const c_void;
     let c_ptr = c.data_ptr() as *mut c_void;
-    let null_stream = std::ptr::null_mut();
+    let null_stream = xserv_cuda::current_stream_raw();
 
     match backend {
-        GemmBackend::Naive => {
-            unsafe {
-                match dtype {
-                    DType::F32 => launch_gemm_naive_f32(a_ptr, b_ptr, c_ptr, m as i32, n as i32, k as i32, null_stream),
-                    DType::BF16 => launch_gemm_naive_bf16(a_ptr, b_ptr, c_ptr, m as i32, n as i32, k as i32, null_stream),
-                    _ => panic!("unsupported dtype for naive GEMM"),
-                }
+        GemmBackend::Naive => unsafe {
+            match dtype {
+                DType::F32 => launch_gemm_naive_f32(
+                    a_ptr,
+                    b_ptr,
+                    c_ptr,
+                    m as i32,
+                    n as i32,
+                    k as i32,
+                    null_stream,
+                ),
+                DType::BF16 => launch_gemm_naive_bf16(
+                    a_ptr,
+                    b_ptr,
+                    c_ptr,
+                    m as i32,
+                    n as i32,
+                    k as i32,
+                    null_stream,
+                ),
+                _ => panic!("unsupported dtype for naive GEMM"),
             }
-            xserv_cuda::device::synchronize().unwrap();
-        }
-        GemmBackend::Tiled => {
-            unsafe {
-                match dtype {
-                    DType::F32 => launch_gemm_tiled_f32(a_ptr, b_ptr, c_ptr, m as i32, n as i32, k as i32, null_stream),
-                    DType::BF16 => launch_gemm_tiled_bf16(a_ptr, b_ptr, c_ptr, m as i32, n as i32, k as i32, null_stream),
-                    _ => panic!("unsupported dtype for tiled GEMM"),
-                }
+        },
+        GemmBackend::Tiled => unsafe {
+            match dtype {
+                DType::F32 => launch_gemm_tiled_f32(
+                    a_ptr,
+                    b_ptr,
+                    c_ptr,
+                    m as i32,
+                    n as i32,
+                    k as i32,
+                    null_stream,
+                ),
+                DType::BF16 => launch_gemm_tiled_bf16(
+                    a_ptr,
+                    b_ptr,
+                    c_ptr,
+                    m as i32,
+                    n as i32,
+                    k as i32,
+                    null_stream,
+                ),
+                _ => panic!("unsupported dtype for tiled GEMM"),
             }
-            xserv_cuda::device::synchronize().unwrap();
-        }
+        },
         GemmBackend::CuBlas => {
-            // cuBLAS uses column-major, but we have row-major tensors.
-            // Trick: compute C^T = B^T @ A^T, which gives us C in row-major.
-            // cuBLAS sees our row-major data as column-major transposed.
-            let ctx = CublasContext::new().unwrap();
-            let alpha = 1.0f32;
-            let beta = 0.0f32;
+            if m == 1 && dtype == DType::BF16 && n >= 256 {
+                let mut fp32_buf =
+                    xserv_cuda::allocator::cached_alloc(gemv_scratch_elems(k, n) * 4).unwrap();
+                unsafe {
+                    launch_gemv_bf16(
+                        a_ptr,
+                        b_ptr,
+                        c_ptr,
+                        fp32_buf.as_mut_ptr() as *mut c_void,
+                        k as i32,
+                        n as i32,
+                        null_stream,
+                    );
+                }
+            } else {
+                let alpha = 1.0f32;
+                let beta = 0.0f32;
 
-            let (a_type, b_type, c_type) = match dtype {
-                DType::F32 => (CUDA_R_32F, CUDA_R_32F, CUDA_R_32F),
-                DType::BF16 => (CUDA_R_16BF, CUDA_R_16BF, CUDA_R_16BF),
-                _ => panic!("unsupported dtype for cuBLAS GEMM"),
-            };
+                let (a_type, b_type, c_type) = match dtype {
+                    DType::F32 => (CUDA_R_32F, CUDA_R_32F, CUDA_R_32F),
+                    DType::BF16 => (CUDA_R_16BF, CUDA_R_16BF, CUDA_R_16BF),
+                    _ => panic!("unsupported dtype for cuBLAS GEMM"),
+                };
 
-            unsafe {
-                cublasSetStream_v2(ctx.handle, null_stream);
-                // Row-major trick: swap A/B and transpose flags
-                // C(row-major) = A @ B  <=>  C^T(col-major) = B^T @ A^T
-                error::check(cublasGemmEx(
-                    ctx.handle,
-                    CUBLAS_OP_N, CUBLAS_OP_N,
-                    n as i32, m as i32, k as i32,
-                    &alpha as *const f32 as *const c_void,
-                    b_ptr, b_type, n as i32,  // B as col-major = B^T
-                    a_ptr, a_type, k as i32,  // A as col-major = A^T
-                    &beta as *const f32 as *const c_void,
-                    c_ptr, c_type, n as i32,  // C as col-major = C^T
-                    CUBLAS_COMPUTE_32F,
-                    -1, // default algo
-                )).expect("cuBLAS GEMM failed");
+                with_cublas(|handle| unsafe {
+                    cublasSetStream_v2(handle, null_stream);
+                    error::check(cublasGemmEx(
+                        handle,
+                        CUBLAS_OP_N,
+                        CUBLAS_OP_N,
+                        n as i32,
+                        m as i32,
+                        k as i32,
+                        &alpha as *const f32 as *const c_void,
+                        b_ptr,
+                        b_type,
+                        n as i32,
+                        a_ptr,
+                        a_type,
+                        k as i32,
+                        &beta as *const f32 as *const c_void,
+                        c_ptr,
+                        c_type,
+                        n as i32,
+                        CUBLAS_COMPUTE_32F,
+                        -1,
+                    ))
+                    .expect("cuBLAS GEMM failed");
+                });
             }
-            xserv_cuda::device::synchronize().unwrap();
         }
     }
 
     c
 }
+
+/// Batched matrix multiplication via cuBLAS: C[b] = A[b] @ B[b]
+/// a: [..., M, K], b: [..., K, N] → [..., M, N]
+/// Leading dimensions must match and tensors must be contiguous.
+pub fn batched_matmul(a: &Tensor, b: &Tensor) -> Tensor {
+    assert!(a.ndim() >= 2 && b.ndim() >= 2);
+    assert_eq!(a.ndim(), b.ndim());
+    assert!(a.is_contiguous() && b.is_contiguous());
+    assert!(matches!(a.device(), Device::Cuda(_)));
+    assert_eq!(a.dtype(), b.dtype());
+
+    let ndim = a.ndim();
+    let m = a.shape()[ndim - 2];
+    let k = a.shape()[ndim - 1];
+    let n = b.shape()[ndim - 1];
+    assert_eq!(b.shape()[ndim - 2], k, "inner dimension mismatch");
+
+    // Compute batch count from leading dimensions
+    let batch: usize = a.shape()[..ndim - 2].iter().product();
+    assert_eq!(
+        b.shape()[..ndim - 2].iter().product::<usize>(),
+        batch,
+        "batch dimensions mismatch"
+    );
+
+    let mut out_shape: Vec<usize> = a.shape()[..ndim - 2].to_vec();
+    out_shape.push(m);
+    out_shape.push(n);
+    // cuBLAS with beta=0 fully overwrites every element of C.
+    let c = Tensor::empty(&out_shape, a.dtype(), a.device());
+
+    let dtype = a.dtype();
+    let (a_type, b_type, c_type) = match dtype {
+        DType::F32 => (CUDA_R_32F, CUDA_R_32F, CUDA_R_32F),
+        DType::BF16 => (CUDA_R_16BF, CUDA_R_16BF, CUDA_R_16BF),
+        _ => panic!("unsupported dtype for batched matmul"),
+    };
+
+    let alpha = 1.0f32;
+    let beta = 0.0f32;
+    // cuBLAS strides are in elements (not bytes)
+    let stride_a = (m * k) as i64;
+    let stride_b = (k * n) as i64;
+    let stride_c = (m * n) as i64;
+
+    with_cublas(|handle| unsafe {
+        cublasSetStream_v2(handle, xserv_cuda::current_stream_raw());
+        // Row-major trick: C = A @ B ⟺ C^T = B^T @ A^T (col-major)
+        error::check(cublasGemmStridedBatchedEx(
+            handle,
+            CUBLAS_OP_N,
+            CUBLAS_OP_N,
+            n as i32,
+            m as i32,
+            k as i32,
+            &alpha as *const f32 as *const c_void,
+            b.data_ptr() as _,
+            b_type,
+            n as i32,
+            stride_b,
+            a.data_ptr() as _,
+            a_type,
+            k as i32,
+            stride_a,
+            &beta as *const f32 as *const c_void,
+            c.data_ptr() as *mut c_void,
+            c_type,
+            n as i32,
+            stride_c,
+            batch as i32,
+            CUBLAS_COMPUTE_32F,
+            -1,
+        ))
+        .expect("cuBLAS batched GEMM failed");
+    });
+    c
+}

crates/xserv-kernels/src/layernorm.rs

@@ -2,10 +2,26 @@ use std::ffi::c_void;
 use xserv_tensor::{DType, Device, Tensor};
 
 unsafe extern "C" {
-    fn launch_layernorm_f32(x: *const c_void, gamma: *const c_void, beta: *const c_void,
-                            out: *mut c_void, rows: i32, hidden_size: i32, eps: f32, stream: *mut c_void);
-    fn launch_layernorm_bf16(x: *const c_void, gamma: *const c_void, beta: *const c_void,
-                             out: *mut c_void, rows: i32, hidden_size: i32, eps: f32, stream: *mut c_void);
+    fn launch_layernorm_f32(
+        x: *const c_void,
+        gamma: *const c_void,
+        beta: *const c_void,
+        out: *mut c_void,
+        rows: i32,
+        hidden_size: i32,
+        eps: f32,
+        stream: *mut c_void,
+    );
+    fn launch_layernorm_bf16(
+        x: *const c_void,
+        gamma: *const c_void,
+        beta: *const c_void,
+        out: *mut c_void,
+        rows: i32,
+        hidden_size: i32,
+        eps: f32,
+        stream: *mut c_void,
+    );
 }
 
 pub fn layernorm(x: &Tensor, gamma: &Tensor, beta: &Tensor, eps: f32) -> Tensor {
@@ -17,23 +33,40 @@ pub fn layernorm(x: &Tensor, gamma: &Tensor, beta: &Tensor, eps: f32) -> Tensor
     assert_eq!(beta.shape(), &[hidden_size]);
 
     let rows = x.numel() / hidden_size;
-    let out = Tensor::zeros(x.shape(), x.dtype(), x.device());
+    assert!(
+        rows <= i32::MAX as usize,
+        "too many rows for i32 kernel param"
+    );
+    assert!(
+        hidden_size <= i32::MAX as usize,
+        "hidden_size too large for i32 kernel param"
+    );
+    let out = Tensor::empty(x.shape(), x.dtype(), x.device());
 
     unsafe {
         match x.dtype() {
             DType::F32 => launch_layernorm_f32(
-                x.data_ptr() as _, gamma.data_ptr() as _, beta.data_ptr() as _,
+                x.data_ptr() as _,
+                gamma.data_ptr() as _,
+                beta.data_ptr() as _,
                 out.data_ptr() as *mut c_void,
-                rows as i32, hidden_size as i32, eps, std::ptr::null_mut(),
+                rows as i32,
+                hidden_size as i32,
+                eps,
+                xserv_cuda::current_stream_raw(),
             ),
             DType::BF16 => launch_layernorm_bf16(
-                x.data_ptr() as _, gamma.data_ptr() as _, beta.data_ptr() as _,
+                x.data_ptr() as _,
+                gamma.data_ptr() as _,
+                beta.data_ptr() as _,
                 out.data_ptr() as *mut c_void,
-                rows as i32, hidden_size as i32, eps, std::ptr::null_mut(),
+                rows as i32,
+                hidden_size as i32,
+                eps,
+                xserv_cuda::current_stream_raw(),
             ),
             _ => panic!("unsupported dtype for layernorm"),
         }
     }
-    xserv_cuda::device::synchronize().unwrap();
     out
 }

crates/xserv-kernels/src/lib.rs

@@ -1,15 +1,36 @@
 pub mod activation;
+pub mod argmax;
+pub mod attention;
+pub mod dispatch;
 pub mod embedding;
 pub mod gemm;
 pub mod layernorm;
+pub mod moe;
+pub mod quantization;
 pub mod rmsnorm;
 pub mod rope;
 pub mod softmax;
+pub mod transpose;
 
-pub use activation::{gelu, silu};
-pub use embedding::embedding;
-pub use gemm::{matmul, GemmBackend};
+pub use activation::{add, bias_add_2d, gelu, gpt_oss_glu, mul, scale, silu, silu_mul};
+pub use argmax::{argmax_bf16_single, argmax_bf16_to_host};
+pub use attention::{
+    attention, copy_kv_position, decode_attention, flash_attention, flash_attention_sinks,
+    paged_decode_attention, paged_decode_attention_sinks, paged_decode_attention_tree,
+    reshape_and_cache_batched_bf16, reshape_and_cache_bf16,
+};
+pub use embedding::{embedding, embedding_device_ids};
+pub use gemm::{GemmBackend, batched_matmul, matmul, matmul_batched_gemv};
 pub use layernorm::layernorm;
-pub use rmsnorm::rmsnorm;
-pub use rope::{rope_inplace, RopeCache};
+pub use rmsnorm::{add_rmsnorm, rmsnorm};
+pub use rope::{RopeCache, rope_inplace, rope_inplace_device_pos};
 pub use softmax::softmax;
+pub use transpose::{
+    merge_heads_gpu, repeat_kv_gpu, reshape_heads_gpu, strided_to_contiguous_gpu,
+    transpose_for_rope_gpu, transpose_from_rope_gpu,
+};
+
+/// Register GPU kernels with the tensor crate. Call once at startup.
+pub fn init() {
+    xserv_tensor::register_gpu_contiguous(strided_to_contiguous_gpu);
+}

crates/xserv-kernels/src/moe.rs

@@ -0,0 +1,474 @@
+use std::ffi::c_void;
+use xserv_tensor::{DType, Tensor};
+
+use crate::gemm::{CublasHandle, cublas_handle};
+
+unsafe extern "C" {
+    fn launch_moe_topk_softmax_bf16(
+        router_logits: *const c_void,
+        topk_ids: *mut c_void,
+        topk_weights: *mut c_void,
+        num_tokens: i32,
+        num_experts: i32,
+        top_k: i32,
+        stream: *mut c_void,
+    );
+    fn launch_moe_replicate_bf16(
+        x: *const c_void,
+        x_rep: *mut c_void,
+        num_tokens: i32,
+        hidden: i32,
+        local_experts: i32,
+        stream: *mut c_void,
+    );
+    fn launch_moe_bias_add_3d_bf16(
+        x: *mut c_void,
+        bias: *const c_void,
+        batch: i32,
+        num_tokens: i32,
+        dim: i32,
+        stream: *mut c_void,
+    );
+    fn launch_moe_weighted_sum_bf16(
+        expert_out: *const c_void,
+        topk_ids: *const c_void,
+        topk_weights: *const c_void,
+        out: *mut c_void,
+        num_tokens: i32,
+        hidden: i32,
+        top_k: i32,
+        expert_start: i32,
+        local_experts: i32,
+        stream: *mut c_void,
+    );
+
+    fn launch_moe_sparse_gemv_fp8_bf16(
+        x: *const c_void,
+        w: *const c_void,
+        w_scales: *const c_void,
+        bias: *const c_void,
+        topk_ids: *const c_void,
+        y: *mut c_void,
+        num_tokens: i32,
+        n: i32,
+        k: i32,
+        top_k: i32,
+        expert_start: i32,
+        local_experts: i32,
+        x_per_slot: i32,
+        stream: *mut c_void,
+    );
+    fn launch_moe_sparse_gemv_mxfp4_bf16(
+        x: *const c_void,
+        w_packed: *const c_void,
+        w_scales: *const c_void,
+        bias: *const c_void,
+        topk_ids: *const c_void,
+        y: *mut c_void,
+        num_tokens: i32,
+        n: i32,
+        k: i32,
+        top_k: i32,
+        expert_start: i32,
+        local_experts: i32,
+        x_per_slot: i32,
+        stream: *mut c_void,
+    );
+    fn launch_moe_weighted_sum_sparse_bf16(
+        down: *const c_void,
+        topk_ids: *const c_void,
+        topk_weights: *const c_void,
+        out: *mut c_void,
+        num_tokens: i32,
+        hidden: i32,
+        top_k: i32,
+        expert_start: i32,
+        local_experts: i32,
+        stream: *mut c_void,
+    );
+
+    fn cublasGemmStridedBatchedEx(
+        handle: CublasHandle,
+        transa: i32,
+        transb: i32,
+        m: i32,
+        n: i32,
+        k: i32,
+        alpha: *const c_void,
+        a: *const c_void,
+        a_type: i32,
+        lda: i32,
+        stride_a: i64,
+        b: *const c_void,
+        b_type: i32,
+        ldb: i32,
+        stride_b: i64,
+        beta: *const c_void,
+        c: *mut c_void,
+        c_type: i32,
+        ldc: i32,
+        stride_c: i64,
+        batch_count: i32,
+        compute_type: i32,
+        algo: i32,
+    ) -> i32;
+
+    fn cublasSetStream_v2(handle: CublasHandle, stream: *mut c_void) -> i32;
+}
+
+const CUDA_R_16BF: i32 = 14;
+const CUBLAS_COMPUTE_32F: i32 = 68;
+const CUBLAS_GEMM_DEFAULT: i32 = -1;
+
+/// GPU top-k selection + softmax over router logits.
+///
+/// Input:  router_logits [num_tokens, num_experts] BF16 on GPU
+/// Output: (topk_ids [num_tokens, top_k] i32, topk_weights [num_tokens, top_k] f32)
+pub fn moe_topk_softmax(
+    router_logits: &Tensor,
+    num_experts: usize,
+    top_k: usize,
+) -> (Tensor, Tensor) {
+    assert_eq!(router_logits.ndim(), 2);
+    assert_eq!(router_logits.dtype(), DType::BF16);
+    assert!(router_logits.is_contiguous());
+    let num_tokens = router_logits.shape()[0];
+    assert_eq!(router_logits.shape()[1], num_experts);
+
+    // NOTE: topk_ids actually holds i32 expert indices; DType has no I32, so
+    // this is a raw 4-byte buffer mislabeled F32. Never read it as floats —
+    // all consumers (weighted-sum / sparse GEMV kernels) cast to int*.
+    let topk_ids = Tensor::empty(&[num_tokens, top_k], DType::F32, router_logits.device());
+    let topk_weights = Tensor::empty(&[num_tokens, top_k], DType::F32, router_logits.device());
+
+    unsafe {
+        launch_moe_topk_softmax_bf16(
+            router_logits.data_ptr() as *const c_void,
+            topk_ids.data_ptr() as *mut c_void,
+            topk_weights.data_ptr() as *mut c_void,
+            num_tokens as i32,
+            num_experts as i32,
+            top_k as i32,
+            xserv_cuda::current_stream_raw(),
+        );
+    }
+
+    (topk_ids, topk_weights)
+}
+
+/// Replicate x [num_tokens, hidden] → [local_experts, num_tokens, hidden].
+pub fn moe_replicate(x: &Tensor, local_experts: usize) -> Tensor {
+    assert_eq!(x.ndim(), 2);
+    assert_eq!(x.dtype(), DType::BF16);
+    assert!(x.is_contiguous());
+    let num_tokens = x.shape()[0];
+    let hidden = x.shape()[1];
+    let out = Tensor::empty(
+        &[local_experts, num_tokens, hidden],
+        DType::BF16,
+        x.device(),
+    );
+
+    unsafe {
+        launch_moe_replicate_bf16(
+            x.data_ptr() as *const c_void,
+            out.data_ptr() as *mut c_void,
+            num_tokens as i32,
+            hidden as i32,
+            local_experts as i32,
+            xserv_cuda::current_stream_raw(),
+        );
+    }
+
+    out
+}
+
+/// In-place 3D bias add: x [batch, num_tokens, dim] += bias [batch, dim].
+pub fn moe_bias_add_3d(x: &Tensor, bias: &Tensor) {
+    assert_eq!(x.ndim(), 3);
+    assert_eq!(bias.ndim(), 2);
+    assert_eq!(x.dtype(), DType::BF16);
+    assert!(x.is_contiguous());
+    let batch = x.shape()[0];
+    let num_tokens = x.shape()[1];
+    let dim = x.shape()[2];
+    assert_eq!(bias.shape(), &[batch, dim]);
+
+    unsafe {
+        launch_moe_bias_add_3d_bf16(
+            x.data_ptr() as *mut c_void,
+            bias.data_ptr() as *const c_void,
+            batch as i32,
+            num_tokens as i32,
+            dim as i32,
+            xserv_cuda::current_stream_raw(),
+        );
+    }
+}
+
+/// Weighted sum of expert outputs → [num_tokens, hidden].
+///
+/// expert_out:   [local_experts, num_tokens, hidden] BF16
+/// topk_ids:     [num_tokens, top_k] i32 (global expert indices)
+/// topk_weights: [num_tokens, top_k] f32
+pub fn moe_weighted_sum(
+    expert_out: &Tensor,
+    topk_ids: &Tensor,
+    topk_weights: &Tensor,
+    expert_start: usize,
+    local_experts: usize,
+    top_k: usize,
+) -> Tensor {
+    assert_eq!(expert_out.ndim(), 3);
+    assert_eq!(expert_out.dtype(), DType::BF16);
+    let num_tokens = expert_out.shape()[1];
+    let hidden = expert_out.shape()[2];
+
+    let out = Tensor::empty(&[num_tokens, hidden], DType::BF16, expert_out.device());
+
+    unsafe {
+        launch_moe_weighted_sum_bf16(
+            expert_out.data_ptr() as *const c_void,
+            topk_ids.data_ptr() as *const c_void,
+            topk_weights.data_ptr() as *const c_void,
+            out.data_ptr() as *mut c_void,
+            num_tokens as i32,
+            hidden as i32,
+            top_k as i32,
+            expert_start as i32,
+            local_experts as i32,
+            xserv_cuda::current_stream_raw(),
+        );
+    }
+
+    out
+}
+
+/// Sparse MoE GEMV (FP8 W8A16): compute only the routed experts.
+///
+/// x:        [num_tokens, K] BF16 (x_per_slot=false, gate_up) or
+///           [num_tokens * top_k, K] BF16 (x_per_slot=true, down)
+/// w_fp8_t:  [local_experts, N, K] FP8E4M3 (transposed weight layout)
+/// w_scales: [local_experts] F32 per-expert scalar scales
+/// bias:     [local_experts, N] BF16 (fused into the epilogue)
+/// topk_ids: [num_tokens, top_k] i32 global expert ids (GPU)
+///
+/// Returns y [num_tokens, top_k, N] BF16. Slots routed to experts NOT
+/// owned by this rank are left UNWRITTEN (uninitialized memory) — the
+/// consumer must skip them (see moe_weighted_sum_sparse).
+#[allow(clippy::too_many_arguments)]
+pub fn moe_sparse_gemv_fp8(
+    x: &Tensor,
+    w_fp8_t: &Tensor,
+    w_scales: &Tensor,
+    bias: &Tensor,
+    topk_ids: &Tensor,
+    num_tokens: usize,
+    top_k: usize,
+    expert_start: usize,
+    local_experts: usize,
+    x_per_slot: bool,
+) -> Tensor {
+    assert_eq!(x.dtype(), DType::BF16);
+    assert!(x.is_contiguous());
+    assert_eq!(w_fp8_t.dtype(), DType::FP8E4M3);
+    let n = w_fp8_t.shape()[1];
+    let k = w_fp8_t.shape()[2];
+    // The kernel reads weights as uint4 (16 FP8 values per lane) and would
+    // silently skip a K%16 tail.
+    assert_eq!(k % 16, 0, "sparse FP8 GEMV requires K % 16 == 0, got {k}");
+    assert_eq!(x.shape()[x.ndim() - 1], k);
+    assert_eq!(
+        x.shape()[0],
+        if x_per_slot {
+            num_tokens * top_k
+        } else {
+            num_tokens
+        }
+    );
+
+    let y = Tensor::empty(&[num_tokens, top_k, n], DType::BF16, x.device());
+    unsafe {
+        launch_moe_sparse_gemv_fp8_bf16(
+            x.data_ptr() as *const c_void,
+            w_fp8_t.data_ptr() as *const c_void,
+            w_scales.data_ptr() as *const c_void,
+            bias.data_ptr() as *const c_void,
+            topk_ids.data_ptr() as *const c_void,
+            y.data_ptr() as *mut c_void,
+            num_tokens as i32,
+            n as i32,
+            k as i32,
+            top_k as i32,
+            expert_start as i32,
+            local_experts as i32,
+            x_per_slot as i32,
+            xserv_cuda::current_stream_raw(),
+        );
+    }
+    y
+}
+
+/// Sparse MoE GEMV (MXFP4 W4A16): same contract as moe_sparse_gemv_fp8,
+/// with packed 4-bit weights [E, N, K/2] + UE8M0 block scales [E, N, K/32].
+#[allow(clippy::too_many_arguments)]
+pub fn moe_sparse_gemv_mxfp4(
+    x: &Tensor,
+    w_packed: &Tensor,
+    w_scales: &Tensor,
+    bias: &Tensor,
+    topk_ids: &Tensor,
+    num_tokens: usize,
+    top_k: usize,
+    n: usize,
+    k: usize,
+    expert_start: usize,
+    local_experts: usize,
+    x_per_slot: bool,
+) -> Tensor {
+    assert_eq!(x.dtype(), DType::BF16);
+    assert!(x.is_contiguous());
+    // 32-element MXFP4 blocks, read as uint4 (32 nibbles) per lane.
+    assert_eq!(k % 32, 0, "sparse MXFP4 GEMV requires K % 32 == 0, got {k}");
+    assert_eq!(x.shape()[x.ndim() - 1], k);
+    assert_eq!(
+        x.shape()[0],
+        if x_per_slot {
+            num_tokens * top_k
+        } else {
+            num_tokens
+        }
+    );
+
+    let y = Tensor::empty(&[num_tokens, top_k, n], DType::BF16, x.device());
+    unsafe {
+        launch_moe_sparse_gemv_mxfp4_bf16(
+            x.data_ptr() as *const c_void,
+            w_packed.data_ptr() as *const c_void,
+            w_scales.data_ptr() as *const c_void,
+            bias.data_ptr() as *const c_void,
+            topk_ids.data_ptr() as *const c_void,
+            y.data_ptr() as *mut c_void,
+            num_tokens as i32,
+            n as i32,
+            k as i32,
+            top_k as i32,
+            expert_start as i32,
+            local_experts as i32,
+            x_per_slot as i32,
+            xserv_cuda::current_stream_raw(),
+        );
+    }
+    y
+}
+
+/// Weighted sum over the slot axis of the sparse GEMV output.
+///
+/// down: [num_tokens, top_k, hidden] BF16 (non-local slots uninitialized
+/// and skipped, never multiplied by zero — NaN * 0 = NaN).
+pub fn moe_weighted_sum_sparse(
+    down: &Tensor,
+    topk_ids: &Tensor,
+    topk_weights: &Tensor,
+    expert_start: usize,
+    local_experts: usize,
+) -> Tensor {
+    assert_eq!(down.ndim(), 3);
+    assert_eq!(down.dtype(), DType::BF16);
+    let num_tokens = down.shape()[0];
+    let top_k = down.shape()[1];
+    let hidden = down.shape()[2];
+
+    let out = Tensor::empty(&[num_tokens, hidden], DType::BF16, down.device());
+    unsafe {
+        launch_moe_weighted_sum_sparse_bf16(
+            down.data_ptr() as *const c_void,
+            topk_ids.data_ptr() as *const c_void,
+            topk_weights.data_ptr() as *const c_void,
+            out.data_ptr() as *mut c_void,
+            num_tokens as i32,
+            hidden as i32,
+            top_k as i32,
+            expert_start as i32,
+            local_experts as i32,
+            xserv_cuda::current_stream_raw(),
+        );
+    }
+    out
+}
+
+/// Strided batched GEMM for MoE expert forward.
+/// C[b] = A[b] @ B[b]  for b in 0..batch
+///
+/// A: [batch, M, K] BF16 contiguous
+/// B: [batch, K, N] BF16 contiguous
+/// Returns C: [batch, M, N] BF16
+#[allow(clippy::too_many_arguments)]
+pub fn batched_gemm_strided(a: &Tensor, b: &Tensor) -> Tensor {
+    assert_eq!(a.ndim(), 3);
+    assert_eq!(b.ndim(), 3);
+    assert_eq!(a.dtype(), DType::BF16);
+    assert_eq!(b.dtype(), DType::BF16);
+    assert!(a.is_contiguous() && b.is_contiguous());
+    assert_eq!(a.shape()[0], b.shape()[0]);
+    assert_eq!(a.shape()[2], b.shape()[1]);
+
+    let batch = a.shape()[0];
+    let m = a.shape()[1];
+    let k = a.shape()[2];
+    let n = b.shape()[2];
+
+    let c = Tensor::empty(&[batch, m, n], DType::BF16, a.device());
+
+    let alpha: f32 = 1.0;
+    let beta: f32 = 0.0;
+
+    // cuBLAS column-major: we compute C^T = B^T @ A^T
+    // A is [batch, M, K] row-major → A^T is [K, M] col-major, lda=K
+    // B is [batch, K, N] row-major → B^T is [N, K] col-major, ldb=N? No...
+    //
+    // Actually for row-major: A[M,K] in memory = col-major A^T[K,M] with lda=K.
+    // So we call cublasGemmStridedBatchedEx with:
+    //   transa=N, transb=N
+    //   m=N, n=M, k=K  (because cuBLAS sees col-major)
+    //   A_cublas = B_row (pointer), lda=N
+    //   B_cublas = A_row (pointer), ldb=K
+    //   C_cublas = C_row (pointer), ldc=N
+
+    let stride_a = (m * k) as i64;
+    let stride_b = (k * n) as i64;
+    let stride_c = (m * n) as i64;
+
+    let handle = cublas_handle();
+    unsafe {
+        cublasSetStream_v2(handle, xserv_cuda::current_stream_raw());
+        let status = cublasGemmStridedBatchedEx(
+            handle,
+            0,
+            0, // CUBLAS_OP_N, CUBLAS_OP_N
+            n as i32,
+            m as i32,
+            k as i32,
+            &alpha as *const f32 as *const c_void,
+            b.data_ptr() as *const c_void,
+            CUDA_R_16BF,
+            n as i32,
+            stride_b,
+            a.data_ptr() as *const c_void,
+            CUDA_R_16BF,
+            k as i32,
+            stride_a,
+            &beta as *const f32 as *const c_void,
+            c.data_ptr() as *mut c_void,
+            CUDA_R_16BF,
+            n as i32,
+            stride_c,
+            batch as i32,
+            CUBLAS_COMPUTE_32F,
+            CUBLAS_GEMM_DEFAULT,
+        );
+        assert_eq!(status, 0, "cublasGemmStridedBatchedEx failed: {status}");
+    }
+
+    c
+}

crates/xserv-kernels/src/quantization.rs

@@ -0,0 +1,603 @@
+use std::cell::RefCell;
+use std::collections::HashMap;
+use std::ffi::c_void;
+use xserv_cuda::GpuBuffer;
+use xserv_tensor::{DType, Tensor};
+
+// ============================================================
+// FFI: custom CUDA kernels
+// ============================================================
+
+unsafe extern "C" {
+    fn launch_dequant_fp8e4m3_to_bf16(
+        src: *const c_void,
+        scales: *const c_void,
+        dst: *mut c_void,
+        num_experts: i32,
+        rows: i32,
+        cols: i32,
+        stream: *mut c_void,
+    );
+    fn launch_quantize_bf16_to_fp8e4m3_rowwise(
+        src: *const c_void,
+        dst: *mut c_void,
+        scales: *mut c_void,
+        num_rows: i32,
+        cols: i32,
+        stream: *mut c_void,
+    );
+    fn launch_rowwise_scale_moe_bf16(
+        data: *mut c_void,
+        a_scales: *const c_void,
+        b_scales: *const c_void,
+        num_rows: i32,
+        cols: i32,
+        tokens: i32,
+        stream: *mut c_void,
+    );
+    fn launch_batched_gemv_mxfp4_bf16(
+        x: *const c_void,
+        w_packed: *const c_void,
+        w_scales: *const c_void,
+        y: *mut c_void,
+        e: i32,
+        n: i32,
+        k: i32,
+        stream: *mut c_void,
+    );
+    fn launch_dequant_mxfp4_to_bf16_t(
+        w_packed: *const c_void,
+        w_scales: *const c_void,
+        out: *mut c_void,
+        e: i32,
+        n: i32,
+        k: i32,
+        stream: *mut c_void,
+    );
+}
+
+// ============================================================
+// FFI: cuBLASLt
+// ============================================================
+
+type CublasLtHandle = *mut c_void;
+type CublasLtMatmulDesc = *mut c_void;
+type CublasLtMatrixLayout = *mut c_void;
+type CublasLtMatmulPreference = *mut c_void;
+
+#[repr(C)]
+#[derive(Clone, Copy)]
+struct CublasLtMatmulAlgo {
+    data: [u64; 8],
+}
+
+#[repr(C)]
+struct CublasLtMatmulHeuristicResult {
+    algo: CublasLtMatmulAlgo,
+    workspace_size: usize,
+    state: i32,
+    _reserved: [f32; 4],
+}
+
+unsafe extern "C" {
+    fn cublasLtCreate(handle: *mut CublasLtHandle) -> i32;
+    fn cublasLtDestroy(handle: CublasLtHandle) -> i32;
+    fn cublasLtMatmulDescCreate(
+        desc: *mut CublasLtMatmulDesc,
+        compute_type: i32,
+        scale_type: i32,
+    ) -> i32;
+    fn cublasLtMatmulDescDestroy(desc: CublasLtMatmulDesc) -> i32;
+    fn cublasLtMatmulDescSetAttribute(
+        desc: CublasLtMatmulDesc,
+        attr: i32,
+        buf: *const c_void,
+        size: usize,
+    ) -> i32;
+    fn cublasLtMatrixLayoutCreate(
+        layout: *mut CublasLtMatrixLayout,
+        dtype: i32,
+        rows: u64,
+        cols: u64,
+        ld: i64,
+    ) -> i32;
+    fn cublasLtMatrixLayoutDestroy(layout: CublasLtMatrixLayout) -> i32;
+    fn cublasLtMatrixLayoutSetAttribute(
+        layout: CublasLtMatrixLayout,
+        attr: i32,
+        buf: *const c_void,
+        size: usize,
+    ) -> i32;
+    fn cublasLtMatmulPreferenceCreate(pref: *mut CublasLtMatmulPreference) -> i32;
+    fn cublasLtMatmulPreferenceDestroy(pref: CublasLtMatmulPreference) -> i32;
+    fn cublasLtMatmulPreferenceSetAttribute(
+        pref: CublasLtMatmulPreference,
+        attr: i32,
+        buf: *const c_void,
+        size: usize,
+    ) -> i32;
+    fn cublasLtMatmulAlgoGetHeuristic(
+        handle: CublasLtHandle,
+        desc: CublasLtMatmulDesc,
+        a_layout: CublasLtMatrixLayout,
+        b_layout: CublasLtMatrixLayout,
+        c_layout: CublasLtMatrixLayout,
+        d_layout: CublasLtMatrixLayout,
+        pref: CublasLtMatmulPreference,
+        requested: i32,
+        results: *mut CublasLtMatmulHeuristicResult,
+        found: *mut i32,
+    ) -> i32;
+    fn cublasLtMatmul(
+        handle: CublasLtHandle,
+        desc: CublasLtMatmulDesc,
+        alpha: *const c_void,
+        a: *const c_void,
+        a_layout: CublasLtMatrixLayout,
+        b: *const c_void,
+        b_layout: CublasLtMatrixLayout,
+        beta: *const c_void,
+        c: *const c_void,
+        c_layout: CublasLtMatrixLayout,
+        d: *mut c_void,
+        d_layout: CublasLtMatrixLayout,
+        algo: *const CublasLtMatmulAlgo,
+        workspace: *mut c_void,
+        workspace_size: usize,
+        stream: *mut c_void,
+    ) -> i32;
+}
+
+// cuBLASLt constants
+const CUBLAS_COMPUTE_32F: i32 = 68;
+const CUDA_R_32F: i32 = 0;
+const CUDA_R_16BF: i32 = 14;
+const CUDA_R_8F_E4M3: i32 = 28;
+
+// MatmulDesc attributes
+const CUBLASLT_MATMUL_DESC_A_SCALE_POINTER: i32 = 17;
+const CUBLASLT_MATMUL_DESC_B_SCALE_POINTER: i32 = 18;
+
+// MatrixLayout attributes
+const CUBLASLT_MATRIX_LAYOUT_BATCH_COUNT: i32 = 5;
+const CUBLASLT_MATRIX_LAYOUT_STRIDED_BATCH_OFFSET: i32 = 6;
+
+// MatmulPreference attributes
+const CUBLASLT_MATMUL_PREF_MAX_WORKSPACE_BYTES: i32 = 1;
+
+const WORKSPACE_BYTES: usize = 32 * 1024 * 1024;
+
+const CUBLASLT_MATMUL_DESC_TRANSA: i32 = 3;
+
+/// A fully-prepared FP8 matmul plan for one (M, N, K) shape: the matmul
+/// descriptor, the four matrix layouts, and the heuristically-chosen algo.
+/// Built once per shape and reused across every expert and every forward
+/// pass — the heuristic search and descriptor/layout creation are the
+/// expensive parts, so doing them once instead of per-expert-per-layer is
+/// the difference between FP8 being faster or slower than BF16.
+#[derive(Clone, Copy)]
+struct Fp8Plan {
+    desc: CublasLtMatmulDesc,
+    a_layout: CublasLtMatrixLayout,
+    b_layout: CublasLtMatrixLayout,
+    c_layout: CublasLtMatrixLayout,
+    d_layout: CublasLtMatrixLayout,
+    algo: CublasLtMatmulAlgo,
+    workspace_size: usize,
+}
+
+struct CublasLtContext {
+    handle: CublasLtHandle,
+    workspace: GpuBuffer,
+    /// Persistent device scalar holding 1.0, used as the A/B scale pointer.
+    /// Scales are applied post-GEMM, so the in-GEMM scales stay 1.0.
+    one_buf: GpuBuffer,
+    /// Cache of prepared matmul plans keyed by (M, N, K, batch).
+    plans: HashMap<(usize, usize, usize, usize), Fp8Plan>,
+}
+
+impl CublasLtContext {
+    fn new() -> Self {
+        let mut handle = std::ptr::null_mut();
+        let status = unsafe { cublasLtCreate(&mut handle) };
+        assert_eq!(status, 0, "cublasLtCreate failed: {status}");
+        let workspace = GpuBuffer::alloc(WORKSPACE_BYTES).expect("alloc cublasLt workspace");
+        let mut one_buf = GpuBuffer::alloc(4).expect("alloc cublasLt fp8 scale");
+        one_buf
+            .copy_from_host(&1.0f32.to_le_bytes())
+            .expect("init fp8 scale");
+        Self {
+            handle,
+            workspace,
+            one_buf,
+            plans: HashMap::new(),
+        }
+    }
+
+    /// Get the cached strided-batched plan for (m, n, k, batch), building it on
+    /// first use.
+    fn plan(&mut self, m: usize, n: usize, k: usize, batch: usize) -> Fp8Plan {
+        if let Some(p) = self.plans.get(&(m, n, k, batch)) {
+            return *p;
+        }
+        let one_ptr = self.one_buf.as_ptr() as *const c_void;
+        let plan = unsafe { build_fp8_plan(self.handle, one_ptr, m, n, k, batch) };
+        self.plans.insert((m, n, k, batch), plan);
+        plan
+    }
+}
+
+impl Drop for CublasLtContext {
+    fn drop(&mut self) {
+        // Tear down cached plans before destroying the handle.
+        for (_, p) in self.plans.drain() {
+            unsafe {
+                cublasLtMatrixLayoutDestroy(p.a_layout);
+                cublasLtMatrixLayoutDestroy(p.b_layout);
+                cublasLtMatrixLayoutDestroy(p.c_layout);
+                cublasLtMatrixLayoutDestroy(p.d_layout);
+                cublasLtMatmulDescDestroy(p.desc);
+            }
+        }
+        if !self.handle.is_null() {
+            unsafe { cublasLtDestroy(self.handle) };
+        }
+    }
+}
+
+/// Build a strided-batched FP8 matmul plan for `batch` experts of one
+/// (m, n, k) shape. Row-major → cuBLASLt col-major mapping (transA=T,
+/// transB=N, m_lt=N, n_lt=M, k_lt=K). A/B scale pointers stay at 1.0 — both
+/// the per-expert weight scale and the per-token activation scale are applied
+/// post-GEMM in a fused kernel, which lets all experts run in one matmul.
+unsafe fn build_fp8_plan(
+    handle: CublasLtHandle,
+    one_ptr: *const c_void,
+    m: usize,
+    n: usize,
+    k: usize,
+    batch: usize,
+) -> Fp8Plan {
+    let m_lt = n as u64;
+    let n_lt = m as u64;
+    let k_lt = k as u64;
+
+    let mut desc: CublasLtMatmulDesc = std::ptr::null_mut();
+    cublasLtMatmulDescCreate(&mut desc, CUBLAS_COMPUTE_32F, CUDA_R_32F);
+
+    // transA=T (required for FP8 on Blackwell)
+    let trans_a: i32 = 1;
+    cublasLtMatmulDescSetAttribute(
+        desc,
+        CUBLASLT_MATMUL_DESC_TRANSA,
+        &trans_a as *const i32 as _,
+        4,
+    );
+    let ptr_sz = std::mem::size_of::<*const c_void>();
+    cublasLtMatmulDescSetAttribute(
+        desc,
+        CUBLASLT_MATMUL_DESC_A_SCALE_POINTER,
+        &one_ptr as *const _ as _,
+        ptr_sz,
+    );
+    cublasLtMatmulDescSetAttribute(
+        desc,
+        CUBLASLT_MATMUL_DESC_B_SCALE_POINTER,
+        &one_ptr as *const _ as _,
+        ptr_sz,
+    );
+
+    // Per-expert strides in ELEMENTS for the strided-batch layout.
+    let stride_a = (n * k) as i64; // weights [N, K]
+    let stride_b = (m * k) as i64; // activations [M, K]
+    let stride_c = (m * n) as i64; // output [M, N]
+    let bc = batch as i32;
+    let set_batch = |layout: CublasLtMatrixLayout, stride: i64| {
+        cublasLtMatrixLayoutSetAttribute(
+            layout,
+            CUBLASLT_MATRIX_LAYOUT_BATCH_COUNT,
+            &bc as *const i32 as _,
+            4,
+        );
+        cublasLtMatrixLayoutSetAttribute(
+            layout,
+            CUBLASLT_MATRIX_LAYOUT_STRIDED_BATCH_OFFSET,
+            &stride as *const i64 as _,
+            8,
+        );
+    };
+
+    // "A" layout (weights, transposed): physical (K, N) col-major, ld=K
+    let mut a_layout: CublasLtMatrixLayout = std::ptr::null_mut();
+    cublasLtMatrixLayoutCreate(&mut a_layout, CUDA_R_8F_E4M3, k_lt, m_lt, k as i64);
+    set_batch(a_layout, stride_a);
+    // "B" layout (activations): physical (K, M) col-major, ld=K
+    let mut b_layout: CublasLtMatrixLayout = std::ptr::null_mut();
+    cublasLtMatrixLayoutCreate(&mut b_layout, CUDA_R_8F_E4M3, k_lt, n_lt, k as i64);
+    set_batch(b_layout, stride_b);
+    // "C"/"D" layout (output): physical (N, M) col-major, ld=N
+    let mut c_layout: CublasLtMatrixLayout = std::ptr::null_mut();
+    cublasLtMatrixLayoutCreate(&mut c_layout, CUDA_R_16BF, m_lt, n_lt, m_lt as i64);
+    set_batch(c_layout, stride_c);
+    let mut d_layout: CublasLtMatrixLayout = std::ptr::null_mut();
+    cublasLtMatrixLayoutCreate(&mut d_layout, CUDA_R_16BF, m_lt, n_lt, m_lt as i64);
+    set_batch(d_layout, stride_c);
+
+    let mut pref: CublasLtMatmulPreference = std::ptr::null_mut();
+    cublasLtMatmulPreferenceCreate(&mut pref);
+    let ws_bytes = WORKSPACE_BYTES as u64;
+    cublasLtMatmulPreferenceSetAttribute(
+        pref,
+        CUBLASLT_MATMUL_PREF_MAX_WORKSPACE_BYTES,
+        &ws_bytes as *const u64 as _,
+        8,
+    );
+
+    let mut heuristic = std::mem::zeroed::<CublasLtMatmulHeuristicResult>();
+    let mut found: i32 = 0;
+    let status = cublasLtMatmulAlgoGetHeuristic(
+        handle,
+        desc,
+        a_layout,
+        b_layout,
+        c_layout,
+        d_layout,
+        pref,
+        1,
+        &mut heuristic,
+        &mut found,
+    );
+    assert!(
+        status == 0 && found > 0,
+        "cublasLtMatmulAlgoGetHeuristic failed for batched FP8 GEMM (m={m}, n={n}, k={k}, batch={batch}): status={status}, found={found}"
+    );
+    cublasLtMatmulPreferenceDestroy(pref);
+
+    Fp8Plan {
+        desc,
+        a_layout,
+        b_layout,
+        c_layout,
+        d_layout,
+        algo: heuristic.algo,
+        workspace_size: heuristic.workspace_size,
+    }
+}
+
+thread_local! {
+    static CUBLASLT_CTX: RefCell<CublasLtContext> = RefCell::new(CublasLtContext::new());
+}
+
+// ============================================================
+// Public API
+// ============================================================
+
+/// Dequantize a 3D FP8 E4M3 tensor to BF16 using per-expert FP32 scales.
+///
+/// src:    [num_experts, rows, cols] FP8E4M3, contiguous, GPU
+/// scales: [num_experts] F32, contiguous, GPU
+///
+/// Returns: [num_experts, rows, cols] BF16
+pub fn dequant_fp8_to_bf16(src: &Tensor, scales: &Tensor) -> Tensor {
+    assert_eq!(src.ndim(), 3, "dequant_fp8_to_bf16: src must be 3D");
+    assert_eq!(src.dtype(), DType::FP8E4M3);
+    assert!(src.is_contiguous());
+    assert_eq!(scales.ndim(), 1);
+    assert_eq!(scales.dtype(), DType::F32);
+    assert!(scales.is_contiguous());
+
+    let num_experts = src.shape()[0];
+    let rows = src.shape()[1];
+    let cols = src.shape()[2];
+    assert_eq!(scales.shape()[0], num_experts);
+
+    let out = Tensor::empty(&[num_experts, rows, cols], DType::BF16, src.device());
+
+    unsafe {
+        launch_dequant_fp8e4m3_to_bf16(
+            src.data_ptr() as *const c_void,
+            scales.data_ptr() as *const c_void,
+            out.data_ptr() as *mut c_void,
+            num_experts as i32,
+            rows as i32,
+            cols as i32,
+            xserv_cuda::current_stream_raw(),
+        );
+    }
+
+    out
+}
+
+/// Dynamically quantize a contiguous BF16 tensor to FP8 E4M3 with per-row scales.
+///
+/// src: [num_rows, cols] or [batch, rows, cols] BF16, contiguous, GPU
+/// Treats the tensor as 2D (flattens leading dims into num_rows).
+///
+/// Returns: (fp8_data [same shape] FP8E4M3, scales [total_rows] F32)
+pub fn quantize_bf16_to_fp8_rowwise(src: &Tensor) -> (Tensor, Tensor) {
+    assert_eq!(src.dtype(), DType::BF16);
+    assert!(src.is_contiguous());
+    assert!(src.ndim() >= 2);
+
+    let cols = src.shape()[src.ndim() - 1];
+    let num_rows: usize = src.shape()[..src.ndim() - 1].iter().product();
+
+    let fp8_out = Tensor::empty(src.shape(), DType::FP8E4M3, src.device());
+    let scales = Tensor::empty(&[num_rows], DType::F32, src.device());
+
+    unsafe {
+        launch_quantize_bf16_to_fp8e4m3_rowwise(
+            src.data_ptr() as *const c_void,
+            fp8_out.data_ptr() as *mut c_void,
+            scales.data_ptr() as *mut c_void,
+            num_rows as i32,
+            cols as i32,
+            xserv_cuda::current_stream_raw(),
+        );
+    }
+
+    (fp8_out, scales)
+}
+
+/// FP8 batched GEMM via cuBLASLt (transA=T required on Blackwell).
+///
+/// Computes: C[b] = scale_a[b] * scale_b[b] * (A_fp8[b] @ B_fp8_T[b]^T)
+/// effectively C[b] = A[b, M, K] @ W[b, K, N] but W is stored transposed
+/// as [b, N, K] for cuBLASLt FP8 compatibility.
+///
+/// a_fp8:    [batch, M, K] FP8E4M3  (activations, quantized per-row)
+/// a_scales: [batch * M] F32        (per-token activation scales, applied post-GEMM)
+/// b_fp8_t:  [batch, N, K] FP8E4M3  (weights, TRANSPOSED for cuBLASLt)
+/// b_scales: [batch] F32            (per-expert scalar weight scales, applied in-GEMM)
+///
+/// Returns:  [batch, M, N] BF16
+pub fn batched_gemm_fp8(
+    a_fp8: &Tensor,
+    a_scales: &Tensor,
+    b_fp8_t: &Tensor,
+    b_scales: &Tensor,
+) -> Tensor {
+    assert_eq!(a_fp8.ndim(), 3);
+    assert_eq!(b_fp8_t.ndim(), 3);
+    assert_eq!(a_fp8.dtype(), DType::FP8E4M3);
+    assert_eq!(b_fp8_t.dtype(), DType::FP8E4M3);
+    assert!(a_fp8.is_contiguous() && b_fp8_t.is_contiguous());
+    assert_eq!(a_fp8.shape()[0], b_fp8_t.shape()[0]);
+    // b_fp8_t is [batch, N, K] transposed, so b_fp8_t.shape[2] == K == a_fp8.shape[2]
+    assert_eq!(a_fp8.shape()[2], b_fp8_t.shape()[2]);
+
+    let batch = a_fp8.shape()[0];
+    let m = a_fp8.shape()[1]; // tokens
+    let k = a_fp8.shape()[2]; // hidden
+    let n = b_fp8_t.shape()[1]; // out_dim (from transposed weight)
+
+    // a_scales: [batch * M] per-token activation scales (applied post-GEMM, per row).
+    // b_scales: [batch] per-expert scalar weight scales (applied in-GEMM via B-scale ptr).
+    assert_eq!(a_scales.shape()[0], batch * m);
+    assert_eq!(b_scales.shape()[0], batch);
+
+    let c = Tensor::empty(&[batch, m, n], DType::BF16, a_fp8.device());
+
+    CUBLASLT_CTX.with(|cell| {
+        let mut ctx = cell.borrow_mut();
+        let handle = ctx.handle;
+        let ws_ptr = ctx.workspace.as_ptr() as *mut c_void;
+        // Cached strided-batched plan: heuristic + descriptor/layout creation
+        // happen once per (m, n, k, batch). All experts run in ONE matmul.
+        let plan = ctx.plan(m, n, k, batch);
+
+        // alpha=1, beta=0, in-GEMM scales=1.0. The unscaled result
+        //   D_raw[e] = A_fp8[e] @ B_fp8[e]^T
+        // is recovered to the real value by the fused post-scale kernel below.
+        let alpha: f32 = 1.0;
+        let beta: f32 = 0.0;
+
+        unsafe {
+            let status = cublasLtMatmul(
+                handle,
+                plan.desc,
+                &alpha as *const f32 as _,
+                b_fp8_t.data_ptr() as *const c_void, // cuBLASLt "A" = weights
+                plan.a_layout,
+                a_fp8.data_ptr() as *const c_void, // cuBLASLt "B" = activations
+                plan.b_layout,
+                &beta as *const f32 as _,
+                c.data_ptr() as *const c_void, // C (unused with beta=0)
+                plan.c_layout,
+                c.data_ptr() as *mut c_void, // D = output
+                plan.d_layout,
+                &plan.algo,
+                ws_ptr,
+                plan.workspace_size,
+                xserv_cuda::current_stream_raw(),
+            );
+            assert_eq!(
+                status, 0,
+                "batched cublasLtMatmul FP8 failed: status={status}"
+            );
+        }
+    });
+
+    // Post-GEMM: recover the real result in one pass.
+    //   c[e, t, :] *= a_scales[e*M + t] * b_scales[e]
+    // (per-token activation scale × per-expert weight scale). BF16's relative
+    // error is scale-invariant, so applying the scale here is precision-
+    // equivalent to folding it into the GEMM epilogue.
+    let total_rows = (batch * m) as i32;
+    unsafe {
+        launch_rowwise_scale_moe_bf16(
+            c.data_ptr() as *mut c_void,
+            a_scales.data_ptr() as *const c_void,
+            b_scales.data_ptr() as *const c_void,
+            total_rows,
+            n as i32,
+            m as i32,
+            xserv_cuda::current_stream_raw(),
+        );
+    }
+
+    c
+}
+
+// ============================================================
+// MXFP4 W4A16 (weight-only 4-bit) for MoE experts
+// ============================================================
+
+/// MXFP4 W4A16 batched GEMV for decode (M=1).
+///
+/// x:        [E, K] BF16 (per-expert activation; replicated across experts)
+/// w_packed: [E, N, K/2] byte tensor — two E2M1 nibbles per byte (lo = even k)
+/// w_scales: [E, N, K/32] byte tensor — UE8M0 scale per 32-element block
+///
+/// Returns:  [E, N] BF16, where y[e,n] = sum_k x[e,k] * dequant(W[e,n,k]).
+pub fn batched_gemv_mxfp4(
+    x: &Tensor,
+    w_packed: &Tensor,
+    w_scales: &Tensor,
+    n: usize,
+    k: usize,
+) -> Tensor {
+    assert_eq!(x.dtype(), DType::BF16);
+    assert!(x.is_contiguous());
+    let e = x.shape()[0];
+    assert_eq!(x.shape()[x.ndim() - 1], k, "GEMV K mismatch");
+
+    let y = Tensor::empty(&[e, n], DType::BF16, x.device());
+    unsafe {
+        launch_batched_gemv_mxfp4_bf16(
+            x.data_ptr() as *const c_void,
+            w_packed.data_ptr() as *const c_void,
+            w_scales.data_ptr() as *const c_void,
+            y.data_ptr() as *mut c_void,
+            e as i32,
+            n as i32,
+            k as i32,
+            xserv_cuda::current_stream_raw(),
+        );
+    }
+    y
+}
+
+/// Dequantize MXFP4 weights [E, N, K] → BF16 [E, K, N] for the prefill GEMM path
+/// (the BF16 batched GEMM expects weights as [E, K, N]).
+pub fn dequant_mxfp4_to_bf16_t(
+    w_packed: &Tensor,
+    w_scales: &Tensor,
+    e: usize,
+    n: usize,
+    k: usize,
+) -> Tensor {
+    let out = Tensor::empty(&[e, k, n], DType::BF16, w_packed.device());
+    unsafe {
+        launch_dequant_mxfp4_to_bf16_t(
+            w_packed.data_ptr() as *const c_void,
+            w_scales.data_ptr() as *const c_void,
+            out.data_ptr() as *mut c_void,
+            e as i32,
+            n as i32,
+            k as i32,
+            xserv_cuda::current_stream_raw(),
+        );
+    }
+    out
+}

crates/xserv-kernels/src/rmsnorm.rs

@@ -2,10 +2,35 @@ use std::ffi::c_void;
 use xserv_tensor::{DType, Device, Tensor};
 
 unsafe extern "C" {
-    fn launch_rmsnorm_f32(x: *const c_void, gamma: *const c_void, out: *mut c_void,
-                          rows: i32, hidden_size: i32, eps: f32, stream: *mut c_void);
-    fn launch_rmsnorm_bf16(x: *const c_void, gamma: *const c_void, out: *mut c_void,
-                           rows: i32, hidden_size: i32, eps: f32, stream: *mut c_void);
+    fn launch_rmsnorm_f32(
+        x: *const c_void,
+        gamma: *const c_void,
+        out: *mut c_void,
+        rows: i32,
+        hidden_size: i32,
+        eps: f32,
+        stream: *mut c_void,
+    );
+    fn launch_rmsnorm_bf16(
+        x: *const c_void,
+        gamma: *const c_void,
+        out: *mut c_void,
+        rows: i32,
+        hidden_size: i32,
+        eps: f32,
+        stream: *mut c_void,
+    );
+    fn launch_add_rmsnorm_bf16(
+        x: *const c_void,
+        residual: *const c_void,
+        gamma: *const c_void,
+        normed_out: *mut c_void,
+        sum_out: *mut c_void,
+        rows: i32,
+        hidden_size: i32,
+        eps: f32,
+        stream: *mut c_void,
+    );
 }
 
 pub fn rmsnorm(x: &Tensor, gamma: &Tensor, eps: f32) -> Tensor {
@@ -17,21 +42,82 @@ pub fn rmsnorm(x: &Tensor, gamma: &Tensor, eps: f32) -> Tensor {
     assert_eq!(x.dtype(), gamma.dtype());
 
     let rows = x.numel() / hidden_size;
-    let out = Tensor::zeros(x.shape(), x.dtype(), x.device());
+    assert!(
+        rows <= i32::MAX as usize,
+        "too many rows for i32 kernel param"
+    );
+    assert!(
+        hidden_size <= i32::MAX as usize,
+        "hidden_size too large for i32 kernel param"
+    );
+    let out = Tensor::empty(x.shape(), x.dtype(), x.device());
 
     unsafe {
         match x.dtype() {
             DType::F32 => launch_rmsnorm_f32(
-                x.data_ptr() as _, gamma.data_ptr() as _, out.data_ptr() as *mut c_void,
-                rows as i32, hidden_size as i32, eps, std::ptr::null_mut(),
+                x.data_ptr() as _,
+                gamma.data_ptr() as _,
+                out.data_ptr() as *mut c_void,
+                rows as i32,
+                hidden_size as i32,
+                eps,
+                xserv_cuda::current_stream_raw(),
             ),
             DType::BF16 => launch_rmsnorm_bf16(
-                x.data_ptr() as _, gamma.data_ptr() as _, out.data_ptr() as *mut c_void,
-                rows as i32, hidden_size as i32, eps, std::ptr::null_mut(),
+                x.data_ptr() as _,
+                gamma.data_ptr() as _,
+                out.data_ptr() as *mut c_void,
+                rows as i32,
+                hidden_size as i32,
+                eps,
+                xserv_cuda::current_stream_raw(),
             ),
             _ => panic!("unsupported dtype for rmsnorm"),
         }
     }
-    xserv_cuda::device::synchronize().unwrap();
     out
 }
+
+/// Fused Add + RMSNorm: computes sum = x + residual, then normed = rmsnorm(sum, gamma, eps).
+/// Returns (normed, sum). BF16 only.
+/// Saves one kernel launch and one full HBM round-trip per layer.
+pub fn add_rmsnorm(x: &Tensor, residual: &Tensor, gamma: &Tensor, eps: f32) -> (Tensor, Tensor) {
+    assert!(x.ndim() >= 1);
+    assert_eq!(x.shape(), residual.shape());
+    assert!(x.is_contiguous() && residual.is_contiguous() && gamma.is_contiguous());
+    assert!(matches!(x.device(), Device::Cuda(_)));
+    assert_eq!(x.dtype(), DType::BF16, "add_rmsnorm requires BF16");
+    assert_eq!(residual.dtype(), DType::BF16);
+    assert_eq!(gamma.dtype(), DType::BF16);
+
+    let hidden_size = *x.shape().last().unwrap();
+    assert_eq!(gamma.shape(), &[hidden_size]);
+
+    let rows = x.numel() / hidden_size;
+    assert!(
+        rows <= i32::MAX as usize,
+        "too many rows for i32 kernel param"
+    );
+    assert!(
+        hidden_size <= i32::MAX as usize,
+        "hidden_size too large for i32 kernel param"
+    );
+    let normed_out = Tensor::empty(x.shape(), DType::BF16, x.device());
+    let sum_out = Tensor::empty(x.shape(), DType::BF16, x.device());
+
+    unsafe {
+        launch_add_rmsnorm_bf16(
+            x.data_ptr() as *const c_void,
+            residual.data_ptr() as *const c_void,
+            gamma.data_ptr() as *const c_void,
+            normed_out.data_ptr() as *mut c_void,
+            sum_out.data_ptr() as *mut c_void,
+            rows as i32,
+            hidden_size as i32,
+            eps,
+            xserv_cuda::current_stream_raw(),
+        );
+    }
+
+    (normed_out, sum_out)
+}

crates/xserv-kernels/src/rope.rs

@@ -3,15 +3,34 @@ use xserv_cuda::GpuBuffer;
 use xserv_tensor::{DType, Device, Tensor};
 
 unsafe extern "C" {
-    fn launch_rope_f32(x: *mut c_void, cos_cache: *const c_void, sin_cache: *const c_void,
-                       positions: *const c_void, num_tokens: i32, num_heads: i32,
-                       head_dim: i32, stream: *mut c_void);
-    fn launch_rope_bf16(x: *mut c_void, cos_cache: *const c_void, sin_cache: *const c_void,
-                        positions: *const c_void, num_tokens: i32, num_heads: i32,
-                        head_dim: i32, stream: *mut c_void);
-    fn launch_compute_rope_cache(cos_cache: *mut c_void, sin_cache: *mut c_void,
-                                 max_seq_len: i32, half_dim: i32, theta: f32,
-                                 stream: *mut c_void);
+    fn launch_rope_f32(
+        x: *mut c_void,
+        cos_cache: *const c_void,
+        sin_cache: *const c_void,
+        positions: *const c_void,
+        num_tokens: i32,
+        num_heads: i32,
+        head_dim: i32,
+        stream: *mut c_void,
+    );
+    fn launch_rope_bf16(
+        x: *mut c_void,
+        cos_cache: *const c_void,
+        sin_cache: *const c_void,
+        positions: *const c_void,
+        num_tokens: i32,
+        num_heads: i32,
+        head_dim: i32,
+        stream: *mut c_void,
+    );
+    fn launch_compute_rope_cache(
+        cos_cache: *mut c_void,
+        sin_cache: *mut c_void,
+        max_seq_len: i32,
+        half_dim: i32,
+        theta: f32,
+        stream: *mut c_void,
+    );
 }
 
 pub struct RopeCache {
@@ -30,13 +49,99 @@ impl RopeCache {
 
         unsafe {
             launch_compute_rope_cache(
-                cos.as_mut_ptr() as _, sin.as_mut_ptr() as _,
-                max_seq_len as i32, half_dim as i32, theta, std::ptr::null_mut(),
+                cos.as_mut_ptr() as _,
+                sin.as_mut_ptr() as _,
+                max_seq_len as i32,
+                half_dim as i32,
+                theta,
+                xserv_cuda::current_stream_raw(),
             );
         }
-        xserv_cuda::device::synchronize().unwrap();
 
-        Self { cos, sin, max_seq_len, half_dim }
+        Self {
+            cos,
+            sin,
+            max_seq_len,
+            half_dim,
+        }
+    }
+
+    /// YaRN (Yet another RoPE extensioN) RoPE cache. Applies frequency-dependent
+    /// interpolation so the model can extrapolate beyond its training context.
+    pub fn new_yarn(
+        max_seq_len: usize,
+        head_dim: usize,
+        theta: f64,
+        factor: f64,
+        original_max_pos: usize,
+        beta_fast: f64,
+        beta_slow: f64,
+    ) -> Self {
+        let half_dim = head_dim / 2;
+        let dim = head_dim as f64;
+
+        // find_correction_dim: inverse formula to find dimension from number of rotations
+        let find_correction_dim = |num_rotations: f64| -> f64 {
+            dim * (original_max_pos as f64 / (num_rotations * 2.0 * std::f64::consts::PI)).ln()
+                / (2.0 * theta.ln())
+        };
+
+        let low_raw = find_correction_dim(beta_fast);
+        let high_raw = find_correction_dim(beta_slow);
+        // config has truncate=false, so use raw values (no floor/ceil)
+        let low = low_raw.max(0.0);
+        let high = high_raw.min((half_dim - 1) as f64);
+
+        // Compute inv_freq with YaRN interpolation
+        let mut inv_freq = vec![0.0f64; half_dim];
+        for i in 0..half_dim {
+            let pos_freq = theta.powf((2 * i) as f64 / dim);
+            let inv_freq_extrapolation = 1.0 / pos_freq; // original
+            let inv_freq_interpolation = 1.0 / (factor * pos_freq); // scaled
+
+            // Linear ramp: 0 where we keep original, 1 where we interpolate
+            let ramp = if (high - low).abs() < 0.001 {
+                0.5
+            } else {
+                ((i as f64 - low) / (high - low)).clamp(0.0, 1.0)
+            };
+            let extrapolation_factor = 1.0 - ramp;
+
+            inv_freq[i] = inv_freq_interpolation * (1.0 - extrapolation_factor)
+                + inv_freq_extrapolation * extrapolation_factor;
+        }
+
+        // Attention scaling factor for YaRN: 0.1 * ln(factor) + 1.0
+        let attn_factor = 0.1 * factor.ln() + 1.0;
+
+        // Build cos/sin cache on CPU then upload
+        let total = max_seq_len * half_dim;
+        let mut cos_host = vec![0.0f32; total];
+        let mut sin_host = vec![0.0f32; total];
+        for pos in 0..max_seq_len {
+            for i in 0..half_dim {
+                let angle = pos as f64 * inv_freq[i];
+                cos_host[pos * half_dim + i] = (angle.cos() * attn_factor) as f32;
+                sin_host[pos * half_dim + i] = (angle.sin() * attn_factor) as f32;
+            }
+        }
+
+        let nbytes = total * std::mem::size_of::<f32>();
+        let mut cos = GpuBuffer::alloc(nbytes).expect("alloc yarn cos_cache");
+        let mut sin = GpuBuffer::alloc(nbytes).expect("alloc yarn sin_cache");
+        let cos_bytes =
+            unsafe { std::slice::from_raw_parts(cos_host.as_ptr() as *const u8, nbytes) };
+        let sin_bytes =
+            unsafe { std::slice::from_raw_parts(sin_host.as_ptr() as *const u8, nbytes) };
+        cos.copy_from_host(cos_bytes).unwrap();
+        sin.copy_from_host(sin_bytes).unwrap();
+
+        Self {
+            cos,
+            sin,
+            max_seq_len,
+            half_dim,
+        }
     }
 }
 
@@ -59,27 +164,48 @@ pub fn rope_inplace(x: &Tensor, cache: &RopeCache, positions: &[u32]) {
             num_tokens * std::mem::size_of::<u32>(),
         )
     };
-    let mut pos_gpu = GpuBuffer::alloc(pos_bytes.len()).expect("alloc positions");
+    let mut pos_gpu =
+        xserv_cuda::allocator::cached_alloc(pos_bytes.len()).expect("alloc positions");
     pos_gpu.copy_from_host(pos_bytes).unwrap();
 
+    rope_inplace_device_pos(x, cache, pos_gpu.as_ptr() as *const c_void);
+}
+
+/// RoPE in-place with positions already on the GPU (u32, [num_tokens]).
+/// Used by the CUDA-graph decode path, where the position lives in a
+/// persistent device buffer updated outside the captured region.
+pub fn rope_inplace_device_pos(x: &Tensor, cache: &RopeCache, pos_gpu: *const c_void) {
+    assert_eq!(x.ndim(), 3);
+    assert!(x.is_contiguous());
+    assert!(matches!(x.device(), Device::Cuda(_)));
+    let num_tokens = x.shape()[0];
+    let num_heads = x.shape()[1];
+    let head_dim = x.shape()[2];
+    assert_eq!(head_dim / 2, cache.half_dim);
+
     unsafe {
         match x.dtype() {
             DType::F32 => launch_rope_f32(
                 x.data_ptr() as *mut c_void,
-                cache.cos.as_ptr() as _, cache.sin.as_ptr() as _,
-                pos_gpu.as_ptr() as _,
-                num_tokens as i32, num_heads as i32, head_dim as i32,
-                std::ptr::null_mut(),
+                cache.cos.as_ptr() as _,
+                cache.sin.as_ptr() as _,
+                pos_gpu,
+                num_tokens as i32,
+                num_heads as i32,
+                head_dim as i32,
+                xserv_cuda::current_stream_raw(),
             ),
             DType::BF16 => launch_rope_bf16(
                 x.data_ptr() as *mut c_void,
-                cache.cos.as_ptr() as _, cache.sin.as_ptr() as _,
-                pos_gpu.as_ptr() as _,
-                num_tokens as i32, num_heads as i32, head_dim as i32,
-                std::ptr::null_mut(),
+                cache.cos.as_ptr() as _,
+                cache.sin.as_ptr() as _,
+                pos_gpu,
+                num_tokens as i32,
+                num_heads as i32,
+                head_dim as i32,
+                xserv_cuda::current_stream_raw(),
             ),
             _ => panic!("unsupported dtype for rope"),
         }
     }
-    xserv_cuda::device::synchronize().unwrap();
 }

crates/xserv-kernels/src/softmax.rs

@@ -2,8 +2,20 @@ use std::ffi::c_void;
 use xserv_tensor::{DType, Device, Tensor};
 
 unsafe extern "C" {
-    fn launch_softmax_f32(x: *const c_void, out: *mut c_void, rows: i32, cols: i32, stream: *mut c_void);
-    fn launch_softmax_bf16(x: *const c_void, out: *mut c_void, rows: i32, cols: i32, stream: *mut c_void);
+    fn launch_softmax_f32(
+        x: *const c_void,
+        out: *mut c_void,
+        rows: i32,
+        cols: i32,
+        stream: *mut c_void,
+    );
+    fn launch_softmax_bf16(
+        x: *const c_void,
+        out: *mut c_void,
+        rows: i32,
+        cols: i32,
+        stream: *mut c_void,
+    );
 }
 
 /// Softmax along the last dimension.
@@ -14,21 +26,34 @@ pub fn softmax(x: &Tensor) -> Tensor {
 
     let cols = *x.shape().last().unwrap();
     let rows = x.numel() / cols;
-    let out = Tensor::zeros(x.shape(), x.dtype(), x.device());
+    assert!(
+        rows <= i32::MAX as usize,
+        "too many rows for i32 kernel param"
+    );
+    assert!(
+        cols <= i32::MAX as usize,
+        "cols too large for i32 kernel param"
+    );
+    let out = Tensor::empty(x.shape(), x.dtype(), x.device());
 
     unsafe {
         match x.dtype() {
             DType::F32 => launch_softmax_f32(
-                x.data_ptr() as _, out.data_ptr() as *mut c_void,
-                rows as i32, cols as i32, std::ptr::null_mut(),
+                x.data_ptr() as _,
+                out.data_ptr() as *mut c_void,
+                rows as i32,
+                cols as i32,
+                xserv_cuda::current_stream_raw(),
             ),
             DType::BF16 => launch_softmax_bf16(
-                x.data_ptr() as _, out.data_ptr() as *mut c_void,
-                rows as i32, cols as i32, std::ptr::null_mut(),
+                x.data_ptr() as _,
+                out.data_ptr() as *mut c_void,
+                rows as i32,
+                cols as i32,
+                xserv_cuda::current_stream_raw(),
             ),
             _ => panic!("unsupported dtype for softmax"),
         }
     }
-    xserv_cuda::device::synchronize().unwrap();
     out
 }

crates/xserv-kernels/src/transpose.rs

@@ -0,0 +1,256 @@
+use std::ffi::c_void;
+use xserv_tensor::{DType, Device, Tensor};
+
+unsafe extern "C" {
+    fn launch_reshape_heads_bf16(
+        inp: *const c_void,
+        out: *mut c_void,
+        seq_len: i32,
+        num_heads: i32,
+        head_dim: i32,
+        stream: *mut c_void,
+    );
+    fn launch_merge_heads_bf16(
+        inp: *const c_void,
+        out: *mut c_void,
+        seq_len: i32,
+        num_heads: i32,
+        head_dim: i32,
+        stream: *mut c_void,
+    );
+    fn launch_transpose_hsd_to_shd_bf16(
+        inp: *const c_void,
+        out: *mut c_void,
+        seq_len: i32,
+        num_heads: i32,
+        head_dim: i32,
+        stream: *mut c_void,
+    );
+    fn launch_transpose_shd_to_hsd_bf16(
+        inp: *const c_void,
+        out: *mut c_void,
+        seq_len: i32,
+        num_heads: i32,
+        head_dim: i32,
+        stream: *mut c_void,
+    );
+    fn launch_repeat_kv_bf16(
+        inp: *const c_void,
+        out: *mut c_void,
+        kv_heads: i32,
+        n_rep: i32,
+        seq_len: i32,
+        head_dim: i32,
+        stream: *mut c_void,
+    );
+    fn launch_strided_copy_bf16(
+        inp: *const c_void,
+        out: *mut c_void,
+        numel: i32,
+        ndim: i32,
+        shape0: i32,
+        shape1: i32,
+        shape2: i32,
+        shape3: i32,
+        in_stride0: i32,
+        in_stride1: i32,
+        in_stride2: i32,
+        in_stride3: i32,
+        in_offset: i32,
+        stream: *mut c_void,
+    );
+    fn launch_strided_copy_f32(
+        inp: *const c_void,
+        out: *mut c_void,
+        numel: i32,
+        ndim: i32,
+        shape0: i32,
+        shape1: i32,
+        shape2: i32,
+        shape3: i32,
+        in_stride0: i32,
+        in_stride1: i32,
+        in_stride2: i32,
+        in_stride3: i32,
+        in_offset: i32,
+        stream: *mut c_void,
+    );
+}
+
+/// [S, H*D] → [1, H, S, D] on GPU (BF16)
+pub fn reshape_heads_gpu(x: &Tensor, seq_len: usize, num_heads: usize, head_dim: usize) -> Tensor {
+    assert_eq!(x.dtype(), DType::BF16);
+    assert!(x.is_contiguous() && matches!(x.device(), Device::Cuda(_)));
+    let out = Tensor::empty(&[1, num_heads, seq_len, head_dim], DType::BF16, x.device());
+    unsafe {
+        launch_reshape_heads_bf16(
+            x.data_ptr() as _,
+            out.data_ptr() as *mut c_void,
+            seq_len as i32,
+            num_heads as i32,
+            head_dim as i32,
+            xserv_cuda::current_stream_raw(),
+        );
+    }
+    out
+}
+
+/// [1, H, S, D] → [S, H*D] on GPU (BF16)
+pub fn merge_heads_gpu(x: &Tensor, seq_len: usize, num_heads: usize, head_dim: usize) -> Tensor {
+    assert_eq!(x.dtype(), DType::BF16);
+    assert!(x.is_contiguous() && matches!(x.device(), Device::Cuda(_)));
+    let hidden = num_heads * head_dim;
+    let out = Tensor::empty(&[seq_len, hidden], DType::BF16, x.device());
+    unsafe {
+        launch_merge_heads_bf16(
+            x.data_ptr() as _,
+            out.data_ptr() as *mut c_void,
+            seq_len as i32,
+            num_heads as i32,
+            head_dim as i32,
+            xserv_cuda::current_stream_raw(),
+        );
+    }
+    out
+}
+
+/// [1, H, S, D] → [S, H, D] for RoPE on GPU (BF16)
+pub fn transpose_for_rope_gpu(
+    x: &Tensor,
+    seq_len: usize,
+    num_heads: usize,
+    head_dim: usize,
+) -> Tensor {
+    assert_eq!(x.dtype(), DType::BF16);
+    assert!(x.is_contiguous() && matches!(x.device(), Device::Cuda(_)));
+    let out = Tensor::empty(&[seq_len, num_heads, head_dim], DType::BF16, x.device());
+    unsafe {
+        launch_transpose_hsd_to_shd_bf16(
+            x.data_ptr() as _,
+            out.data_ptr() as *mut c_void,
+            seq_len as i32,
+            num_heads as i32,
+            head_dim as i32,
+            xserv_cuda::current_stream_raw(),
+        );
+    }
+    out
+}
+
+/// [S, H, D] → [1, H, S, D] after RoPE on GPU (BF16)
+pub fn transpose_from_rope_gpu(
+    x: &Tensor,
+    seq_len: usize,
+    num_heads: usize,
+    head_dim: usize,
+) -> Tensor {
+    assert_eq!(x.dtype(), DType::BF16);
+    assert!(x.is_contiguous() && matches!(x.device(), Device::Cuda(_)));
+    let out = Tensor::empty(&[1, num_heads, seq_len, head_dim], DType::BF16, x.device());
+    unsafe {
+        launch_transpose_shd_to_hsd_bf16(
+            x.data_ptr() as _,
+            out.data_ptr() as *mut c_void,
+            seq_len as i32,
+            num_heads as i32,
+            head_dim as i32,
+            xserv_cuda::current_stream_raw(),
+        );
+    }
+    out
+}
+
+/// [1, KV_H, S, D] → [1, KV_H*n_rep, S, D] on GPU (BF16)
+pub fn repeat_kv_gpu(x: &Tensor, n_rep: usize) -> Tensor {
+    if n_rep == 1 {
+        return x.clone();
+    }
+    assert_eq!(x.dtype(), DType::BF16);
+    assert!(x.is_contiguous() && matches!(x.device(), Device::Cuda(_)));
+    let kv_heads = x.shape()[1];
+    let seq_len = x.shape()[2];
+    let head_dim = x.shape()[3];
+    let new_heads = kv_heads * n_rep;
+    let out = Tensor::empty(&[1, new_heads, seq_len, head_dim], DType::BF16, x.device());
+    unsafe {
+        launch_repeat_kv_bf16(
+            x.data_ptr() as _,
+            out.data_ptr() as *mut c_void,
+            kv_heads as i32,
+            n_rep as i32,
+            seq_len as i32,
+            head_dim as i32,
+            xserv_cuda::current_stream_raw(),
+        );
+    }
+    out
+}
+
+/// Make a non-contiguous GPU tensor contiguous via a strided copy kernel.
+/// Supports BF16 and F32, up to 4D tensors (padded to 4D internally).
+pub fn strided_to_contiguous_gpu(x: &Tensor) -> Tensor {
+    assert!(matches!(x.device(), Device::Cuda(_)), "expected GPU tensor");
+    assert!(!x.is_contiguous(), "tensor is already contiguous");
+    assert!(x.ndim() <= 4, "strided_to_contiguous_gpu supports up to 4D");
+
+    let ndim = x.ndim();
+    let numel = x.numel();
+
+    // Pad shape and strides to 4D (prepend 1s for shape, 0s for strides)
+    let mut shape4 = [1i32; 4];
+    let mut strides4 = [0i32; 4];
+    let pad = 4 - ndim;
+    for i in 0..ndim {
+        shape4[pad + i] = x.shape()[i] as i32;
+        strides4[pad + i] = x.strides()[i] as i32;
+    }
+
+    let out = Tensor::empty(x.shape(), x.dtype(), x.device());
+
+    // Use storage base pointer + element offset, because strides are relative to
+    // element 0 of the storage, not the data_ptr() (which already adds byte offset).
+    let storage_ptr = x.storage().gpu_buffer().as_ptr();
+    let in_offset = x.offset() as i32;
+
+    unsafe {
+        match x.dtype() {
+            DType::BF16 => launch_strided_copy_bf16(
+                storage_ptr as _,
+                out.data_ptr() as *mut c_void,
+                numel as i32,
+                ndim as i32,
+                shape4[0],
+                shape4[1],
+                shape4[2],
+                shape4[3],
+                strides4[0],
+                strides4[1],
+                strides4[2],
+                strides4[3],
+                in_offset,
+                xserv_cuda::current_stream_raw(),
+            ),
+            DType::F32 => launch_strided_copy_f32(
+                storage_ptr as _,
+                out.data_ptr() as *mut c_void,
+                numel as i32,
+                ndim as i32,
+                shape4[0],
+                shape4[1],
+                shape4[2],
+                shape4[3],
+                strides4[0],
+                strides4[1],
+                strides4[2],
+                strides4[3],
+                in_offset,
+                xserv_cuda::current_stream_raw(),
+            ),
+            _ => panic!(
+                "strided_to_contiguous_gpu: unsupported dtype {:?}",
+                x.dtype()
+            ),
+        }
+    }
+    out
+}

crates/xserv-kernels/tests/attention_test.rs

@@ -0,0 +1,232 @@
+use xserv_kernels::*;
+use xserv_tensor::{Device, Tensor};
+
+fn init() {
+    xserv_cuda::device::set_device(0).unwrap();
+}
+
+fn cpu_attention(
+    q: &[f32],
+    k: &[f32],
+    v: &[f32],
+    batch: usize,
+    heads: usize,
+    q_len: usize,
+    kv_len: usize,
+    head_dim: usize,
+    causal: bool,
+) -> Vec<f32> {
+    let mut out = vec![0.0f32; batch * heads * q_len * head_dim];
+    let scale = 1.0 / (head_dim as f32).sqrt();
+
+    for b in 0..batch {
+        for h in 0..heads {
+            // scores = Q @ K^T, scaled
+            let mut scores = vec![0.0f32; q_len * kv_len];
+            for i in 0..q_len {
+                for j in 0..kv_len {
+                    let mut s = 0.0f32;
+                    for d in 0..head_dim {
+                        let qi = q[((b * heads + h) * q_len + i) * head_dim + d];
+                        let ki = k[((b * heads + h) * kv_len + j) * head_dim + d];
+                        s += qi * ki;
+                    }
+                    scores[i * kv_len + j] = s * scale;
+                }
+            }
+            // causal mask
+            if causal {
+                let offset = kv_len - q_len;
+                for i in 0..q_len {
+                    for j in 0..kv_len {
+                        if j > i + offset {
+                            scores[i * kv_len + j] = f32::NEG_INFINITY;
+                        }
+                    }
+                }
+            }
+            // softmax per row
+            for i in 0..q_len {
+                let row = &mut scores[i * kv_len..(i + 1) * kv_len];
+                let max = row.iter().cloned().fold(f32::NEG_INFINITY, f32::max);
+                let mut sum = 0.0f32;
+                for v in row.iter_mut() {
+                    *v = (*v - max).exp();
+                    sum += *v;
+                }
+                for v in row.iter_mut() {
+                    *v /= sum;
+                }
+            }
+            // output = weights @ V
+            for i in 0..q_len {
+                for d in 0..head_dim {
+                    let mut s = 0.0f32;
+                    for j in 0..kv_len {
+                        let w = scores[i * kv_len + j];
+                        let vi = v[((b * heads + h) * kv_len + j) * head_dim + d];
+                        s += w * vi;
+                    }
+                    out[((b * heads + h) * q_len + i) * head_dim + d] = s;
+                }
+            }
+        }
+    }
+    out
+}
+
+fn check_close(a: &[f32], b: &[f32], atol: f32, name: &str) {
+    assert_eq!(a.len(), b.len(), "{name}: length mismatch");
+    let mut max_err = 0.0f32;
+    for (i, (x, y)) in a.iter().zip(b).enumerate() {
+        let err = (x - y).abs();
+        if err > max_err {
+            max_err = err;
+        }
+        assert!(
+            err <= atol,
+            "{name}: mismatch at [{i}]: got {x}, expected {y}, err {err}"
+        );
+    }
+    println!("{name}: max_err = {max_err:.6e}");
+}
+
+fn make_data(n: usize) -> Vec<f32> {
+    (0..n).map(|i| ((i % 17) as f32 - 8.0) * 0.05).collect()
+}
+
+#[test]
+fn test_batched_matmul() {
+    init();
+    let batch = 4;
+    let heads = 8;
+    let m = 32;
+    let k = 64;
+    let n = 32;
+
+    let a_data = make_data(batch * heads * m * k);
+    let b_data = make_data(batch * heads * k * n);
+
+    let a = Tensor::from_slice(&a_data, &[batch, heads, m, k]).to_device(Device::Cuda(0));
+    let b = Tensor::from_slice(&b_data, &[batch, heads, k, n]).to_device(Device::Cuda(0));
+    let c = batched_matmul(&a, &b).to_device(Device::Cpu);
+
+    assert_eq!(c.shape(), &[batch, heads, m, n]);
+
+    // Verify one batch element
+    let a_cpu = &a_data[0..m * k];
+    let b_cpu = &b_data[0..k * n];
+    let mut expected = vec![0.0f32; m * n];
+    for i in 0..m {
+        for j in 0..n {
+            let mut s = 0.0f32;
+            for kk in 0..k {
+                s += a_cpu[i * k + kk] * b_cpu[kk * n + j];
+            }
+            expected[i * n + j] = s;
+        }
+    }
+    let result = c.as_slice::<f32>();
+    check_close(&result[0..m * n], &expected, 1e-3, "batched_matmul[0]");
+}
+
+#[test]
+fn test_attention_no_causal() {
+    init();
+    let b = 1;
+    let h = 2;
+    let s = 8;
+    let d = 16;
+    let q_data = make_data(b * h * s * d);
+    let k_data = make_data(b * h * s * d);
+    let v_data = make_data(b * h * s * d);
+    let expected = cpu_attention(&q_data, &k_data, &v_data, b, h, s, s, d, false);
+
+    let q = Tensor::from_slice(&q_data, &[b, h, s, d]).to_device(Device::Cuda(0));
+    let k = Tensor::from_slice(&k_data, &[b, h, s, d]).to_device(Device::Cuda(0));
+    let v = Tensor::from_slice(&v_data, &[b, h, s, d]).to_device(Device::Cuda(0));
+    let out = attention(&q, &k, &v, false).to_device(Device::Cpu);
+    check_close(
+        out.as_slice::<f32>(),
+        &expected,
+        1e-4,
+        "attention_no_causal",
+    );
+}
+
+#[test]
+fn test_attention_causal() {
+    init();
+    let b = 1;
+    let h = 2;
+    let s = 16;
+    let d = 32;
+    let q_data = make_data(b * h * s * d);
+    let k_data = make_data(b * h * s * d);
+    let v_data = make_data(b * h * s * d);
+    let expected = cpu_attention(&q_data, &k_data, &v_data, b, h, s, s, d, true);
+
+    let q = Tensor::from_slice(&q_data, &[b, h, s, d]).to_device(Device::Cuda(0));
+    let k = Tensor::from_slice(&k_data, &[b, h, s, d]).to_device(Device::Cuda(0));
+    let v = Tensor::from_slice(&v_data, &[b, h, s, d]).to_device(Device::Cuda(0));
+    let out = attention(&q, &k, &v, true).to_device(Device::Cpu);
+    check_close(out.as_slice::<f32>(), &expected, 1e-3, "attention_causal");
+}
+
+#[test]
+fn test_attention_causal_larger() {
+    init();
+    let b = 2;
+    let h = 4;
+    let s = 64;
+    let d = 64;
+    let q_data = make_data(b * h * s * d);
+    let k_data = make_data(b * h * s * d);
+    let v_data = make_data(b * h * s * d);
+    let expected = cpu_attention(&q_data, &k_data, &v_data, b, h, s, s, d, true);
+
+    let q = Tensor::from_slice(&q_data, &[b, h, s, d]).to_device(Device::Cuda(0));
+    let k = Tensor::from_slice(&k_data, &[b, h, s, d]).to_device(Device::Cuda(0));
+    let v = Tensor::from_slice(&v_data, &[b, h, s, d]).to_device(Device::Cuda(0));
+    let out = attention(&q, &k, &v, true).to_device(Device::Cpu);
+    check_close(
+        out.as_slice::<f32>(),
+        &expected,
+        1e-2,
+        "attention_causal_larger",
+    );
+}
+
+#[test]
+fn test_attention_causal_first_row_sees_only_first_token() {
+    init();
+    let b = 1;
+    let h = 1;
+    let s = 4;
+    let d = 8;
+    let q_data = make_data(b * h * s * d);
+    let k_data = make_data(b * h * s * d);
+    let v_data: Vec<f32> = (0..s * d)
+        .map(|i| {
+            if i < d { 1.0 } else { 0.0 } // only first V row is nonzero
+        })
+        .collect();
+
+    let q = Tensor::from_slice(&q_data, &[b, h, s, d]).to_device(Device::Cuda(0));
+    let k = Tensor::from_slice(&k_data, &[b, h, s, d]).to_device(Device::Cuda(0));
+    let v = Tensor::from_slice(&v_data, &[b, h, s, d]).to_device(Device::Cuda(0));
+    let out = attention(&q, &k, &v, true).to_device(Device::Cpu);
+
+    // First row (position 0) with causal mask can only see position 0.
+    // So attention weight for position 0 is 1.0 for token 0 only.
+    // output[0] should be exactly V[0] = [1, 1, 1, ...1]
+    let result = out.as_slice::<f32>();
+    for i in 0..d {
+        assert!(
+            (result[i] - 1.0).abs() < 1e-5,
+            "first row should equal V[0], got {} at dim {}",
+            result[i],
+            i
+        );
+    }
+}

crates/xserv-kernels/tests/gemm_test.rs

@@ -1,5 +1,5 @@
 use half::bf16;
-use xserv_kernels::{matmul, GemmBackend};
+use xserv_kernels::{GemmBackend, matmul};
 use xserv_tensor::{Device, Tensor};
 
 fn cpu_matmul_f32(a: &[f32], b: &[f32], m: usize, n: usize, k: usize) -> Vec<f32> {
@@ -75,56 +75,110 @@ fn run_gemm_test_bf16(backend: GemmBackend, m: usize, n: usize, k: usize) {
 // --- F32 tests ---
 
 #[test]
-fn test_gemm_naive_f32_small() { run_gemm_test_f32(GemmBackend::Naive, 4, 4, 4); }
+fn test_gemm_naive_f32_small() {
+    run_gemm_test_f32(GemmBackend::Naive, 4, 4, 4);
+}
 
 #[test]
-fn test_gemm_naive_f32_medium() { run_gemm_test_f32(GemmBackend::Naive, 64, 64, 64); }
+fn test_gemm_naive_f32_medium() {
+    run_gemm_test_f32(GemmBackend::Naive, 64, 64, 64);
+}
 
 #[test]
-fn test_gemm_naive_f32_rect() { run_gemm_test_f32(GemmBackend::Naive, 32, 64, 48); }
+fn test_gemm_naive_f32_rect() {
+    run_gemm_test_f32(GemmBackend::Naive, 32, 64, 48);
+}
 
 #[test]
-fn test_gemm_tiled_f32_small() { run_gemm_test_f32(GemmBackend::Tiled, 4, 4, 4); }
+fn test_gemm_tiled_f32_small() {
+    run_gemm_test_f32(GemmBackend::Tiled, 4, 4, 4);
+}
 
 #[test]
-fn test_gemm_tiled_f32_medium() { run_gemm_test_f32(GemmBackend::Tiled, 128, 128, 128); }
+fn test_gemm_tiled_f32_medium() {
+    run_gemm_test_f32(GemmBackend::Tiled, 128, 128, 128);
+}
 
 #[test]
-fn test_gemm_tiled_f32_rect() { run_gemm_test_f32(GemmBackend::Tiled, 65, 33, 97); }
+fn test_gemm_tiled_f32_rect() {
+    run_gemm_test_f32(GemmBackend::Tiled, 65, 33, 97);
+}
 
 #[test]
-fn test_gemm_cublas_f32_small() { run_gemm_test_f32(GemmBackend::CuBlas, 4, 4, 4); }
+fn test_gemm_cublas_f32_small() {
+    run_gemm_test_f32(GemmBackend::CuBlas, 4, 4, 4);
+}
 
 #[test]
-fn test_gemm_cublas_f32_medium() { run_gemm_test_f32(GemmBackend::CuBlas, 256, 256, 256); }
+fn test_gemm_cublas_f32_medium() {
+    run_gemm_test_f32(GemmBackend::CuBlas, 256, 256, 256);
+}
 
 #[test]
-fn test_gemm_cublas_f32_rect() { run_gemm_test_f32(GemmBackend::CuBlas, 65, 33, 97); }
+fn test_gemm_cublas_f32_rect() {
+    run_gemm_test_f32(GemmBackend::CuBlas, 65, 33, 97);
+}
 
 // --- BF16 tests ---
 
 #[test]
-fn test_gemm_naive_bf16_small() { run_gemm_test_bf16(GemmBackend::Naive, 4, 4, 4); }
+fn test_gemm_naive_bf16_small() {
+    run_gemm_test_bf16(GemmBackend::Naive, 4, 4, 4);
+}
 
 #[test]
-fn test_gemm_naive_bf16_medium() { run_gemm_test_bf16(GemmBackend::Naive, 64, 64, 64); }
+fn test_gemm_naive_bf16_medium() {
+    run_gemm_test_bf16(GemmBackend::Naive, 64, 64, 64);
+}
 
 #[test]
-fn test_gemm_tiled_bf16_small() { run_gemm_test_bf16(GemmBackend::Tiled, 4, 4, 4); }
+fn test_gemm_tiled_bf16_small() {
+    run_gemm_test_bf16(GemmBackend::Tiled, 4, 4, 4);
+}
 
 #[test]
-fn test_gemm_tiled_bf16_medium() { run_gemm_test_bf16(GemmBackend::Tiled, 128, 128, 128); }
+fn test_gemm_tiled_bf16_medium() {
+    run_gemm_test_bf16(GemmBackend::Tiled, 128, 128, 128);
+}
 
 #[test]
-fn test_gemm_cublas_bf16_small() { run_gemm_test_bf16(GemmBackend::CuBlas, 4, 4, 4); }
+fn test_gemm_cublas_bf16_small() {
+    run_gemm_test_bf16(GemmBackend::CuBlas, 4, 4, 4);
+}
 
 #[test]
-fn test_gemm_cublas_bf16_medium() { run_gemm_test_bf16(GemmBackend::CuBlas, 256, 256, 256); }
+fn test_gemm_cublas_bf16_medium() {
+    run_gemm_test_bf16(GemmBackend::CuBlas, 256, 256, 256);
+}
+
+// --- Custom GEMV tests (M=1, BF16 fast path) ---
+
+#[test]
+fn test_gemv_bf16_small() {
+    run_gemm_test_bf16(GemmBackend::CuBlas, 1, 64, 64);
+}
+
+#[test]
+fn test_gemv_bf16_medium() {
+    run_gemm_test_bf16(GemmBackend::CuBlas, 1, 256, 256);
+}
+
+#[test]
+fn test_gemv_bf16_4096() {
+    run_gemm_test_bf16(GemmBackend::CuBlas, 1, 4096, 4096);
+}
+
+#[test]
+fn test_gemv_bf16_rect() {
+    run_gemm_test_bf16(GemmBackend::CuBlas, 1, 512, 4096);
+}
 
 // --- Larger benchmark-style tests ---
 
 #[test]
-fn test_gemm_cublas_f32_1024() { run_gemm_test_f32(GemmBackend::CuBlas, 1024, 1024, 1024); }
+fn test_gemm_cublas_f32_1024() {
+    run_gemm_test_f32(GemmBackend::CuBlas, 1024, 1024, 1024);
+}
 
 #[test]
 fn test_gemm_consistency_all_backends() {

crates/xserv-kernels/tests/ops_test.rs

@@ -2,7 +2,9 @@ use half::bf16;
 use xserv_kernels::*;
 use xserv_tensor::{Device, Tensor};
 
-fn init() { xserv_cuda::device::set_device(0).unwrap(); }
+fn init() {
+    xserv_cuda::device::set_device(0).unwrap();
+}
 
 // --- CPU reference implementations ---
 
@@ -37,10 +39,12 @@ fn cpu_layernorm(x: &[f32], gamma: &[f32], beta: &[f32], eps: f32, hidden: usize
 
 fn cpu_gelu(x: &[f32]) -> Vec<f32> {
     let sqrt_2_over_pi = 0.7978845608f32;
-    x.iter().map(|&v| {
-        let inner = sqrt_2_over_pi * (v + 0.044715 * v * v * v);
-        0.5 * v * (1.0 + inner.tanh())
-    }).collect()
+    x.iter()
+        .map(|&v| {
+            let inner = sqrt_2_over_pi * (v + 0.044715 * v * v * v);
+            0.5 * v * (1.0 + inner.tanh())
+        })
+        .collect()
 }
 
 fn cpu_silu(x: &[f32]) -> Vec<f32> {
@@ -74,10 +78,10 @@ fn cpu_rope(x: &mut [f32], positions: &[u32], num_heads: usize, head_dim: usize,
                 let cos_val = angle.cos();
                 let sin_val = angle.sin();
                 let base = (t * num_heads + h) * head_dim;
-                let x0 = x[base + 2 * i];
-                let x1 = x[base + 2 * i + 1];
-                x[base + 2 * i]     = x0 * cos_val - x1 * sin_val;
-                x[base + 2 * i + 1] = x0 * sin_val + x1 * cos_val;
+                let x0 = x[base + i];
+                let x1 = x[base + i + half_dim];
+                x[base + i] = x0 * cos_val - x1 * sin_val;
+                x[base + i + half_dim] = x1 * cos_val + x0 * sin_val;
             }
         }
     }
@@ -88,8 +92,13 @@ fn check_close(result: &[f32], expected: &[f32], atol: f32, name: &str) {
     let mut max_err = 0.0f32;
     for (i, (r, e)) in result.iter().zip(expected).enumerate() {
         let err = (r - e).abs();
-        if err > max_err { max_err = err; }
-        assert!(err <= atol, "{name}: mismatch at [{i}]: got {r}, expected {e}, err {err}");
+        if err > max_err {
+            max_err = err;
+        }
+        assert!(
+            err <= atol,
+            "{name}: mismatch at [{i}]: got {r}, expected {e}, err {err}"
+        );
     }
     println!("{name}: max_err = {max_err:.6e}");
 }
@@ -208,13 +217,18 @@ fn test_softmax_sum_to_one() {
     init();
     let rows = 4;
     let cols = 2048;
-    let data: Vec<f32> = (0..rows * cols).map(|i| ((i % 31) as f32 - 15.0) * 0.5).collect();
+    let data: Vec<f32> = (0..rows * cols)
+        .map(|i| ((i % 31) as f32 - 15.0) * 0.5)
+        .collect();
     let x = Tensor::from_slice(&data, &[rows, cols]).to_device(Device::Cuda(0));
     let out = softmax(&x).to_device(Device::Cpu);
     let result = out.as_slice::<f32>();
     for r in 0..rows {
         let row_sum: f32 = result[r * cols..(r + 1) * cols].iter().sum();
-        assert!((row_sum - 1.0).abs() < 1e-5, "softmax row {r} sum = {row_sum}");
+        assert!(
+            (row_sum - 1.0).abs() < 1e-5,
+            "softmax row {r} sum = {row_sum}"
+        );
     }
 }
 
@@ -247,8 +261,10 @@ fn test_embedding_f32() {
         for i in 0..hidden {
             let expected = table_data[tid as usize * hidden + i];
             let got = result[seq_idx * hidden + i];
-            assert!((got - expected).abs() < 1e-6,
-                "embedding mismatch at [{seq_idx},{i}]: got {got}, expected {expected}");
+            assert!(
+                (got - expected).abs() < 1e-6,
+                "embedding mismatch at [{seq_idx},{i}]: got {got}, expected {expected}"
+            );
         }
     }
 }
@@ -270,8 +286,8 @@ fn test_rope_f32() {
     let mut expected = x_data.clone();
     cpu_rope(&mut expected, &positions, num_heads, head_dim, theta);
 
-    let x = Tensor::from_slice(&x_data, &[num_tokens, num_heads, head_dim])
-        .to_device(Device::Cuda(0));
+    let x =
+        Tensor::from_slice(&x_data, &[num_tokens, num_heads, head_dim]).to_device(Device::Cuda(0));
     let cache = RopeCache::new(64, head_dim, theta);
     rope_inplace(&x, &cache, &positions);
 
@@ -292,8 +308,8 @@ fn test_rope_position_0_identity() {
         .map(|i| (i as f32 + 1.0) * 0.1)
         .collect();
 
-    let x = Tensor::from_slice(&x_data, &[num_tokens, num_heads, head_dim])
-        .to_device(Device::Cuda(0));
+    let x =
+        Tensor::from_slice(&x_data, &[num_tokens, num_heads, head_dim]).to_device(Device::Cuda(0));
     let cache = RopeCache::new(64, head_dim, 10000.0);
     rope_inplace(&x, &cache, &positions);
 

crates/xserv-model/Cargo.toml

@@ -0,0 +1,18 @@
+[package]
+name = "xserv-model"
+version.workspace = true
+edition.workspace = true
+
+[dependencies]
+xserv-cuda = { path = "../xserv-cuda" }
+xserv-tensor = { path = "../xserv-tensor" }
+xserv-kernels = { path = "../xserv-kernels" }
+xserv-tokenizer = { path = "../xserv-tokenizer" }
+xserv-distributed = { path = "../xserv-distributed" }
+half.workspace = true
+libc.workspace = true
+smallvec.workspace = true
+serde.workspace = true
+serde_json.workspace = true
+safetensors.workspace = true
+rand.workspace = true

crates/xserv-model/src/bin/bench-gpt-oss.rs

@@ -0,0 +1,421 @@
+use std::path::PathBuf;
+use std::sync::Arc;
+use std::time::Instant;
+
+use xserv_distributed::{TpContext, UniqueId, get_unique_id};
+use xserv_model::{BLOCK_SIZE, GptOss, GraphedGptOssDecoder, ModelConfig, PagedKVCache, loader};
+use xserv_tensor::{DType, Device};
+use xserv_tokenizer::Tokenizer;
+
+fn main() {
+    let args: Vec<String> = std::env::args().collect();
+    if args.len() < 2 {
+        eprintln!("Usage: bench-gpt-oss <model-dir> [--max-tokens N] [--tp N]");
+        std::process::exit(1);
+    }
+
+    let model_dir = PathBuf::from(&args[1]);
+    let max_tokens: usize = get_arg(&args, "--max-tokens").unwrap_or(32);
+    let world: usize = get_arg(&args, "--tp").unwrap_or(2);
+
+    let config = ModelConfig::from_file(&model_dir.join("config.json"));
+    let tokenizer = Tokenizer::from_file(&model_dir.join("tokenizer.json"));
+
+    eprintln!(
+        "gpt-oss-20b: layers={}, hidden={}, heads={}/{} kv, experts={}, top_k={}, vocab={}",
+        config.num_layers(),
+        config.hidden(),
+        config.num_heads(),
+        config.num_kv_heads(),
+        config.num_experts(),
+        config.experts_per_token(),
+        config.vocab_size
+    );
+    eprintln!("TP world={world}, max_tokens={max_tokens}");
+
+    let max_seq_len: usize = 2048;
+    let max_blocks_per_seq = (max_seq_len + BLOCK_SIZE - 1) / BLOCK_SIZE;
+
+    // TP setup
+    let uid = get_unique_id();
+    let local_kv = config.num_kv_heads() / world;
+
+    // Spawn worker threads for ranks 1..world
+    let mut worker_handles = Vec::new();
+    let mut worker_txs = Vec::new();
+    for rank in 1..world {
+        let (tx, rx) = std::sync::mpsc::channel::<WorkerCmd>();
+        let (ack_tx, ack_rx) = std::sync::mpsc::channel::<()>();
+        let cfg = config.clone();
+        let md = model_dir.clone();
+        let uid_copy = uid;
+        worker_handles.push((
+            std::thread::spawn(move || {
+                worker_loop(rank, world, uid_copy, md, cfg, max_seq_len, rx, ack_tx);
+            }),
+            ack_rx,
+        ));
+        worker_txs.push(tx);
+    }
+
+    // Rank 0 setup
+    xserv_cuda::device::set_device(0).unwrap();
+    let tp0 = Arc::new(TpContext::init(0, world, uid, 0));
+    eprintln!("[rank 0] Loading weights...");
+    let weights = loader::load_model_dir(&model_dir, Device::Cpu);
+    eprintln!(
+        "[rank 0] Loaded {} tensors, building model...",
+        weights.len()
+    );
+    let model = GptOss::from_weights_tp(config.clone(), weights, 0, world, 0, Some(tp0));
+    let total_blocks = max_blocks_per_seq + 64;
+    let mut cache = PagedKVCache::new_tp(
+        &config,
+        local_kv,
+        total_blocks,
+        0,
+        4,
+        max_blocks_per_seq,
+        DType::BF16,
+        0,
+    );
+    eprintln!("[rank 0] Ready.");
+
+    // Prompt
+    let prompt_arg = get_arg::<String>(&args, "--prompt");
+    let prompt = prompt_arg
+        .as_deref()
+        .unwrap_or("What is the meaning of life?");
+    let token_ids = tokenizer.encode(prompt);
+    eprintln!("Prompt ({} tokens): {prompt}", token_ids.len());
+
+    // Register sequence
+    let slot = 0;
+    cache.register_sequence(slot).unwrap();
+    broadcast_cmd(&worker_txs, &worker_handles, WorkerCmd::Register(slot));
+
+    // Teacher-forced diagnostic: prefill (prompt + forced ids) in one shot and
+    // report, for each forced position, whether xserv's argmax == the forced
+    // (oracle) next token. Removes free-running compounding so it isolates
+    // whether per-position logits agree with the llama.cpp trajectory.
+    if let Some(forced) = get_arg::<String>(&args, "--forced") {
+        let forced_ids: Vec<u32> = forced
+            .split(',')
+            .filter_map(|s| s.trim().parse().ok())
+            .collect();
+        let mut seq = token_ids.clone();
+        seq.extend_from_slice(&forced_ids);
+        // Workers must run the same prefill in lockstep (TP AllReduces match up).
+        broadcast_cmd(
+            &worker_txs,
+            &worker_handles,
+            WorkerCmd::Prefill {
+                tokens: seq.clone(),
+                slot,
+            },
+        );
+        let logits = model.forward_prefill_paged(&seq, slot, &mut cache);
+        wait_workers(&worker_handles);
+        let logits_cpu = logits.to_device(Device::Cpu);
+        let vocab = logits.shape()[1];
+        let data = logits_cpu.as_slice::<half::bf16>();
+        let plen = token_ids.len();
+        let mut matches = 0usize;
+        let mut total = 0usize;
+        // position i predicts seq[i+1]; we check the forced region
+        for i in (plen - 1)..(seq.len() - 1) {
+            let row = &data[i * vocab..(i + 1) * vocab];
+            let argmax = row
+                .iter()
+                .enumerate()
+                .max_by(|a, b| a.1.to_f32().partial_cmp(&b.1.to_f32()).unwrap())
+                .map(|(j, _)| j as u32)
+                .unwrap();
+            let expected = seq[i + 1];
+            let ok = argmax == expected;
+            if ok {
+                matches += 1;
+            }
+            total += 1;
+            eprintln!(
+                "pos {i}: xserv_argmax={argmax} oracle={expected} {}",
+                if ok { "OK" } else { "DIFF" }
+            );
+        }
+        eprintln!(
+            "\nTeacher-forced top-1 agreement: {matches}/{total} = {:.1}%",
+            100.0 * matches as f64 / total as f64
+        );
+        broadcast_cmd(&worker_txs, &worker_handles, WorkerCmd::Shutdown);
+        for (h, _) in worker_handles {
+            h.join().unwrap();
+        }
+        return;
+    }
+
+    // Teacher-forced DECODE diagnostic: prefill the prompt, then walk the oracle
+    // trajectory through the autoregressive decode path (NOT prefill), recording
+    // per-position top-1 agreement bucketed by position. Localizes long-context
+    // decode degradation (which prefill teacher-forcing cannot see).
+    if let Some(forced) = get_arg::<String>(&args, "--forced-decode") {
+        let forced_ids: Vec<u32> = forced
+            .split(',')
+            .filter_map(|s| s.trim().parse().ok())
+            .collect();
+        broadcast_cmd(
+            &worker_txs,
+            &worker_handles,
+            WorkerCmd::Prefill {
+                tokens: token_ids.clone(),
+                slot,
+            },
+        );
+        let logits = model.forward_prefill_paged(&token_ids, slot, &mut cache);
+        wait_workers(&worker_handles);
+        let mut pred = sample_greedy_last(&logits); // prediction for forced[0]
+        let bucket = 50usize;
+        let mut buckets: Vec<(usize, usize)> = Vec::new();
+        let (mut matches, mut total) = (0usize, 0usize);
+        for (i, &f) in forced_ids.iter().enumerate() {
+            let ok = pred == f;
+            matches += ok as usize;
+            total += 1;
+            let b = i / bucket;
+            if buckets.len() <= b {
+                buckets.push((0, 0));
+            }
+            buckets[b].0 += ok as usize;
+            buckets[b].1 += 1;
+            // Teacher-force: feed the oracle token through the decode path.
+            let pos = cache.seq_len(slot);
+            broadcast_cmd(
+                &worker_txs,
+                &worker_handles,
+                WorkerCmd::Decode {
+                    tokens: vec![f],
+                    positions: vec![pos],
+                    slots: vec![slot],
+                },
+            );
+            let logits = model.forward_decode_paged(&[f], &[pos], &[slot], &mut cache);
+            wait_workers(&worker_handles);
+            pred = sample_greedy_last(&logits);
+        }
+        eprintln!(
+            "Teacher-forced DECODE agreement: {matches}/{total} = {:.1}%",
+            100.0 * matches as f64 / total as f64
+        );
+        for (b, (m, t)) in buckets.iter().enumerate() {
+            eprintln!(
+                "  pos[{:>4}..{:<4}]: {m:>3}/{t:<3} = {:.0}%",
+                b * bucket,
+                b * bucket + t,
+                100.0 * (*m as f64) / (*t as f64)
+            );
+        }
+        broadcast_cmd(&worker_txs, &worker_handles, WorkerCmd::Shutdown);
+        for (h, _) in worker_handles {
+            h.join().unwrap();
+        }
+        return;
+    }
+
+    // Prefill
+    let t0 = Instant::now();
+    broadcast_cmd(
+        &worker_txs,
+        &worker_handles,
+        WorkerCmd::Prefill {
+            tokens: token_ids.clone(),
+            slot,
+        },
+    );
+    let logits = model.forward_prefill_paged(&token_ids, slot, &mut cache);
+    wait_workers(&worker_handles);
+    let ttft = t0.elapsed();
+
+    let mut next = sample_greedy_last(&logits);
+    let mut output_tokens = vec![next];
+
+    eprintln!("TTFT: {:.1}ms", ttft.as_secs_f64() * 1000.0);
+    print!("{prompt}");
+
+    // Decode
+    let mut decoder = GraphedGptOssDecoder::new();
+    let decode_start = Instant::now();
+    for _ in 1..max_tokens {
+        let text = tokenizer.decode(&[next]);
+        print!("{text}");
+
+        if tokenizer.eos_token_id() == Some(next) {
+            break;
+        }
+
+        let pos = cache.seq_len(slot);
+        broadcast_cmd(
+            &worker_txs,
+            &worker_handles,
+            WorkerCmd::Decode {
+                tokens: vec![next],
+                positions: vec![pos],
+                slots: vec![slot],
+            },
+        );
+        let logits = decoder.decode(&model, &[next], &[pos], &[slot], &mut cache);
+        wait_workers(&worker_handles);
+
+        next = sample_greedy_last(&logits);
+        output_tokens.push(next);
+    }
+    let decode_elapsed = decode_start.elapsed();
+    println!();
+
+    let gen_tokens = output_tokens.len();
+    let full_text = tokenizer.decode(&output_tokens);
+    eprintln!("\nGenerated text: {full_text}");
+    eprintln!(
+        "Token IDs: {:?}",
+        &output_tokens[..output_tokens.len().min(20)]
+    );
+    let tpot = if gen_tokens > 1 {
+        decode_elapsed.as_secs_f64() * 1000.0 / (gen_tokens - 1) as f64
+    } else {
+        0.0
+    };
+    let tok_s = if gen_tokens > 1 {
+        (gen_tokens - 1) as f64 / decode_elapsed.as_secs_f64()
+    } else {
+        0.0
+    };
+
+    eprintln!("\n--- Performance ---");
+    eprintln!("Generated: {} tokens", gen_tokens);
+    eprintln!("TTFT: {:.1}ms", ttft.as_secs_f64() * 1000.0);
+    eprintln!("TPOT: {:.1}ms", tpot);
+    eprintln!("Throughput: {:.1} tok/s", tok_s);
+
+    // Cleanup
+    broadcast_cmd(&worker_txs, &worker_handles, WorkerCmd::Shutdown);
+    for (h, _) in worker_handles {
+        h.join().unwrap();
+    }
+}
+
+// --- Worker infrastructure ---
+
+#[derive(Clone)]
+enum WorkerCmd {
+    Register(usize),
+    Prefill {
+        tokens: Vec<u32>,
+        slot: usize,
+    },
+    Decode {
+        tokens: Vec<u32>,
+        positions: Vec<usize>,
+        slots: Vec<usize>,
+    },
+    Shutdown,
+}
+
+fn worker_loop(
+    rank: usize,
+    world: usize,
+    uid: UniqueId,
+    model_dir: PathBuf,
+    config: ModelConfig,
+    max_seq_len: usize,
+    rx: std::sync::mpsc::Receiver<WorkerCmd>,
+    ack_tx: std::sync::mpsc::Sender<()>,
+) {
+    xserv_cuda::device::set_device(rank as u32).unwrap();
+    let tp = Arc::new(TpContext::init(rank, world, uid, rank as u32));
+    eprintln!("[rank {rank}] Loading weights...");
+    let weights = loader::load_model_dir(&model_dir, Device::Cpu);
+    let model =
+        GptOss::from_weights_tp(config.clone(), weights, rank, world, rank as u32, Some(tp));
+    let local_kv = config.num_kv_heads() / world;
+    let max_blocks_per_seq = (max_seq_len + BLOCK_SIZE - 1) / BLOCK_SIZE;
+    let total_blocks = max_blocks_per_seq + 64;
+    let mut cache = PagedKVCache::new_tp(
+        &config,
+        local_kv,
+        total_blocks,
+        0,
+        4,
+        max_blocks_per_seq,
+        DType::BF16,
+        rank as u32,
+    );
+    eprintln!("[rank {rank}] Ready.");
+    ack_tx.send(()).unwrap();
+
+    let mut decoder = GraphedGptOssDecoder::new();
+    while let Ok(cmd) = rx.recv() {
+        match cmd {
+            WorkerCmd::Register(slot) => {
+                let _ = cache.register_sequence(slot);
+            }
+            WorkerCmd::Prefill { tokens, slot } => {
+                let _ = model.forward_prefill_paged(&tokens, slot, &mut cache);
+            }
+            WorkerCmd::Decode {
+                tokens,
+                positions,
+                slots,
+            } => {
+                let _ = decoder.decode(&model, &tokens, &positions, &slots, &mut cache);
+            }
+            WorkerCmd::Shutdown => break,
+        }
+        ack_tx.send(()).unwrap();
+    }
+}
+
+fn broadcast_cmd(
+    txs: &[std::sync::mpsc::Sender<WorkerCmd>],
+    _handles: &[(std::thread::JoinHandle<()>, std::sync::mpsc::Receiver<()>)],
+    cmd: WorkerCmd,
+) {
+    for tx in txs {
+        tx.send(cmd.clone()).unwrap();
+    }
+}
+
+fn wait_workers(handles: &[(std::thread::JoinHandle<()>, std::sync::mpsc::Receiver<()>)]) {
+    for (_, rx) in handles {
+        rx.recv().unwrap();
+    }
+}
+
+fn sample_greedy_last(logits: &xserv_tensor::Tensor) -> u32 {
+    use half::bf16;
+    assert_eq!(logits.ndim(), 2);
+    // GPU argmax fast path (4-byte D2H instead of the full logits row).
+    if logits.dtype() == xserv_tensor::DType::BF16 && logits.is_contiguous() {
+        let ids = xserv_kernels::argmax_bf16_to_host(logits);
+        return *ids.last().unwrap();
+    }
+    let logits_cpu = logits.to_device(Device::Cpu);
+    let vocab_size = logits.shape()[1];
+    let seq_len = logits.shape()[0];
+    let data = logits_cpu.as_slice::<bf16>();
+    let last = &data[(seq_len - 1) * vocab_size..seq_len * vocab_size];
+
+    last.iter()
+        .enumerate()
+        .max_by(|a, b| {
+            let af = a.1.to_f32();
+            let bf = b.1.to_f32();
+            af.partial_cmp(&bf).unwrap_or(std::cmp::Ordering::Equal)
+        })
+        .map(|(i, _)| i as u32)
+        .unwrap()
+}
+
+fn get_arg<T: std::str::FromStr>(args: &[String], flag: &str) -> Option<T> {
+    args.iter()
+        .position(|a| a == flag)
+        .and_then(|i| args.get(i + 1))
+        .and_then(|s| s.parse().ok())
+}

crates/xserv-model/src/bin/bench-gpt2.rs

@@ -0,0 +1,221 @@
+use std::path::PathBuf;
+use std::time::Instant;
+use xserv_model::gpt2::{KVCache, sample_greedy};
+use xserv_model::{GPT2, ModelConfig, loader};
+use xserv_tensor::Device;
+use xserv_tokenizer::Tokenizer;
+
+fn main() {
+    let args: Vec<String> = std::env::args().collect();
+    if args.len() < 2 {
+        eprintln!("Usage: bench-gpt2 <model-dir> [--gen-tokens N] [--no-cache]");
+        std::process::exit(1);
+    }
+    let model_dir = PathBuf::from(&args[1]);
+    let gen_tokens: usize = args
+        .iter()
+        .position(|a| a == "--gen-tokens")
+        .and_then(|i| args.get(i + 1))
+        .and_then(|s| s.parse().ok())
+        .unwrap_or(20);
+    let use_cache = !args.iter().any(|a| a == "--no-cache");
+
+    xserv_cuda::device::set_device(0).unwrap();
+
+    let config = ModelConfig::from_file(&model_dir.join("config.json"));
+    let weights = loader::load_model_dir(&model_dir, Device::Cuda(0));
+    let model = GPT2::from_weights(config.clone(), weights);
+    let tokenizer = Tokenizer::from_file(&model_dir.join("tokenizer.json"));
+
+    // Warmup
+    {
+        let ids = tokenizer.encode("warmup");
+        let _ = model.forward(&ids);
+    }
+
+    eprintln!("mode: {}", if use_cache { "KV cache" } else { "no cache" });
+
+    let prompts: Vec<&str> = vec![
+        "The capital of France is",
+        "Once upon a time in a land far away",
+        "Hello, how are you doing today",
+        "In a shocking finding, scientists discovered a",
+        "The weather today is sunny, so I decided to",
+        "Alan Turing was a British mathematician who",
+        "The best way to learn programming is",
+        "Artificial intelligence will change the world because",
+        "The history of the internet began in the",
+        "A good morning routine starts with",
+        "The stock market crashed because investors",
+        "Deep learning is a subset of machine learning that",
+        "The president of the United States announced",
+        "In the year 2050, humans will",
+        "The secret to happiness is",
+        "When I was a child, I used to",
+        "The most important scientific discovery of the century",
+        "Climate change is caused by",
+        "The recipe for chocolate cake requires",
+        "In conclusion, the evidence suggests that",
+        "The cat sat on the mat and",
+        "According to recent studies, exercise can",
+        "The first step in solving any problem is",
+        "Technology has transformed the way we",
+        "The novel begins with the protagonist",
+        "Education is the most powerful weapon",
+        "The ocean covers more than seventy percent of",
+        "Last night I had a dream about",
+        "The company announced its quarterly earnings",
+        "Music has the power to",
+        "The difference between success and failure is",
+        "In the beginning, there was nothing but",
+        "The doctor told me that I should",
+        "Python is a popular programming language because",
+        "The ancient Romans built roads that",
+        "A balanced diet should include",
+        "The movie received mixed reviews from critics",
+        "Space exploration has led to many",
+        "The teacher asked the students to",
+        "Global warming is one of the most",
+        "The bridge collapsed due to structural",
+        "Quantum computing promises to revolutionize",
+        "The new policy will affect millions of",
+        "During the winter months, it is important to",
+        "The human brain contains approximately",
+        "Democracy depends on the active participation of",
+        "The train arrived at the station exactly",
+        "Researchers at MIT have developed a new",
+        "The smartphone has become an essential part of",
+        "After careful consideration, the committee decided to",
+    ];
+
+    println!("[");
+    for (i, prompt) in prompts.iter().enumerate() {
+        let input_ids = tokenizer.encode(prompt);
+        let input_len = input_ids.len();
+
+        let (generated_ids, ttft_us, token_times_us) = if use_cache {
+            generate_with_cache(&model, &config, &tokenizer, &input_ids, gen_tokens)
+        } else {
+            generate_no_cache(&model, &tokenizer, &input_ids, gen_tokens)
+        };
+
+        let num_generated = generated_ids.len();
+        let generated_text = tokenizer.decode(&generated_ids);
+
+        let tbt_us = if !token_times_us.is_empty() {
+            token_times_us.iter().sum::<u128>() / token_times_us.len() as u128
+        } else {
+            0
+        };
+        let total_gen_us: u128 = ttft_us + token_times_us.iter().sum::<u128>();
+        let tpot_us = if num_generated > 0 {
+            total_gen_us / num_generated as u128
+        } else {
+            0
+        };
+
+        let gen_text_escaped = generated_text
+            .replace('\\', "\\\\")
+            .replace('"', "\\\"")
+            .replace('\n', "\\n")
+            .replace('\r', "\\r")
+            .replace('\t', "\\t");
+        let gen_ids_str: Vec<String> = generated_ids.iter().map(|id| id.to_string()).collect();
+
+        print!("  {{\"prompt\": \"{}\", ", prompt.replace('"', "\\\""));
+        print!("\"input_len\": {input_len}, ");
+        print!("\"num_generated\": {num_generated}, ");
+        print!("\"generated_ids\": [{}], ", gen_ids_str.join(", "));
+        print!("\"generated_text\": \"{gen_text_escaped}\", ");
+        print!("\"ttft_us\": {ttft_us}, ");
+        print!("\"tbt_us\": {tbt_us}, ");
+        print!("\"tpot_us\": {tpot_us}}}");
+        if i < prompts.len() - 1 {
+            println!(",");
+        } else {
+            println!();
+        }
+
+        eprintln!(
+            "[{}/{}] input={input_len}tok gen={num_generated}tok ttft={:.1}ms tbt={:.1}ms | {}",
+            i + 1,
+            prompts.len(),
+            ttft_us as f64 / 1000.0,
+            tbt_us as f64 / 1000.0,
+            &generated_text.replace('\n', " ")[..generated_text.len().min(60)]
+        );
+    }
+    println!("]");
+}
+
+fn generate_with_cache(
+    model: &GPT2,
+    config: &ModelConfig,
+    tokenizer: &Tokenizer,
+    input_ids: &[u32],
+    gen_tokens: usize,
+) -> (Vec<u32>, u128, Vec<u128>) {
+    let mut cache = KVCache::new(
+        config.num_layers(),
+        config.num_heads(),
+        config.head_dim(),
+        xserv_tensor::DType::F32,
+        Device::Cuda(0),
+    );
+
+    // Prefill
+    let t0 = Instant::now();
+    let logits = model.forward_with_cache(input_ids, &mut cache);
+    let first_token = sample_greedy(&logits);
+    let ttft_us = t0.elapsed().as_micros();
+
+    let mut generated = vec![first_token];
+    let mut token_times = Vec::new();
+
+    // Decode
+    for _ in 1..gen_tokens {
+        let last = *generated.last().unwrap();
+        let t_start = Instant::now();
+        let logits = model.forward_with_cache(&[last], &mut cache);
+        let next = sample_greedy(&logits);
+        token_times.push(t_start.elapsed().as_micros());
+        generated.push(next);
+        if tokenizer.eos_token_id() == Some(next) {
+            break;
+        }
+    }
+
+    (generated, ttft_us, token_times)
+}
+
+fn generate_no_cache(
+    model: &GPT2,
+    tokenizer: &Tokenizer,
+    input_ids: &[u32],
+    gen_tokens: usize,
+) -> (Vec<u32>, u128, Vec<u128>) {
+    let mut all_ids = input_ids.to_vec();
+
+    let t0 = Instant::now();
+    let logits = model.forward(&all_ids);
+    let first_token = sample_greedy(&logits);
+    let ttft_us = t0.elapsed().as_micros();
+    all_ids.push(first_token);
+
+    let mut generated = vec![first_token];
+    let mut token_times = Vec::new();
+
+    for _ in 1..gen_tokens {
+        let t_start = Instant::now();
+        let logits = model.forward(&all_ids);
+        let next = sample_greedy(&logits);
+        token_times.push(t_start.elapsed().as_micros());
+        all_ids.push(next);
+        generated.push(next);
+        if tokenizer.eos_token_id() == Some(next) {
+            break;
+        }
+    }
+
+    (generated, ttft_us, token_times)
+}

crates/xserv-model/src/bin/bench-qwen3.rs

@@ -0,0 +1,232 @@
+use std::path::PathBuf;
+use std::time::Instant;
+use xserv_model::qwen3::sample_greedy;
+use xserv_model::{DecodeGraphState, GpuKVCache, ModelConfig, Qwen3, loader};
+use xserv_tensor::{DType, Device};
+use xserv_tokenizer::Tokenizer;
+
+fn main() {
+    let args: Vec<String> = std::env::args().collect();
+    if args.len() < 2 {
+        eprintln!("Usage: bench-qwen3 <model-dir> [--gen-tokens N] [--cuda-graph]");
+        std::process::exit(1);
+    }
+    let model_dir = PathBuf::from(&args[1]);
+    let gen_tokens: usize = args
+        .iter()
+        .position(|a| a == "--gen-tokens")
+        .and_then(|i| args.get(i + 1))
+        .and_then(|s| s.parse().ok())
+        .unwrap_or(20);
+    let use_cuda_graph = args.iter().any(|a| a == "--cuda-graph");
+
+    xserv_cuda::device::set_device(0).unwrap();
+
+    let config = ModelConfig::from_file(&model_dir.join("config.json"));
+    eprintln!("Loading Qwen3-8B weights...");
+    let weights = loader::load_model_dir(&model_dir, Device::Cuda(0));
+    eprintln!("Loaded {} tensors", weights.len());
+    let model = Qwen3::from_weights(config.clone(), weights);
+    let tokenizer = Tokenizer::from_file(&model_dir.join("tokenizer.json"));
+
+    // Warmup
+    {
+        let ids = tokenizer.encode("warmup");
+        let mut cache = GpuKVCache::new(&config, 256, DType::BF16, 0);
+        let _ = model.forward_gpu_cache(&ids, &mut cache);
+    }
+
+    // CUDA Graph setup
+    let layer_ptrs = model.layer_weight_ptrs();
+    let (norm_w, lm_head, embed, cos, sin) = model.graph_capture_ptrs();
+    let mut decode_graph = if use_cuda_graph {
+        eprintln!("CUDA Graph mode enabled");
+        Some(DecodeGraphState::new(&config))
+    } else {
+        None
+    };
+    let mut graph_captured = false;
+
+    eprintln!("Warmup done. Running benchmark...");
+
+    let prompts: Vec<&str> = vec![
+        "The capital of France is",
+        "Once upon a time in a land far away",
+        "Hello, how are you doing today",
+        "In a shocking finding, scientists discovered a",
+        "The weather today is sunny, so I decided to",
+        "Alan Turing was a British mathematician who",
+        "The best way to learn programming is",
+        "Artificial intelligence will change the world because",
+        "The history of the internet began in the",
+        "A good morning routine starts with",
+        "The stock market crashed because investors",
+        "Deep learning is a subset of machine learning that",
+        "The president of the United States announced",
+        "In the year 2050, humans will",
+        "The secret to happiness is",
+        "When I was a child, I used to",
+        "The most important scientific discovery of the century",
+        "Climate change is caused by",
+        "The recipe for chocolate cake requires",
+        "In conclusion, the evidence suggests that",
+        "The cat sat on the mat and",
+        "According to recent studies, exercise can",
+        "The first step in solving any problem is",
+        "Technology has transformed the way we",
+        "The novel begins with the protagonist",
+        "Education is the most powerful weapon",
+        "The ocean covers more than seventy percent of",
+        "Last night I had a dream about",
+        "The company announced its quarterly earnings",
+        "Music has the power to",
+        "The difference between success and failure is",
+        "In the beginning, there was nothing but",
+        "The doctor told me that I should",
+        "Python is a popular programming language because",
+        "The ancient Romans built roads that",
+        "A balanced diet should include",
+        "The movie received mixed reviews from critics",
+        "Space exploration has led to many",
+        "The teacher asked the students to",
+        "Global warming is one of the most",
+        "The bridge collapsed due to structural",
+        "Quantum computing promises to revolutionize",
+        "The new policy will affect millions of",
+        "During the winter months, it is important to",
+        "The human brain contains approximately",
+        "Democracy depends on the active participation of",
+        "The train arrived at the station exactly",
+        "Researchers at MIT have developed a new",
+        "The smartphone has become an essential part of",
+        "After careful consideration, the committee decided to",
+    ];
+
+    println!("[");
+    for (i, prompt) in prompts.iter().enumerate() {
+        let input_ids = tokenizer.encode(prompt);
+        let input_len = input_ids.len();
+
+        let mut cache = GpuKVCache::new(&config, 256, DType::BF16, 0);
+
+        // Reset graph state for new prompt
+        graph_captured = false;
+        if let Some(ref mut g) = decode_graph {
+            g.invalidate();
+        }
+
+        // Prefill
+        let t0 = Instant::now();
+        let logits = model.forward_gpu_cache(&input_ids, &mut cache);
+        let first_token = sample_greedy(&logits);
+        let ttft_us = t0.elapsed().as_micros();
+
+        let mut generated = vec![first_token];
+        let mut token_times = Vec::new();
+
+        // Decode
+        for _ in 1..gen_tokens {
+            let last = *generated.last().unwrap();
+            let t_start = Instant::now();
+
+            let next = if let Some(ref mut graph) = decode_graph {
+                if !graph_captured {
+                    // First decode token: run ungraphed, then capture
+                    let logits = model.forward_gpu_cache(&[last], &mut cache);
+                    graph_captured = true;
+                    graph.capture(&layer_ptrs, norm_w, lm_head, embed, cos, sin);
+                    sample_greedy(&logits)
+                } else {
+                    // Replay captured graphs
+                    let pos = cache.seq_len() as u32;
+                    graph.execute(
+                        last,
+                        pos,
+                        &mut cache,
+                        &layer_ptrs,
+                        embed,
+                        config.vocab_size as i32,
+                        config.hidden() as i32,
+                    );
+                    cache.advance_seq_len(1);
+                    // Read logits from graph buffer
+                    let vocab_size = config.vocab_size;
+                    let mut logits_bytes = vec![0u8; vocab_size * 2];
+                    graph
+                        .logits_buffer()
+                        .copy_to_host(&mut logits_bytes)
+                        .unwrap();
+                    let logits_data: &[half::bf16] = unsafe {
+                        std::slice::from_raw_parts(
+                            logits_bytes.as_ptr() as *const half::bf16,
+                            vocab_size,
+                        )
+                    };
+                    logits_data
+                        .iter()
+                        .enumerate()
+                        .max_by(|a, b| a.1.to_f32().partial_cmp(&b.1.to_f32()).unwrap())
+                        .map(|(idx, _)| idx as u32)
+                        .unwrap()
+                }
+            } else {
+                let logits = model.forward_gpu_cache(&[last], &mut cache);
+                sample_greedy(&logits)
+            };
+
+            token_times.push(t_start.elapsed().as_micros());
+            generated.push(next);
+            if tokenizer.eos_token_id() == Some(next) {
+                break;
+            }
+        }
+
+        let num_generated = generated.len();
+        let generated_text = tokenizer.decode(&generated);
+        let tbt_us = if !token_times.is_empty() {
+            token_times.iter().sum::<u128>() / token_times.len() as u128
+        } else {
+            0
+        };
+        let total_gen_us: u128 = ttft_us + token_times.iter().sum::<u128>();
+        let tpot_us = if num_generated > 0 {
+            total_gen_us / num_generated as u128
+        } else {
+            0
+        };
+
+        let gen_text_escaped = generated_text
+            .replace('\\', "\\\\")
+            .replace('"', "\\\"")
+            .replace('\n', "\\n")
+            .replace('\r', "\\r")
+            .replace('\t', "\\t");
+        let gen_ids_str: Vec<String> = generated.iter().map(|id| id.to_string()).collect();
+
+        print!("  {{\"prompt\": \"{}\", ", prompt.replace('"', "\\\""));
+        print!("\"input_len\": {input_len}, ");
+        print!("\"num_generated\": {num_generated}, ");
+        print!("\"generated_ids\": [{}], ", gen_ids_str.join(", "));
+        print!("\"generated_text\": \"{gen_text_escaped}\", ");
+        print!("\"ttft_us\": {ttft_us}, ");
+        print!("\"tbt_us\": {tbt_us}, ");
+        print!("\"tpot_us\": {tpot_us}}}");
+        if i < prompts.len() - 1 {
+            println!(",");
+        } else {
+            println!();
+        }
+
+        let display_text = generated_text.replace('\n', " ");
+        let truncated: String = display_text.chars().take(60).collect();
+        eprintln!(
+            "[{}/{}] input={input_len}tok gen={num_generated}tok ttft={:.1}ms tbt={:.1}ms | {}",
+            i + 1,
+            prompts.len(),
+            ttft_us as f64 / 1000.0,
+            tbt_us as f64 / 1000.0,
+            truncated
+        );
+    }
+    println!("]");
+}

crates/xserv-model/src/bin/bench-speculative.rs

@@ -0,0 +1,976 @@
+//! Draft-model speculative decoding benchmark for Qwen3.
+//!
+//! v0 scope:
+//! - target + draft are Qwen3-family models with the same tokenizer/vocab;
+//! - batch=1;
+//! - greedy exact-match acceptance;
+//! - no probabilistic rejection sampling.
+
+use half::bf16;
+use std::path::{Path, PathBuf};
+use std::time::Instant;
+
+use xserv_model::qwen3_graph::GraphedQwen3Decoder;
+use xserv_model::{BLOCK_SIZE, ModelConfig, PagedKVCache, Qwen3, loader};
+use xserv_tensor::{DType, Device, Tensor};
+use xserv_tokenizer::Tokenizer;
+
+const DEFAULT_GAMMA: usize = 4;
+const DEFAULT_GEN_TOKENS: usize = 64;
+const DEFAULT_MAX_SEQ_LEN: usize = 2048;
+
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+enum VerifyPath {
+    Flash,
+    PagedDecode,
+}
+
+impl VerifyPath {
+    fn as_str(self) -> &'static str {
+        match self {
+            VerifyPath::Flash => "flash",
+            VerifyPath::PagedDecode => "paged-decode",
+        }
+    }
+}
+
+const PROMPTS: [&str; 50] = [
+    "The capital of France is",
+    "Once upon a time in a land far away",
+    "Hello, how are you doing today",
+    "In a shocking finding, scientists discovered a",
+    "The weather today is sunny, so I decided to",
+    "Alan Turing was a British mathematician who",
+    "The best way to learn programming is",
+    "Artificial intelligence will change the world because",
+    "The history of the internet began in the",
+    "A good morning routine starts with",
+    "The stock market crashed because investors",
+    "Deep learning is a subset of machine learning that",
+    "The president of the United States announced",
+    "In the year 2050, humans will",
+    "The secret to happiness is",
+    "When I was a child, I used to",
+    "The most important scientific discovery of the century",
+    "Climate change is caused by",
+    "The recipe for chocolate cake requires",
+    "In conclusion, the evidence suggests that",
+    "The cat sat on the mat and",
+    "According to recent studies, exercise can",
+    "The first step in solving any problem is",
+    "Technology has transformed the way we",
+    "The novel begins with the protagonist",
+    "Education is the most powerful weapon",
+    "The ocean covers more than seventy percent of",
+    "Last night I had a dream about",
+    "The company announced its quarterly earnings",
+    "Music has the power to",
+    "The difference between success and failure is",
+    "In the beginning, there was nothing but",
+    "The doctor told me that I should",
+    "Python is a popular programming language because",
+    "The ancient Romans built roads that",
+    "A balanced diet should include",
+    "The movie received mixed reviews from critics",
+    "Space exploration has led to many",
+    "The teacher asked the students to",
+    "Global warming is one of the most",
+    "The bridge collapsed due to structural",
+    "Quantum computing promises to revolutionize",
+    "The new policy will affect millions of",
+    "During the winter months, it is important to",
+    "The human brain contains approximately",
+    "Democracy depends on the active participation of",
+    "The train arrived at the station exactly",
+    "Researchers at MIT have developed a new",
+    "The smartphone has become an essential part of",
+    "After careful consideration, the committee decided to",
+];
+
+#[derive(Default)]
+struct RunStats {
+    ids: Vec<u32>,
+    total_s: f64,
+    prefill_s: f64,
+    decode_s: f64,
+    target_steps: usize,
+    accepted: usize,
+    proposed: usize,
+    verify_steps: usize,
+    mirror_steps: usize,
+    commit_steps: usize,
+    correction_steps: usize,
+    verify_decode_mismatches: usize,
+}
+
+#[derive(Default)]
+struct Totals {
+    prompts: usize,
+    baseline_generated: usize,
+    spec_generated: usize,
+    baseline_total_s: f64,
+    baseline_prefill_s: f64,
+    baseline_decode_s: f64,
+    spec_total_s: f64,
+    spec_prefill_s: f64,
+    spec_decode_s: f64,
+    spec_target_steps: usize,
+    spec_accepted: usize,
+    spec_proposed: usize,
+    spec_verify_steps: usize,
+    spec_mirror_steps: usize,
+    spec_commit_steps: usize,
+    spec_correction_steps: usize,
+    spec_verify_decode_mismatches: usize,
+    mismatches: usize,
+}
+
+fn main() {
+    let args: Vec<String> = std::env::args().collect();
+    if args.len() < 3 {
+        eprintln!(
+            "Usage: bench-speculative <target-model-dir> <draft-model-dir> \
+             [--gen-tokens N] [--gamma N] [--prompts N] [--max-seq-len N] [--device N] \
+             [--use-verify-logits] [--verify-path flash|paged-decode] [--dump-verify-mismatches]"
+        );
+        std::process::exit(1);
+    }
+
+    let target_dir = PathBuf::from(&args[1]);
+    let draft_dir = PathBuf::from(&args[2]);
+    let gen_tokens = arg_usize(&args, "--gen-tokens", DEFAULT_GEN_TOKENS);
+    let gamma = arg_usize(&args, "--gamma", DEFAULT_GAMMA);
+    let prompt_count = arg_usize(&args, "--prompts", PROMPTS.len()).min(PROMPTS.len());
+    let max_seq_len = arg_usize(&args, "--max-seq-len", DEFAULT_MAX_SEQ_LEN);
+    let device = arg_usize(&args, "--device", 0) as u32;
+    let use_verify_logits = args.iter().any(|a| a == "--use-verify-logits");
+    let verify_path = parse_verify_path(&args, use_verify_logits);
+    let dump_verify_mismatches = args.iter().any(|a| a == "--dump-verify-mismatches");
+
+    assert!(gen_tokens > 0, "--gen-tokens must be > 0");
+    assert!(gamma > 0, "--gamma must be > 0");
+
+    xserv_cuda::device::set_device(device).unwrap();
+    let info = xserv_cuda::device::device_info(device).unwrap();
+    eprintln!(
+        "GPU {device}: {} ({} MB free)",
+        info.name,
+        info.free_memory / 1024 / 1024
+    );
+
+    let target_config = ModelConfig::from_file(&target_dir.join("config.json"));
+    let draft_config = ModelConfig::from_file(&draft_dir.join("config.json"));
+    assert_qwen3(&target_config, "target");
+    assert_qwen3(&draft_config, "draft");
+    assert_eq!(
+        target_config.vocab_size, draft_config.vocab_size,
+        "target and draft vocab_size must match"
+    );
+
+    warn_if_tokenizers_differ(&target_dir, &draft_dir);
+    let tokenizer = Tokenizer::from_file(&target_dir.join("tokenizer.json"));
+    if tokenizer.vocab_size() != target_config.vocab_size {
+        eprintln!(
+            "WARNING: tokenizer decoder len {} differs from config vocab_size {}; continuing because token ids come from the shared tokenizer.json",
+            tokenizer.vocab_size(),
+            target_config.vocab_size
+        );
+    }
+
+    eprintln!(
+        "Loading target Qwen3: layers={} hidden={} heads={}/{} vocab={}",
+        target_config.num_layers(),
+        target_config.hidden(),
+        target_config.num_heads(),
+        target_config.num_kv_heads(),
+        target_config.vocab_size
+    );
+    let target_weights = loader::load_model_dir(&target_dir, Device::Cuda(device));
+    let target = Qwen3::from_weights(target_config.clone(), target_weights);
+    xserv_cuda::allocator::cached_trim();
+
+    eprintln!(
+        "Loading draft Qwen3: layers={} hidden={} heads={}/{} vocab={}",
+        draft_config.num_layers(),
+        draft_config.hidden(),
+        draft_config.num_heads(),
+        draft_config.num_kv_heads(),
+        draft_config.vocab_size
+    );
+    let draft_weights = loader::load_model_dir(&draft_dir, Device::Cuda(device));
+    let draft = Qwen3::from_weights(draft_config.clone(), draft_weights);
+    xserv_cuda::allocator::cached_trim();
+
+    let warm_ids = tokenizer.encode("warmup");
+    let warm_tokens = gen_tokens.min(4);
+    {
+        let mut target_cache = new_cache(&target_config, max_seq_len, device);
+        let _ = run_baseline(
+            &target,
+            &mut target_cache,
+            &tokenizer,
+            &warm_ids,
+            warm_tokens,
+        );
+    }
+    {
+        let mut target_cache = new_cache_with_rows(
+            &target_config,
+            max_seq_len,
+            device,
+            if use_verify_logits { gamma } else { 1 },
+        );
+        let mut target_verify_cache =
+            new_cache_with_rows(&target_config, max_seq_len, device, gamma);
+        let mut draft_cache = new_cache(&draft_config, max_seq_len, device);
+        let mut draft_decoder = GraphedQwen3Decoder::new();
+        let _ = run_speculative(
+            &target,
+            &draft,
+            &mut target_cache,
+            &mut target_verify_cache,
+            &mut draft_cache,
+            &mut draft_decoder,
+            &tokenizer,
+            &warm_ids,
+            warm_tokens,
+            gamma,
+            use_verify_logits,
+            verify_path,
+            dump_verify_mismatches,
+        );
+    }
+    eprintln!(
+        "Warmup done. Running {prompt_count} prompts, gen_tokens={gen_tokens}, gamma={gamma}, acceptance_mode={}, verify_path={}",
+        if use_verify_logits {
+            "verify_logits"
+        } else {
+            "decode"
+        },
+        verify_path.as_str()
+    );
+
+    let mut totals = Totals::default();
+
+    // Persistent per-benchmark caches so the draft CUDA graph (Phase 24) can be
+    // captured once and replayed across every prompt. Freeing and re-registering
+    // slot 0 between prompts keeps block_table_gpu / context_lens_gpu addresses
+    // stable, which is exactly what the graph captured.
+    let mut target_cache = new_cache_with_rows(
+        &target_config,
+        max_seq_len,
+        device,
+        if use_verify_logits { gamma } else { 1 },
+    );
+    let mut target_verify_cache = new_cache_with_rows(&target_config, max_seq_len, device, gamma);
+    let mut draft_cache = new_cache(&draft_config, max_seq_len, device);
+    let mut draft_decoder = GraphedQwen3Decoder::new();
+
+    for (i, prompt) in PROMPTS.iter().take(prompt_count).enumerate() {
+        let ids = tokenizer.encode(prompt);
+        validate_length_budget(&ids, gen_tokens, max_seq_len, prompt);
+        let mut baseline_cache = new_cache(&target_config, max_seq_len, device);
+        let baseline = run_baseline(&target, &mut baseline_cache, &tokenizer, &ids, gen_tokens);
+        drop(baseline_cache);
+
+        let spec = run_speculative(
+            &target,
+            &draft,
+            &mut target_cache,
+            &mut target_verify_cache,
+            &mut draft_cache,
+            &mut draft_decoder,
+            &tokenizer,
+            &ids,
+            gen_tokens,
+            gamma,
+            use_verify_logits,
+            verify_path,
+            dump_verify_mismatches,
+        );
+
+        let matched = baseline.ids == spec.ids;
+        if !matched {
+            totals.mismatches += 1;
+            eprintln!("MISMATCH prompt {i}: {prompt}");
+            eprintln!("  baseline: {:?}", baseline.ids);
+            eprintln!("  spec:     {:?}", spec.ids);
+        }
+
+        println!(
+            "prompt={:02} match={} gen={} accept={}/{} target_steps={} \
+             baseline_e2e_tpot_ms={:.3} spec_e2e_tpot_ms={:.3}",
+            i,
+            matched,
+            spec.ids.len(),
+            spec.accepted,
+            spec.proposed,
+            spec.target_steps,
+            per_token_ms(baseline.total_s, baseline.ids.len()),
+            per_token_ms(spec.total_s, spec.ids.len()),
+        );
+
+        totals.prompts += 1;
+        totals.baseline_generated += baseline.ids.len();
+        totals.spec_generated += spec.ids.len();
+        totals.baseline_total_s += baseline.total_s;
+        totals.baseline_prefill_s += baseline.prefill_s;
+        totals.baseline_decode_s += baseline.decode_s;
+        totals.spec_total_s += spec.total_s;
+        totals.spec_prefill_s += spec.prefill_s;
+        totals.spec_decode_s += spec.decode_s;
+        totals.spec_target_steps += spec.target_steps;
+        totals.spec_accepted += spec.accepted;
+        totals.spec_proposed += spec.proposed;
+        totals.spec_verify_steps += spec.verify_steps;
+        totals.spec_mirror_steps += spec.mirror_steps;
+        totals.spec_commit_steps += spec.commit_steps;
+        totals.spec_correction_steps += spec.correction_steps;
+        totals.spec_verify_decode_mismatches += spec.verify_decode_mismatches;
+    }
+
+    let baseline_decode_tokens = totals.baseline_generated;
+    let spec_decode_tokens = totals.spec_generated;
+    let acceptance = ratio(totals.spec_accepted, totals.spec_proposed);
+    let tokens_per_target_step = ratio(totals.spec_generated, totals.spec_target_steps);
+    let matched =
+        totals.mismatches == 0 && (!use_verify_logits || totals.spec_verify_decode_mismatches == 0);
+
+    println!("--- SUMMARY ---");
+    println!("prompts={} matched={matched}", totals.prompts);
+    println!(
+        "acceptance_mode={}",
+        if use_verify_logits {
+            "verify_logits"
+        } else {
+            "decode"
+        }
+    );
+    println!("verify_path={}", verify_path.as_str());
+    println!(
+        "acceptance_rate={:.4} accepted={} proposed={}",
+        acceptance, totals.spec_accepted, totals.spec_proposed
+    );
+    println!(
+        "tokens_per_target_step={:.4} target_steps={} verify_steps={} mirror_decode_steps={} commit_decode_steps={} correction_steps={}",
+        tokens_per_target_step,
+        totals.spec_target_steps,
+        totals.spec_verify_steps,
+        totals.spec_mirror_steps,
+        totals.spec_commit_steps,
+        totals.spec_correction_steps
+    );
+    println!(
+        "verify_decode_mismatches={}",
+        totals.spec_verify_decode_mismatches
+    );
+    println!(
+        "baseline_e2e_tpot_ms={:.3} baseline_e2e_tok_s={:.3}",
+        per_token_ms(totals.baseline_total_s, totals.baseline_generated),
+        tok_s(totals.baseline_generated, totals.baseline_total_s)
+    );
+    println!(
+        "spec_e2e_tpot_ms={:.3} spec_e2e_tok_s={:.3} speedup_e2e={:.4}",
+        per_token_ms(totals.spec_total_s, totals.spec_generated),
+        tok_s(totals.spec_generated, totals.spec_total_s),
+        speedup(totals.baseline_total_s, totals.spec_total_s)
+    );
+    println!(
+        "baseline_decode_tpot_ms={:.3} baseline_decode_tok_s={:.3}",
+        per_token_ms(totals.baseline_decode_s, baseline_decode_tokens),
+        tok_s(baseline_decode_tokens, totals.baseline_decode_s)
+    );
+    println!(
+        "spec_decode_tpot_ms={:.3} spec_decode_tok_s={:.3} speedup_decode={:.4}",
+        per_token_ms(totals.spec_decode_s, spec_decode_tokens),
+        tok_s(spec_decode_tokens, totals.spec_decode_s),
+        speedup(totals.baseline_decode_s, totals.spec_decode_s)
+    );
+    println!(
+        "decode_token_counts baseline={} spec={}",
+        baseline_decode_tokens, spec_decode_tokens
+    );
+
+    if !matched {
+        std::process::exit(2);
+    }
+}
+
+fn run_baseline(
+    model: &Qwen3,
+    cache: &mut PagedKVCache,
+    tokenizer: &Tokenizer,
+    prompt_ids: &[u32],
+    gen_tokens: usize,
+) -> RunStats {
+    let slot = 0;
+    cache.register_sequence(slot).expect("register target slot");
+
+    let t0 = Instant::now();
+    let prefill_start = Instant::now();
+    let logits = model.forward_prefill_paged(prompt_ids, slot, cache);
+    sync_device();
+    let prefill_s = prefill_start.elapsed().as_secs_f64();
+
+    let mut generated = Vec::with_capacity(gen_tokens);
+    let mut next = last_argmax(&logits);
+    generated.push(next);
+
+    let decode_start = Instant::now();
+    let mut target_steps = 0usize;
+    while generated.len() < gen_tokens && !tokenizer.is_eos(next) {
+        let pos = cache.seq_len(slot);
+        let logits = model.forward_decode_paged(&[next], &[pos], &[slot], cache);
+        target_steps += 1;
+        next = last_argmax(&logits);
+        generated.push(next);
+    }
+    sync_device();
+    let decode_s = decode_start.elapsed().as_secs_f64();
+    sync_device();
+    let total_s = t0.elapsed().as_secs_f64();
+
+    cache.free_sequence(slot);
+    RunStats {
+        ids: generated,
+        total_s,
+        prefill_s,
+        decode_s,
+        target_steps,
+        ..Default::default()
+    }
+}
+
+#[allow(clippy::too_many_arguments)]
+fn run_speculative(
+    target: &Qwen3,
+    draft: &Qwen3,
+    target_cache: &mut PagedKVCache,
+    target_verify_cache: &mut PagedKVCache,
+    draft_cache: &mut PagedKVCache,
+    draft_decoder: &mut GraphedQwen3Decoder,
+    tokenizer: &Tokenizer,
+    prompt_ids: &[u32],
+    gen_tokens: usize,
+    gamma: usize,
+    use_verify_logits: bool,
+    verify_path: VerifyPath,
+    dump_verify_mismatches: bool,
+) -> RunStats {
+    let slot = 0;
+    target_cache
+        .register_sequence(slot)
+        .expect("register target slot");
+    target_verify_cache
+        .register_sequence(slot)
+        .expect("register target verify slot");
+    draft_cache
+        .register_sequence(slot)
+        .expect("register draft slot");
+
+    let t0 = Instant::now();
+    let prefill_start = Instant::now();
+    let target_logits = target.forward_prefill_paged(prompt_ids, slot, target_cache);
+    if !use_verify_logits {
+        let _ = target.forward_prefill_paged(prompt_ids, slot, target_verify_cache);
+    }
+    let draft_logits = draft.forward_prefill_paged(prompt_ids, slot, draft_cache);
+    sync_device();
+    let prefill_s = prefill_start.elapsed().as_secs_f64();
+
+    let mut target_next = last_argmax(&target_logits);
+    let mut draft_next = last_argmax(&draft_logits);
+    let mut generated = Vec::with_capacity(gen_tokens);
+    let mut accepted_total = 0usize;
+    let mut proposed_total = 0usize;
+    let mut verify_steps = 0usize;
+    let mut mirror_steps = 0usize;
+    let mut commit_steps = 0usize;
+    let mut correction_steps = 0usize;
+    let mut verify_decode_mismatches = 0usize;
+
+    let decode_start = Instant::now();
+    while generated.len() < gen_tokens {
+        let remaining = gen_tokens - generated.len();
+        let round_gamma = gamma.min(remaining);
+        let round_start_len = target_cache.seq_len(slot);
+        assert_eq!(
+            round_start_len,
+            draft_cache.seq_len(slot),
+            "target and draft cache lengths diverged"
+        );
+        if !use_verify_logits {
+            assert_eq!(
+                round_start_len,
+                target_verify_cache.seq_len(slot),
+                "target verify cache length diverged"
+            );
+        }
+
+        let mut draft_tokens = Vec::with_capacity(round_gamma);
+        for _ in 0..round_gamma {
+            let token = draft_next;
+            draft_tokens.push(token);
+            if tokenizer.is_eos(token) {
+                break;
+            }
+            let pos = draft_cache.seq_len(slot);
+            let logits = draft_decoder.decode(draft, &[token], &[pos], &[slot], draft_cache);
+            draft_next = last_argmax(&logits);
+        }
+        proposed_total += draft_tokens.len();
+
+        if use_verify_logits {
+            verify_steps += 1;
+            let verify_logits =
+                target.forward_verify_paged_decode_attention(&draft_tokens, slot, target_cache);
+            let verify_argmax = argmax_rows(&verify_logits);
+            assert_eq!(
+                verify_argmax.len(),
+                draft_tokens.len(),
+                "verify logits rows must match draft token count"
+            );
+
+            let mut accepted = 0usize;
+            let mut done = false;
+            while accepted < draft_tokens.len() {
+                let expected = if accepted > 0 {
+                    verify_argmax[accepted - 1]
+                } else {
+                    target_next
+                };
+                if draft_tokens[accepted] != expected {
+                    break;
+                }
+                let token = draft_tokens[accepted];
+                generated.push(token);
+                accepted_total += 1;
+                accepted += 1;
+
+                if generated.len() >= gen_tokens || tokenizer.is_eos(token) {
+                    done = true;
+                    break;
+                }
+            }
+
+            if accepted > 0 {
+                target_next = verify_argmax[accepted - 1];
+            }
+            target_cache
+                .truncate_sequence(slot, round_start_len + accepted)
+                .unwrap();
+
+            if done {
+                draft_cache
+                    .truncate_sequence(slot, target_cache.seq_len(slot))
+                    .unwrap();
+                break;
+            }
+
+            if accepted == draft_tokens.len() {
+                continue;
+            }
+
+            let correction = if accepted > 0 {
+                verify_argmax[accepted - 1]
+            } else {
+                target_next
+            };
+            generated.push(correction);
+
+            draft_cache
+                .truncate_sequence(slot, round_start_len)
+                .unwrap();
+            replay_draft_tokens(
+                draft,
+                draft_decoder,
+                draft_cache,
+                slot,
+                &draft_tokens[..accepted],
+                &mut draft_next,
+            );
+
+            if generated.len() >= gen_tokens || tokenizer.is_eos(correction) {
+                break;
+            }
+
+            let pos = target_cache.seq_len(slot);
+            let logits = target.forward_decode_paged(&[correction], &[pos], &[slot], target_cache);
+            target_next = last_argmax(&logits);
+            commit_steps += 1;
+
+            let pos = draft_cache.seq_len(slot);
+            let logits = draft_decoder.decode(draft, &[correction], &[pos], &[slot], draft_cache);
+            draft_next = last_argmax(&logits);
+            correction_steps += 1;
+            continue;
+        }
+
+        verify_steps += 1;
+        let verify_logits = match verify_path {
+            VerifyPath::Flash => {
+                target.forward_prefill_paged(&draft_tokens, slot, target_verify_cache)
+            }
+            VerifyPath::PagedDecode => target.forward_verify_paged_decode_attention(
+                &draft_tokens,
+                slot,
+                target_verify_cache,
+            ),
+        };
+        let verify_argmax = argmax_rows(&verify_logits);
+        assert_eq!(
+            verify_argmax.len(),
+            draft_tokens.len(),
+            "verify logits rows must match draft token count"
+        );
+
+        target_verify_cache
+            .truncate_sequence(slot, round_start_len)
+            .unwrap();
+
+        let mut accepted = 0usize;
+        let mut done = false;
+        while accepted < draft_tokens.len() {
+            let expected = if use_verify_logits && accepted > 0 {
+                verify_argmax[accepted - 1]
+            } else {
+                target_next
+            };
+            if draft_tokens[accepted] != expected {
+                break;
+            }
+            let token_idx = accepted;
+            let token = draft_tokens[token_idx];
+            generated.push(token);
+            accepted_total += 1;
+            accepted += 1;
+
+            if generated.len() >= gen_tokens || tokenizer.is_eos(token) {
+                done = true;
+                break;
+            }
+
+            let pos = target_cache.seq_len(slot);
+            let logits = target.forward_decode_paged(&[token], &[pos], &[slot], target_cache);
+            let decode_next = last_argmax(&logits);
+            if verify_argmax[token_idx] != decode_next {
+                verify_decode_mismatches += 1;
+                eprintln!(
+                    "VERIFY/DECODE MISMATCH at cache_len={} accepted_idx={}: verify={} decode={}",
+                    target_cache.seq_len(slot),
+                    token_idx,
+                    verify_argmax[token_idx],
+                    decode_next
+                );
+                if dump_verify_mismatches {
+                    eprintln!(
+                        "  verify_top5={} decode_top5={}",
+                        format_topk(&verify_logits, token_idx, 5),
+                        format_topk(&logits, 0, 5)
+                    );
+                }
+            }
+            target_next = decode_next;
+            commit_steps += 1;
+
+            advance_target_cache(target, target_verify_cache, slot, token);
+            mirror_steps += 1;
+        }
+        if done {
+            draft_cache
+                .truncate_sequence(slot, target_cache.seq_len(slot))
+                .unwrap();
+            target_verify_cache
+                .truncate_sequence(slot, target_cache.seq_len(slot))
+                .unwrap();
+            break;
+        }
+
+        if accepted == draft_tokens.len() {
+            continue;
+        }
+
+        let correction = if use_verify_logits && accepted > 0 {
+            verify_argmax[accepted - 1]
+        } else {
+            target_next
+        };
+        generated.push(correction);
+
+        draft_cache
+            .truncate_sequence(slot, round_start_len)
+            .unwrap();
+        replay_draft_tokens(
+            draft,
+            draft_decoder,
+            draft_cache,
+            slot,
+            &draft_tokens[..accepted],
+            &mut draft_next,
+        );
+
+        if generated.len() >= gen_tokens || tokenizer.is_eos(correction) {
+            break;
+        }
+
+        let pos = target_cache.seq_len(slot);
+        let logits = target.forward_decode_paged(&[correction], &[pos], &[slot], target_cache);
+        target_next = last_argmax(&logits);
+        commit_steps += 1;
+
+        advance_target_cache(target, target_verify_cache, slot, correction);
+        mirror_steps += 1;
+
+        let pos = draft_cache.seq_len(slot);
+        let logits = draft_decoder.decode(draft, &[correction], &[pos], &[slot], draft_cache);
+        draft_next = last_argmax(&logits);
+        correction_steps += 1;
+    }
+    sync_device();
+    let decode_s = decode_start.elapsed().as_secs_f64();
+    sync_device();
+    let total_s = t0.elapsed().as_secs_f64();
+
+    target_cache.free_sequence(slot);
+    target_verify_cache.free_sequence(slot);
+    draft_cache.free_sequence(slot);
+
+    RunStats {
+        ids: generated,
+        total_s,
+        prefill_s,
+        decode_s,
+        target_steps: verify_steps + mirror_steps + commit_steps + correction_steps,
+        accepted: accepted_total,
+        proposed: proposed_total,
+        verify_steps,
+        mirror_steps,
+        commit_steps,
+        correction_steps,
+        verify_decode_mismatches,
+    }
+}
+
+fn advance_target_cache(target: &Qwen3, cache: &mut PagedKVCache, slot: usize, token: u32) {
+    let pos = cache.seq_len(slot);
+    let _ = target.forward_decode_paged(&[token], &[pos], &[slot], cache);
+}
+
+fn replay_draft_tokens(
+    draft: &Qwen3,
+    draft_decoder: &mut GraphedQwen3Decoder,
+    cache: &mut PagedKVCache,
+    slot: usize,
+    tokens: &[u32],
+    next: &mut u32,
+) {
+    for &token in tokens {
+        let pos = cache.seq_len(slot);
+        let logits = draft_decoder.decode(draft, &[token], &[pos], &[slot], cache);
+        *next = last_argmax(&logits);
+    }
+}
+
+fn new_cache(config: &ModelConfig, max_seq_len: usize, device: u32) -> PagedKVCache {
+    new_cache_with_rows(config, max_seq_len, device, 1)
+}
+
+fn new_cache_with_rows(
+    config: &ModelConfig,
+    max_seq_len: usize,
+    device: u32,
+    max_rows: usize,
+) -> PagedKVCache {
+    let max_blocks_per_seq = max_seq_len.div_ceil(BLOCK_SIZE);
+    let total_blocks = max_blocks_per_seq + 8;
+    PagedKVCache::new(
+        config,
+        total_blocks,
+        0,
+        max_rows.max(1),
+        max_blocks_per_seq,
+        DType::BF16,
+        device,
+    )
+}
+
+fn argmax_rows(logits: &Tensor) -> Vec<u32> {
+    assert_eq!(logits.ndim(), 2);
+    if logits.dtype() == DType::BF16
+        && matches!(logits.device(), Device::Cuda(_))
+        && logits.is_contiguous()
+    {
+        return xserv_kernels::argmax_bf16_to_host(logits);
+    }
+
+    let vocab_size = logits.shape()[1];
+    let rows = logits.shape()[0];
+    let logits_cpu = logits.to_device(Device::Cpu);
+    match logits.dtype() {
+        DType::F32 => logits_cpu
+            .as_slice::<f32>()
+            .chunks_exact(vocab_size)
+            .take(rows)
+            .map(argmax_f32)
+            .collect(),
+        DType::BF16 => logits_cpu
+            .as_slice::<bf16>()
+            .chunks_exact(vocab_size)
+            .take(rows)
+            .map(|row| {
+                row.iter()
+                    .enumerate()
+                    .max_by(|a, b| a.1.to_f32().partial_cmp(&b.1.to_f32()).unwrap())
+                    .map(|(i, _)| i as u32)
+                    .unwrap()
+            })
+            .collect(),
+        _ => panic!("unsupported dtype for argmax: {:?}", logits.dtype()),
+    }
+}
+
+fn last_argmax(logits: &Tensor) -> u32 {
+    *argmax_rows(logits).last().unwrap()
+}
+
+fn argmax_f32(row: &[f32]) -> u32 {
+    row.iter()
+        .enumerate()
+        .max_by(|a, b| a.1.partial_cmp(b.1).unwrap())
+        .map(|(i, _)| i as u32)
+        .unwrap()
+}
+
+fn format_topk(logits: &Tensor, row: usize, k: usize) -> String {
+    let vals = topk_row(logits, row, k);
+    vals.iter()
+        .map(|(id, val)| format!("{id}:{val:.3}"))
+        .collect::<Vec<_>>()
+        .join(",")
+}
+
+fn topk_row(logits: &Tensor, row: usize, k: usize) -> Vec<(u32, f32)> {
+    assert_eq!(logits.ndim(), 2);
+    let vocab_size = logits.shape()[1];
+    assert!(row < logits.shape()[0], "topk row out of bounds");
+    let logits_cpu = logits.to_device(Device::Cpu);
+    let mut vals: Vec<(u32, f32)> = match logits.dtype() {
+        DType::F32 => logits_cpu.as_slice::<f32>()[row * vocab_size..(row + 1) * vocab_size]
+            .iter()
+            .enumerate()
+            .map(|(i, &v)| (i as u32, v))
+            .collect(),
+        DType::BF16 => logits_cpu.as_slice::<bf16>()[row * vocab_size..(row + 1) * vocab_size]
+            .iter()
+            .enumerate()
+            .map(|(i, &v)| (i as u32, v.to_f32()))
+            .collect(),
+        _ => panic!("unsupported dtype for topk: {:?}", logits.dtype()),
+    };
+    vals.select_nth_unstable_by(k, |a, b| b.1.partial_cmp(&a.1).unwrap());
+    vals.truncate(k);
+    vals.sort_unstable_by(|a, b| b.1.partial_cmp(&a.1).unwrap());
+    vals
+}
+
+fn assert_qwen3(config: &ModelConfig, name: &str) {
+    let model_type = config.model_type.as_deref().unwrap_or("unknown");
+    assert!(
+        model_type.contains("qwen"),
+        "{name} model_type must be qwen-like, got {model_type}"
+    );
+}
+
+fn warn_if_tokenizers_differ(target_dir: &Path, draft_dir: &Path) {
+    let target = std::fs::read(target_dir.join("tokenizer.json"));
+    let draft = std::fs::read(draft_dir.join("tokenizer.json"));
+    if let (Ok(target), Ok(draft)) = (target, draft) {
+        if target != draft {
+            eprintln!(
+                "WARNING: target and draft tokenizer.json differ; v0 assumes identical token ids"
+            );
+        }
+    }
+}
+
+fn arg_usize(args: &[String], flag: &str, default: usize) -> usize {
+    args.iter()
+        .position(|a| a == flag)
+        .and_then(|i| args.get(i + 1))
+        .and_then(|s| s.parse().ok())
+        .unwrap_or(default)
+}
+
+fn parse_verify_path(args: &[String], use_verify_logits: bool) -> VerifyPath {
+    let default = if use_verify_logits {
+        VerifyPath::PagedDecode
+    } else {
+        VerifyPath::Flash
+    };
+    let Some(value) = args
+        .iter()
+        .position(|a| a == "--verify-path")
+        .and_then(|i| args.get(i + 1))
+    else {
+        return default;
+    };
+    match value.as_str() {
+        "flash" => VerifyPath::Flash,
+        "paged-decode" => VerifyPath::PagedDecode,
+        _ => {
+            eprintln!("unknown --verify-path {value:?}; expected flash or paged-decode");
+            std::process::exit(1);
+        }
+    }
+}
+
+fn validate_length_budget(prompt_ids: &[u32], gen_tokens: usize, max_seq_len: usize, prompt: &str) {
+    let required = prompt_ids.len() + gen_tokens;
+    if required > max_seq_len {
+        eprintln!(
+            "prompt requires prompt_len({}) + gen_tokens({}) = {} tokens, exceeding --max-seq-len {}: {:?}",
+            prompt_ids.len(),
+            gen_tokens,
+            required,
+            max_seq_len,
+            prompt
+        );
+        std::process::exit(2);
+    }
+}
+
+fn sync_device() {
+    xserv_cuda::device::synchronize().expect("cuda device synchronize");
+}
+
+fn ratio(num: usize, den: usize) -> f64 {
+    if den == 0 {
+        0.0
+    } else {
+        num as f64 / den as f64
+    }
+}
+
+fn speedup(baseline_s: f64, spec_s: f64) -> f64 {
+    if spec_s == 0.0 {
+        0.0
+    } else {
+        baseline_s / spec_s
+    }
+}
+
+fn tok_s(tokens: usize, seconds: f64) -> f64 {
+    if seconds == 0.0 {
+        0.0
+    } else {
+        tokens as f64 / seconds
+    }
+}
+
+fn per_token_ms(seconds: f64, tokens: usize) -> f64 {
+    if tokens == 0 {
+        0.0
+    } else {
+        seconds * 1000.0 / tokens as f64
+    }
+}

crates/xserv-model/src/bin/bench-tp.rs

@@ -0,0 +1,244 @@
+//! Tensor-parallel E2E benchmark for Qwen3.
+//!
+//! Spawns one thread per TP rank (each bound to one GPU), loads the sharded
+//! model, and runs greedy autoregressive generation. Because lm_head is
+//! replicated and the post-AllReduce hidden state is identical on every rank,
+//! all ranks compute identical logits and pick the same greedy token — so the
+//! rank threads stay in lockstep via the per-layer AllReduces without any
+//! token broadcast. Rank 0 records output + timings.
+//!
+//! Usage: bench-tp <model-dir> [--tp N] [--gen-tokens N] [--devices 0,1,2,3]
+//!
+//! Run with --tp 1 / 2 / 4 and compare the printed text (correctness) and
+//! tok/s (performance).
+
+use std::path::PathBuf;
+use std::sync::Arc;
+use std::thread;
+use std::time::Instant;
+
+use xserv_model::qwen3::sample_greedy;
+use xserv_model::{BLOCK_SIZE, ModelConfig, PagedKVCache, Qwen3, loader};
+use xserv_tensor::{DType, Device};
+use xserv_tokenizer::Tokenizer;
+
+struct PromptResult {
+    gen_ids: Vec<u32>,
+    ttft_ms: f64,
+    decode_tok_s: f64,
+}
+
+fn main() {
+    let args: Vec<String> = std::env::args().collect();
+    if args.len() < 2 {
+        eprintln!("Usage: bench-tp <model-dir> [--tp N] [--gen-tokens N] [--devices 0,1,2,3]");
+        std::process::exit(1);
+    }
+    let model_dir = PathBuf::from(&args[1]);
+    let world: usize = arg(&args, "--tp")
+        .and_then(|s| s.parse().ok())
+        .unwrap_or(1)
+        .max(1);
+    let gen_tokens: usize = arg(&args, "--gen-tokens")
+        .and_then(|s| s.parse().ok())
+        .unwrap_or(64);
+    let devices: Vec<u32> = match arg(&args, "--devices") {
+        Some(s) => s.split(',').filter_map(|d| d.trim().parse().ok()).collect(),
+        None => (0..world as u32).collect(),
+    };
+    assert_eq!(devices.len(), world, "--devices count must equal --tp");
+
+    let config = ModelConfig::from_file(&model_dir.join("config.json"));
+    assert!(
+        config.num_kv_heads() % world == 0,
+        "num_kv_heads {} not divisible by tp {world}",
+        config.num_kv_heads()
+    );
+    let tokenizer = Tokenizer::from_file(&model_dir.join("tokenizer.json"));
+    let eos = tokenizer.eos_token_id();
+
+    let prompts: Vec<&str> = vec![
+        "The capital of France is",
+        "Explain photosynthesis in one sentence.",
+        "Write a haiku about the ocean.",
+        "List three uses of a hammer.",
+        "What is the speed of light?",
+        "Describe the water cycle briefly.",
+        "Who wrote Romeo and Juliet?",
+        "Translate 'good morning' into Spanish.",
+    ];
+    let prompt_ids: Vec<Vec<u32>> = prompts.iter().map(|p| tokenizer.encode(p)).collect();
+
+    // Tensors are not Send (their Storage holds a raw GPU pointer), so each rank
+    // thread loads its own CPU copy of the weights and shards in-thread. Loading
+    // is not part of the timed region.
+    let id = if world > 1 {
+        Some(xserv_distributed::get_unique_id())
+    } else {
+        None
+    };
+
+    let handles: Vec<_> = (0..world)
+        .map(|rank| {
+            let model_dir = model_dir.clone();
+            let config = config.clone();
+            let prompt_ids = prompt_ids.clone();
+            let device = devices[rank];
+            thread::spawn(move || {
+                run_rank(
+                    rank, world, device, id, config, model_dir, prompt_ids, gen_tokens, eos,
+                )
+            })
+        })
+        .collect();
+
+    let mut rank0: Option<Vec<PromptResult>> = None;
+    for (rank, h) in handles.into_iter().enumerate() {
+        let r = h.join().expect("rank thread panicked");
+        if rank == 0 {
+            rank0 = r;
+        }
+    }
+
+    let results = rank0.expect("rank 0 produced no results");
+    println!("\n=== TP={world} (devices {devices:?}) — Qwen3 E2E benchmark ===");
+    println!(
+        "{:<45} {:>10} {:>12} {:>8}",
+        "prompt", "TTFT(ms)", "decode tok/s", "gen"
+    );
+    let mut tps_sum = 0.0;
+    for (i, r) in results.iter().enumerate() {
+        let text = tokenizer.decode(&r.gen_ids).replace('\n', " ");
+        let short: String = text.chars().take(50).collect();
+        let p: String = prompts[i].chars().take(43).collect();
+        println!(
+            "{:<45} {:>10.1} {:>12.1} {:>8} | {}",
+            p,
+            r.ttft_ms,
+            r.decode_tok_s,
+            r.gen_ids.len(),
+            short
+        );
+        tps_sum += r.decode_tok_s;
+    }
+    println!(
+        "--- mean decode throughput: {:.1} tok/s ---",
+        tps_sum / results.len() as f64
+    );
+
+    // Machine-readable line for cross-TP correctness diffing (rank 0 token ids).
+    let all_ids: Vec<String> = results
+        .iter()
+        .map(|r| {
+            r.gen_ids
+                .iter()
+                .map(|x| x.to_string())
+                .collect::<Vec<_>>()
+                .join(",")
+        })
+        .collect();
+    println!("CORRECTNESS_IDS tp={world} {}", all_ids.join(" | "));
+}
+
+fn run_rank(
+    rank: usize,
+    world: usize,
+    device: u32,
+    id: Option<xserv_distributed::UniqueId>,
+    config: ModelConfig,
+    model_dir: PathBuf,
+    prompt_ids: Vec<Vec<u32>>,
+    gen_tokens: usize,
+    eos: Option<u32>,
+) -> Option<Vec<PromptResult>> {
+    // Bind this thread to its GPU and set up the TP communicator.
+    let tp = if world > 1 {
+        Some(Arc::new(xserv_distributed::TpContext::init(
+            rank,
+            world,
+            id.unwrap(),
+            device,
+        )))
+    } else {
+        xserv_cuda::device::set_device(device).unwrap();
+        None
+    };
+
+    // Load this rank's own CPU copy of the weights and shard in-thread.
+    let weights = loader::load_model_dir(&model_dir, Device::Cpu);
+    let model = Qwen3::from_weights_tp(config.clone(), weights, rank, world, device, tp.clone());
+
+    // Per-rank paged KV cache holds only this rank's local KV heads.
+    let local_kv = config.num_kv_heads() / world;
+    let max_seq = 2048usize;
+    let max_blocks_per_seq = max_seq.div_ceil(BLOCK_SIZE);
+    let total_blocks = max_blocks_per_seq + 8;
+    let mut cache = PagedKVCache::new_tp(
+        &config,
+        local_kv,
+        total_blocks,
+        0,
+        1,
+        max_blocks_per_seq,
+        DType::BF16,
+        device,
+    );
+
+    // Warmup (init kernels / allocator / NCCL channels) — not timed.
+    cache.register_sequence(0).unwrap();
+    let _ = model.forward_prefill_paged(&[1u32, 2, 3], 0, &mut cache);
+    cache.free_sequence(0);
+
+    let mut out = Vec::new();
+    for ids in &prompt_ids {
+        cache.register_sequence(0).unwrap();
+
+        // Prefill (TTFT).
+        let t0 = Instant::now();
+        let logits = model.forward_prefill_paged(ids, 0, &mut cache);
+        let first = sample_greedy(&logits);
+        let ttft_ms = t0.elapsed().as_secs_f64() * 1000.0;
+
+        let mut generated = vec![first];
+
+        // Decode.
+        let t1 = Instant::now();
+        let mut steps = 0usize;
+        for _ in 1..gen_tokens {
+            let last = *generated.last().unwrap();
+            if eos == Some(last) {
+                break;
+            }
+            let pos = cache.seq_len(0);
+            let logits = model.forward_decode_paged(&[last], &[pos], &[0], &mut cache);
+            let next = sample_greedy(&logits);
+            generated.push(next);
+            steps += 1;
+        }
+        let decode_s = t1.elapsed().as_secs_f64();
+        let decode_tok_s = if steps > 0 && decode_s > 0.0 {
+            steps as f64 / decode_s
+        } else {
+            0.0
+        };
+
+        cache.free_sequence(0);
+
+        if rank == 0 {
+            out.push(PromptResult {
+                gen_ids: generated,
+                ttft_ms,
+                decode_tok_s,
+            });
+        }
+    }
+
+    if rank == 0 { Some(out) } else { None }
+}
+
+fn arg<'a>(args: &'a [String], flag: &str) -> Option<&'a str> {
+    args.iter()
+        .position(|a| a == flag)
+        .and_then(|i| args.get(i + 1))
+        .map(|s| s.as_str())
+}

crates/xserv-model/src/bin/bench-verify-cost.rs

@@ -0,0 +1,134 @@
+//! Micro-benchmark: measure the cost of forward_verify_paged_decode_attention
+//! at different batch sizes (γ+1 values), to understand where speedup comes
+//! from (or doesn't).
+
+use std::path::PathBuf;
+use std::time::Instant;
+
+use xserv_model::{BLOCK_SIZE, ModelConfig, PagedKVCache, Qwen3, loader};
+use xserv_tensor::{DType, Device};
+use xserv_tokenizer::Tokenizer;
+
+fn main() {
+    let args: Vec<String> = std::env::args().collect();
+    if args.len() < 2 {
+        eprintln!(
+            "Usage: bench-verify-cost <target-dir> [--prompt-len N] [--iters N] [--device N]"
+        );
+        std::process::exit(1);
+    }
+    let target_dir = PathBuf::from(&args[1]);
+    let prompt_len = arg_usize(&args, "--prompt-len", 100);
+    let iters = arg_usize(&args, "--iters", 30);
+    let device = arg_usize(&args, "--device", 0) as u32;
+
+    xserv_cuda::device::set_device(device).unwrap();
+
+    let cfg = ModelConfig::from_file(&target_dir.join("config.json"));
+    eprintln!("Loading target...");
+    let weights = loader::load_model_dir(&target_dir, Device::Cuda(device));
+    let target = Qwen3::from_weights(cfg.clone(), weights);
+    xserv_cuda::allocator::cached_trim();
+
+    let tok = Tokenizer::from_file(&target_dir.join("tokenizer.json"));
+    let ids = tok.encode(&"the ".repeat(prompt_len))[..prompt_len].to_vec();
+
+    let max_seq_len = 2048;
+    let num_blocks = (max_seq_len + BLOCK_SIZE - 1) / BLOCK_SIZE + 4;
+    let mut cache = PagedKVCache::new(&cfg, num_blocks, 0, 16, num_blocks, DType::BF16, device);
+    cache.register_sequence(0).unwrap();
+
+    // Prefill
+    let _ = target.forward_prefill_paged(&ids, 0, &mut cache);
+    sync();
+
+    // Warmup one of each
+    for &n in &[1, 2, 3, 5, 9] {
+        let toks: Vec<u32> = (0..n).map(|_| ids[0]).collect();
+        let _ = target.forward_decode_paged(
+            &toks,
+            &(0..n).map(|i| ids.len() + i).collect::<Vec<_>>(),
+            &vec![0; n],
+            &mut cache,
+        );
+        cache.truncate_sequence(0, ids.len()).unwrap();
+    }
+    sync();
+
+    // Benchmark single-token decode
+    let mut t = 0.0f64;
+    for i in 0..iters {
+        cache.truncate_sequence(0, ids.len()).unwrap();
+        let t0 = Instant::now();
+        let _ = target.forward_decode_paged(&[ids[0]], &[ids.len()], &[0], &mut cache);
+        sync();
+        t += t0.elapsed().as_secs_f64();
+        let _ = i;
+    }
+    let single = t * 1000.0 / iters as f64;
+    println!(
+        "single-token decode: {:.3} ms (mean of {} iters)",
+        single, iters
+    );
+
+    // Benchmark forward_verify_paged_decode_attention at various batch sizes
+    // (batched-GEMV path).
+    for &n in &[1usize, 2, 3, 5, 9] {
+        let toks: Vec<u32> = (0..n).map(|_| ids[0]).collect();
+        let mut t = 0.0f64;
+        for _ in 0..iters {
+            cache.truncate_sequence(0, ids.len()).unwrap();
+            let t0 = Instant::now();
+            let _ = target.forward_verify_paged_decode_attention(&toks, 0, &mut cache);
+            sync();
+            t += t0.elapsed().as_secs_f64();
+        }
+        let ms = t * 1000.0 / iters as f64;
+        println!(
+            "verify (batched-GEMV) batch={}: {:.3} ms ({:.2}× single)",
+            n,
+            ms,
+            ms / single
+        );
+    }
+
+    // Benchmark _with_hidden variant which uses cuBLAS GEMM after Phase 26 fast-verify.
+    let hooks_layers = [2usize, 18, 33];
+    for &n in &[1usize, 2, 3, 5, 9] {
+        let toks: Vec<u32> = (0..n).map(|_| ids[0]).collect();
+        let mut t = 0.0f64;
+        for _ in 0..iters {
+            cache.truncate_sequence(0, ids.len()).unwrap();
+            let t0 = Instant::now();
+            let _ = target.forward_verify_paged_decode_attention_with_hidden(
+                &toks,
+                0,
+                &mut cache,
+                &hooks_layers,
+            );
+            sync();
+            t += t0.elapsed().as_secs_f64();
+        }
+        let ms = t * 1000.0 / iters as f64;
+        println!(
+            "verify (cuBLAS GEMM) batch={}: {:.3} ms ({:.2}× single)",
+            n,
+            ms,
+            ms / single
+        );
+    }
+
+    cache.free_sequence(0);
+}
+
+fn sync() {
+    xserv_cuda::device::synchronize().unwrap();
+}
+
+fn arg_usize(args: &[String], flag: &str, default: usize) -> usize {
+    args.iter()
+        .position(|a| a == flag)
+        .and_then(|i| args.get(i + 1))
+        .and_then(|s| s.parse().ok())
+        .unwrap_or(default)
+}

crates/xserv-model/src/bin/check-eagle3.rs

@@ -0,0 +1,174 @@
+//! EAGLE3 sanity check: load weights, run one draft step, print top-5 predictions.
+//!
+//! This verifies that:
+//! - Eagle3Head weights load without shape mismatches
+//! - Target hidden states can be captured via decode_core_with_hidden
+//! - Eagle3Head::step produces a valid token id (in target vocab)
+//!
+//! Does NOT measure speedup — that requires a full γ≥2 speculative loop, which
+//! is more complex integration work.
+
+use std::path::PathBuf;
+
+use xserv_model::eagle3::{EAGLE_HOOK_LAYERS, Eagle3Head};
+use xserv_model::{BLOCK_SIZE, ModelConfig, PagedKVCache, Qwen3, loader};
+use xserv_tensor::{DType, Device, Tensor};
+use xserv_tokenizer::Tokenizer;
+
+fn main() {
+    let args: Vec<String> = std::env::args().collect();
+    if args.len() < 3 {
+        eprintln!("Usage: check-eagle3 <target-model-dir> <eagle3-model-dir> [prompt]");
+        std::process::exit(1);
+    }
+    let target_dir = PathBuf::from(&args[1]);
+    let eagle_dir = PathBuf::from(&args[2]);
+    let prompt = args
+        .get(3)
+        .cloned()
+        .unwrap_or_else(|| "The capital of France is".to_string());
+    let device: u32 = 0;
+
+    xserv_cuda::device::set_device(device).unwrap();
+
+    let target_config = ModelConfig::from_file(&target_dir.join("config.json"));
+    eprintln!("Loading target Qwen3-8B...");
+    let target_weights = loader::load_model_dir(&target_dir, Device::Cuda(device));
+    let target = Qwen3::from_weights(target_config.clone(), target_weights);
+    xserv_cuda::allocator::cached_trim();
+
+    eprintln!("Loading EAGLE3 head from {}", eagle_dir.display());
+    let mut eagle = Eagle3Head::load(&eagle_dir, device);
+    xserv_cuda::allocator::cached_trim();
+
+    let tokenizer = Tokenizer::from_file(&target_dir.join("tokenizer.json"));
+    let embed_tokens = target.embed_tokens_tensor();
+
+    let ids = tokenizer.encode(&prompt);
+    let max_seq_len = 512;
+
+    let num_blocks = (max_seq_len + BLOCK_SIZE - 1) / BLOCK_SIZE + 2;
+    let mut cache = PagedKVCache::new(
+        &target_config,
+        num_blocks,
+        0,
+        1,
+        num_blocks,
+        DType::BF16,
+        device,
+    );
+    cache.register_sequence(0).unwrap();
+
+    // Prefill target.
+    let logits = target.forward_prefill_paged(&ids, 0, &mut cache);
+    let target_first = *xserv_kernels::argmax_bf16_to_host(&logits).last().unwrap();
+    let target_first_text = tokenizer.decode(&[target_first]);
+    println!("Prompt: {:?}", prompt);
+    println!(
+        "Target argmax after prefill: {} ({:?})",
+        target_first, target_first_text
+    );
+
+    // Now run one target decode step with target_first to get hidden states at the
+    // hook layers.
+    let pos = cache.seq_len(0);
+    target.decode_prepare(&[pos], &[0], &mut cache);
+    let ids_gpu = upload_u32(&[target_first]);
+    let pos_gpu = upload_u32(&[pos as u32]);
+    let (target_next_logits, hooks) = target.decode_core_with_hidden(
+        ids_gpu.as_ptr() as *const std::ffi::c_void,
+        pos_gpu.as_ptr() as *const std::ffi::c_void,
+        1,
+        &[0],
+        &mut cache,
+        &EAGLE_HOOK_LAYERS,
+    );
+    let target_next = xserv_kernels::argmax_bf16_single(&target_next_logits);
+    let target_next_text = tokenizer.decode(&[target_next]);
+    println!(
+        "Target argmax after 1 decode step: {} ({:?})",
+        target_next, target_next_text
+    );
+
+    for (i, h) in hooks.iter().enumerate() {
+        println!(
+            "hook[{}] (layer {}): shape={:?} dtype={:?}",
+            i,
+            EAGLE_HOOK_LAYERS[i],
+            h.shape(),
+            h.dtype()
+        );
+    }
+
+    // Ask EAGLE what it thinks the NEXT token is (given target_first as prev_token
+    // and the hidden states from the position where target_first lives).
+    // EAGLE should predict target_next (or close to it) to be useful.
+    eagle.reset();
+    let (eagle_pred, eagle_logits) = eagle.step(&hooks, embed_tokens, target_first, pos);
+    let eagle_pred_text = tokenizer.decode(&[eagle_pred]);
+    println!(
+        "EAGLE draft prediction (pairing A: prev=target_first): {} ({:?})",
+        eagle_pred, eagle_pred_text
+    );
+
+    if eagle_pred == target_next {
+        println!("MATCH: EAGLE agrees with target on next token.");
+    } else {
+        println!(
+            "MISMATCH: EAGLE draft={} vs target={} (this is fine per-step; check top-5 below)",
+            eagle_pred, target_next
+        );
+    }
+
+    // Show top-5 from eagle logits (in draft vocab space, mapped to target).
+    print_top5(
+        &eagle_logits,
+        "EAGLE draft top-5 (pairing A)",
+        &eagle,
+        &tokenizer,
+    );
+
+    // Alternative pairing B: pair hooks with target_next (the token those hooks produced
+    // via lm_head), predict token after target_next. Position advances by 1.
+    eagle.reset();
+    let (eagle_pred_b, eagle_logits_b) = eagle.step(&hooks, embed_tokens, target_next, pos + 1);
+    let eagle_pred_b_text = tokenizer.decode(&[eagle_pred_b]);
+    println!(
+        "\nEAGLE draft prediction (pairing B: prev=target_next): {} ({:?})",
+        eagle_pred_b, eagle_pred_b_text
+    );
+    print_top5(
+        &eagle_logits_b,
+        "EAGLE draft top-5 (pairing B)",
+        &eagle,
+        &tokenizer,
+    );
+}
+
+fn upload_u32(vals: &[u32]) -> xserv_cuda::GpuBuffer {
+    let bytes = unsafe { std::slice::from_raw_parts(vals.as_ptr() as *const u8, vals.len() * 4) };
+    let mut buf = xserv_cuda::allocator::cached_alloc(bytes.len()).unwrap();
+    buf.copy_from_host(bytes).unwrap();
+    buf
+}
+
+fn print_top5(logits: &Tensor, label: &str, eagle: &Eagle3Head, tokenizer: &Tokenizer) {
+    use half::bf16;
+    let cpu = logits.to_device(Device::Cpu);
+    let data = cpu.as_slice::<bf16>();
+    let mut vals: Vec<(usize, f32)> = data
+        .iter()
+        .enumerate()
+        .map(|(i, v)| (i, v.to_f32()))
+        .collect();
+    vals.sort_by(|a, b| b.1.partial_cmp(&a.1).unwrap());
+    println!("{label}:");
+    for (i, val) in vals.iter().take(5) {
+        let target_id = eagle.map_draft_to_target(*i as u32);
+        let text = tokenizer.decode(&[target_id]);
+        println!(
+            "  draft_id={} target_id={} val={:.3} text={:?}",
+            i, target_id, val, text
+        );
+    }
+}

crates/xserv-model/src/bin/dump-logits.rs

@@ -0,0 +1,49 @@
+use half::bf16;
+use std::path::PathBuf;
+use xserv_model::{KVCache, ModelConfig, Qwen3, loader};
+use xserv_tensor::{DType, Device};
+use xserv_tokenizer::Tokenizer;
+
+fn main() {
+    let args: Vec<String> = std::env::args().collect();
+    let model_dir = PathBuf::from(&args[1]);
+    let prompt = &args[2];
+
+    xserv_cuda::device::set_device(0).unwrap();
+    let config = ModelConfig::from_file(&model_dir.join("config.json"));
+    let weights = loader::load_model_dir(&model_dir, Device::Cuda(0));
+    let model = Qwen3::from_weights(config.clone(), weights);
+    let tokenizer = Tokenizer::from_file(&model_dir.join("tokenizer.json"));
+
+    let token_ids = tokenizer.encode(prompt);
+    eprintln!("Prompt: {prompt}");
+    eprintln!("Token IDs: {token_ids:?}");
+
+    let mut cache = KVCache::new(
+        config.num_layers(),
+        config.num_kv_heads(),
+        config.head_dim(),
+        DType::BF16,
+        Device::Cuda(0),
+    );
+    let logits = model.forward_with_cache(&token_ids, &mut cache);
+    let logits_cpu = logits.to_device(Device::Cpu);
+    let data = logits_cpu.as_slice::<bf16>();
+    let vocab_size = logits.shape()[1];
+    let seq_len = logits.shape()[0];
+
+    // Print top-20 logits for the last position
+    let last_row = &data[(seq_len - 1) * vocab_size..seq_len * vocab_size];
+    let mut indexed: Vec<(usize, f32)> = last_row
+        .iter()
+        .enumerate()
+        .map(|(i, v)| (i, v.to_f32()))
+        .collect();
+    indexed.sort_by(|a, b| b.1.partial_cmp(&a.1).unwrap());
+
+    println!("Top-20 logits (last position):");
+    for (rank, (id, val)) in indexed.iter().take(20).enumerate() {
+        let tok = tokenizer.decode(&[*id as u32]);
+        println!("  [{rank:>2}] id={id:>6} logit={val:>10.4} token={tok:?}");
+    }
+}

crates/xserv-model/src/bin/xserv-cli.rs

@@ -0,0 +1,212 @@
+use std::io::{self, Write};
+use std::path::PathBuf;
+use xserv_model::{
+    BLOCK_SIZE, KVCache, ModelConfig, PagedKVCache, SamplingParams, loader, sample,
+    sample_greedy_penalized,
+};
+use xserv_tensor::{DType, Device};
+use xserv_tokenizer::Tokenizer;
+
+fn flag<T: std::str::FromStr>(args: &[String], name: &str, default: T) -> T {
+    args.iter()
+        .position(|a| a == name)
+        .and_then(|i| args.get(i + 1))
+        .and_then(|s| s.parse().ok())
+        .unwrap_or(default)
+}
+
+fn pick_next(
+    logits: &xserv_tensor::Tensor,
+    sampling: &SamplingParams,
+    history: &[u32],
+    rep_penalty: f32,
+) -> u32 {
+    if rep_penalty > 1.0 && sampling.temperature == 0.0 {
+        sample_greedy_penalized(logits, history, rep_penalty)
+    } else {
+        sample(logits, sampling)
+    }
+}
+
+fn main() {
+    let args: Vec<String> = std::env::args().collect();
+    if args.len() < 2 {
+        eprintln!(
+            "Usage: xserv-cli <model-dir> [--max-tokens N] [--temperature F] [--top-k N] [--top-p F] [--rep-penalty F] [--rep-window N]"
+        );
+        std::process::exit(1);
+    }
+
+    let model_dir = PathBuf::from(&args[1]);
+    let max_tokens = flag(&args, "--max-tokens", 100usize);
+    let sampling = SamplingParams {
+        temperature: flag(&args, "--temperature", 0.0f32),
+        top_k: flag(&args, "--top-k", 0usize),
+        top_p: flag(&args, "--top-p", 1.0f32),
+    };
+    let rep_penalty = flag(&args, "--rep-penalty", 1.0f32);
+    let rep_window = flag(&args, "--rep-window", 512usize);
+
+    xserv_cuda::device::set_device(0).unwrap();
+    let info = xserv_cuda::device::device_info(0).unwrap();
+    eprintln!(
+        "GPU: {} ({} MB free)",
+        info.name,
+        info.free_memory / 1024 / 1024
+    );
+
+    let config = ModelConfig::from_file(&model_dir.join("config.json"));
+    let model_type = config.model_type.as_deref().unwrap_or("unknown");
+    eprintln!(
+        "Model: {model_type}, layers={}, hidden={}, heads={}/{} kv, vocab={}",
+        config.num_layers(),
+        config.hidden(),
+        config.num_heads(),
+        config.num_kv_heads(),
+        config.vocab_size
+    );
+
+    eprintln!("Loading weights...");
+    let weights = loader::load_model_dir(&model_dir, Device::Cuda(0));
+    eprintln!("Loaded {} tensors", weights.len());
+
+    let is_qwen3 = model_type.contains("qwen");
+    let is_gpt_oss = model_type.contains("gpt_oss");
+    let dtype = if is_qwen3 || is_gpt_oss {
+        DType::BF16
+    } else {
+        DType::F32
+    };
+
+    // Build model
+    enum Model {
+        GPT2(xserv_model::GPT2),
+        Qwen3(xserv_model::Qwen3),
+        GptOss(xserv_model::GptOss),
+    }
+    let model = if is_gpt_oss {
+        Model::GptOss(xserv_model::GptOss::from_weights(config.clone(), weights))
+    } else if is_qwen3 {
+        Model::Qwen3(xserv_model::Qwen3::from_weights(config.clone(), weights))
+    } else {
+        Model::GPT2(xserv_model::GPT2::from_weights(config.clone(), weights))
+    };
+
+    let tokenizer = Tokenizer::from_file(&model_dir.join("tokenizer.json"));
+    eprintln!(
+        "Ready (KV cache, dtype={dtype}, temperature={}, top_k={}, top_p={}, rep_penalty={}, rep_window={}).\n",
+        sampling.temperature, sampling.top_k, sampling.top_p, rep_penalty, rep_window
+    );
+
+    loop {
+        print!("xserv> ");
+        io::stdout().flush().unwrap();
+        let mut input = String::new();
+        if io::stdin().read_line(&mut input).unwrap() == 0 {
+            break;
+        }
+        let raw_input = input.trim();
+        if raw_input.is_empty() {
+            continue;
+        }
+        if raw_input == "quit" || raw_input == "exit" {
+            break;
+        }
+        let input = raw_input.replace("\\n", "\n");
+
+        let token_ids = tokenizer.encode(&input);
+
+        if is_gpt_oss {
+            // GptOss uses paged KV cache
+            let max_seq = 2048;
+            let max_blocks_per_seq = (max_seq + BLOCK_SIZE - 1) / BLOCK_SIZE;
+            let total_blocks = max_blocks_per_seq + 64;
+            let mut paged_cache = PagedKVCache::new(
+                &config,
+                total_blocks,
+                0,
+                4,
+                max_blocks_per_seq,
+                DType::BF16,
+                0,
+            );
+            let slot = 0;
+            paged_cache.register_sequence(slot).expect("register slot");
+
+            let model = match &model {
+                Model::GptOss(m) => m,
+                _ => unreachable!(),
+            };
+            let logits = model.forward_prefill_paged(&token_ids, slot, &mut paged_cache);
+            let mut history = token_ids.clone();
+            let start = history.len().saturating_sub(rep_window);
+            let mut next = pick_next(&logits, &sampling, &history[start..], rep_penalty);
+
+            print!("{input}");
+            io::stdout().flush().unwrap();
+
+            for _ in 0..max_tokens {
+                let text = tokenizer.decode(&[next]);
+                print!("{text}");
+                io::stdout().flush().unwrap();
+                history.push(next);
+
+                if tokenizer.eos_token_id() == Some(next) {
+                    break;
+                }
+
+                let pos = paged_cache.seq_len(slot);
+                let logits = model.forward_decode_paged(&[next], &[pos], &[slot], &mut paged_cache);
+                let start = history.len().saturating_sub(rep_window);
+                next = pick_next(&logits, &sampling, &history[start..], rep_penalty);
+            }
+            println!();
+            paged_cache.free_sequence(slot);
+        } else {
+            let kv_heads = if is_qwen3 {
+                config.num_kv_heads()
+            } else {
+                config.num_heads()
+            };
+            let mut cache = KVCache::new(
+                config.num_layers(),
+                kv_heads,
+                config.head_dim(),
+                dtype,
+                Device::Cuda(0),
+            );
+
+            let logits = match &model {
+                Model::GPT2(m) => m.forward_with_cache(&token_ids, &mut cache),
+                Model::Qwen3(m) => m.forward_with_cache(&token_ids, &mut cache),
+                Model::GptOss(_) => unreachable!(),
+            };
+            let mut history = token_ids.clone();
+            let start = history.len().saturating_sub(rep_window);
+            let mut next = pick_next(&logits, &sampling, &history[start..], rep_penalty);
+
+            print!("{input}");
+            io::stdout().flush().unwrap();
+
+            for _ in 0..max_tokens {
+                let text = tokenizer.decode(&[next]);
+                print!("{text}");
+                io::stdout().flush().unwrap();
+                history.push(next);
+
+                if tokenizer.eos_token_id() == Some(next) {
+                    break;
+                }
+
+                let logits = match &model {
+                    Model::GPT2(m) => m.forward_with_cache(&[next], &mut cache),
+                    Model::Qwen3(m) => m.forward_with_cache(&[next], &mut cache),
+                    Model::GptOss(_) => unreachable!(),
+                };
+                let start = history.len().saturating_sub(rep_window);
+                next = pick_next(&logits, &sampling, &history[start..], rep_penalty);
+            }
+            println!();
+        }
+    }
+}

crates/xserv-model/src/config.rs

@@ -0,0 +1,166 @@
+use serde::Deserialize;
+use std::path::Path;
+
+#[derive(Debug, Clone, Deserialize)]
+pub struct RopeScaling {
+    pub rope_type: Option<String>,
+    pub factor: Option<f64>,
+    pub original_max_position_embeddings: Option<usize>,
+    pub beta_fast: Option<f64>,
+    pub beta_slow: Option<f64>,
+}
+
+#[derive(Debug, Clone, Deserialize)]
+pub struct ModelConfig {
+    pub architectures: Option<Vec<String>>,
+    pub model_type: Option<String>,
+
+    // Modern HF naming
+    #[serde(default)]
+    pub hidden_size: Option<usize>,
+    #[serde(default)]
+    pub intermediate_size: Option<usize>,
+    #[serde(default)]
+    pub num_attention_heads: Option<usize>,
+    #[serde(default)]
+    pub num_key_value_heads: Option<usize>,
+    #[serde(default)]
+    pub num_hidden_layers: Option<usize>,
+    pub vocab_size: usize,
+    #[serde(default)]
+    pub max_position_embeddings: Option<usize>,
+
+    // GPT-2 naming
+    #[serde(default)]
+    pub n_embd: Option<usize>,
+    #[serde(default)]
+    pub n_head: Option<usize>,
+    #[serde(default)]
+    pub n_layer: Option<usize>,
+    #[serde(default)]
+    pub n_positions: Option<usize>,
+    #[serde(default)]
+    pub n_inner: Option<usize>,
+
+    // Normalization
+    #[serde(default)]
+    pub layer_norm_eps: Option<f64>,
+    #[serde(default)]
+    pub layer_norm_epsilon: Option<f64>,
+    #[serde(default)]
+    pub rms_norm_eps: Option<f64>,
+
+    // Other
+    #[serde(default)]
+    pub rope_theta: Option<f64>,
+    #[serde(default)]
+    pub tie_word_embeddings: Option<bool>,
+
+    // MoE (gpt-oss)
+    #[serde(default)]
+    pub num_local_experts: Option<usize>,
+    #[serde(default)]
+    pub num_experts_per_tok: Option<usize>,
+    #[serde(default)]
+    pub layer_types: Option<Vec<String>>,
+    #[serde(default)]
+    pub sliding_window: Option<usize>,
+    #[serde(default)]
+    pub attention_bias: Option<bool>,
+    #[serde(default, rename = "head_dim")]
+    pub explicit_head_dim: Option<usize>,
+    #[serde(default)]
+    pub rope_scaling: Option<RopeScaling>,
+    #[serde(default)]
+    pub swiglu_limit: Option<f64>,
+    #[serde(default)]
+    pub geglu_alpha: Option<f64>,
+    #[serde(default)]
+    pub hidden_act: Option<String>,
+}
+
+impl ModelConfig {
+    pub fn from_file(path: &Path) -> Self {
+        let data = std::fs::read_to_string(path)
+            .unwrap_or_else(|e| panic!("failed to read {}: {e}", path.display()));
+        serde_json::from_str(&data)
+            .unwrap_or_else(|e| panic!("failed to parse {}: {e}", path.display()))
+    }
+
+    pub fn hidden(&self) -> usize {
+        self.hidden_size
+            .or(self.n_embd)
+            .expect("hidden_size or n_embd required")
+    }
+
+    pub fn num_heads(&self) -> usize {
+        self.num_attention_heads
+            .or(self.n_head)
+            .expect("num_attention_heads or n_head required")
+    }
+
+    pub fn num_layers(&self) -> usize {
+        self.num_hidden_layers
+            .or(self.n_layer)
+            .expect("num_hidden_layers or n_layer required")
+    }
+
+    pub fn max_seq_len(&self) -> usize {
+        self.max_position_embeddings
+            .or(self.n_positions)
+            .unwrap_or(2048)
+    }
+
+    pub fn ffn_hidden(&self) -> usize {
+        self.intermediate_size
+            .or(self.n_inner)
+            .unwrap_or(self.hidden() * 4)
+    }
+
+    pub fn num_kv_heads(&self) -> usize {
+        self.num_key_value_heads.unwrap_or(self.num_heads())
+    }
+
+    pub fn head_dim(&self) -> usize {
+        self.explicit_head_dim
+            .unwrap_or_else(|| self.hidden() / self.num_heads())
+    }
+
+    pub fn ln_eps(&self) -> f32 {
+        self.layer_norm_eps
+            .or(self.layer_norm_epsilon)
+            .unwrap_or(1e-5) as f32
+    }
+
+    pub fn tied_embeddings(&self) -> bool {
+        self.tie_word_embeddings.unwrap_or(true)
+    }
+
+    pub fn num_experts(&self) -> usize {
+        self.num_local_experts.unwrap_or(0)
+    }
+
+    pub fn experts_per_token(&self) -> usize {
+        self.num_experts_per_tok.unwrap_or(1)
+    }
+
+    pub fn is_moe(&self) -> bool {
+        self.num_local_experts.unwrap_or(0) > 1
+    }
+
+    pub fn is_sliding_layer(&self, layer_idx: usize) -> bool {
+        self.layer_types
+            .as_ref()
+            .and_then(|lt| lt.get(layer_idx))
+            .map(|t| t == "sliding_attention")
+            .unwrap_or(false)
+    }
+
+    pub fn window_size(&self) -> usize {
+        self.sliding_window.unwrap_or(0)
+    }
+
+    pub fn geglu_alpha(&self) -> f32 {
+        self.geglu_alpha.unwrap_or(1.702) as f32
+    }
+}

crates/xserv-model/src/decode_graph.rs

@@ -0,0 +1,649 @@
+//! CUDA Graph integration for batch=1 single-sequence decode.
+//!
+//! Uses a per-layer split graph approach:
+//! - Pre-attention graph: RMSNorm + QKV projections + reshape + QK-norm + RoPE
+//! - Ungraphed: KV cache append + decode attention (variable kv_len)
+//! - Post-attention graph: merge_heads + O-proj + add_rmsnorm + FFN + residual
+//! - Final graph: last RMSNorm + lm_head GEMV
+
+use std::ffi::c_void;
+use xserv_cuda::{CudaGraph, CudaStream, GpuBuffer};
+use xserv_kernels::dispatch;
+use xserv_kernels::gemm::{cublas_handle, gemv_scratch_elems};
+
+use crate::config::ModelConfig;
+use crate::kv_cache::GpuKVCache;
+
+/// Pre-allocated intermediate buffers for decode (batch=1).
+/// All buffers have stable GPU addresses for CUDA Graph replay.
+struct DecodeBuffers {
+    // Hidden-size buffers: [1, hidden]
+    x: GpuBuffer,           // running hidden state
+    normed: GpuBuffer,      // rmsnorm output
+    attn_out: GpuBuffer,    // attention output [1, num_heads, 1, head_dim]
+    attn_merged: GpuBuffer, // merge_heads output [1, hidden]
+    o_proj: GpuBuffer,      // O projection output [1, hidden]
+    normed2: GpuBuffer,     // post-attn norm output [1, hidden]
+    sum_out: GpuBuffer,     // add_rmsnorm sum output [1, hidden]
+    down: GpuBuffer,        // down projection output [1, hidden]
+
+    // QKV projection outputs
+    q_proj: GpuBuffer, // [1, num_heads * head_dim]
+    k_proj: GpuBuffer, // [1, num_kv_heads * head_dim]
+    v_proj: GpuBuffer, // [1, num_kv_heads * head_dim]
+
+    // Reshaped: [1, H, 1, D]
+    q_reshaped: GpuBuffer,
+    k_reshaped: GpuBuffer,
+    v_reshaped: GpuBuffer,
+
+    // After QK-norm (same shape as reshaped)
+    q_normed: GpuBuffer,
+    k_normed: GpuBuffer,
+
+    // RoPE transposed: [1, H, D]
+    q_rope: GpuBuffer,
+    k_rope: GpuBuffer,
+
+    // After RoPE transpose back: [1, H, 1, D]
+    q_final: GpuBuffer,
+    k_final: GpuBuffer,
+
+    // FFN intermediates
+    gate: GpuBuffer,     // [1, intermediate]
+    up: GpuBuffer,       // [1, intermediate]
+    silu_out: GpuBuffer, // [1, intermediate]
+
+    // GEMV fp32 scratch for deterministic K-block partials.
+    fp32_hidden: GpuBuffer,       // for hidden-sized GEMV outputs
+    fp32_q: GpuBuffer,            // for Q projection
+    fp32_kv: GpuBuffer,           // for K/V projection
+    fp32_intermediate: GpuBuffer, // for gate/up projections
+    fp32_vocab: GpuBuffer,        // for lm_head
+
+    // Token ID and position (GPU-resident, updated before replay)
+    token_id_gpu: GpuBuffer, // 4 bytes (u32)
+    position_gpu: GpuBuffer, // 4 bytes (u32)
+
+    // Final output
+    logits: GpuBuffer, // [1, vocab_size]
+}
+
+pub struct DecodeGraphState {
+    stream: CudaStream,
+    buffers: DecodeBuffers,
+
+    // Per-layer graph pairs
+    pre_attn_graphs: Vec<CudaGraph>,
+    post_attn_graphs: Vec<CudaGraph>,
+    final_graph: CudaGraph,
+
+    captured: bool,
+
+    // Model dimensions
+    hidden: usize,
+    num_heads: usize,
+    num_kv_heads: usize,
+    head_dim: usize,
+    intermediate: usize,
+    vocab_size: usize,
+    num_layers: usize,
+    eps: f32,
+}
+
+impl DecodeGraphState {
+    pub fn new(config: &ModelConfig) -> Self {
+        let hidden = config.hidden();
+        let num_heads = config.num_heads();
+        let num_kv_heads = config.num_kv_heads();
+        let head_dim = config.head_dim();
+        let intermediate = config.ffn_hidden();
+        let vocab_size = config.vocab_size;
+        let num_layers = config.num_layers();
+        let eps = config.rms_norm_eps.unwrap_or(1e-6) as f32;
+        let es = 2usize; // BF16 = 2 bytes
+
+        let stream = CudaStream::new().expect("create CUDA stream for graph");
+
+        let alloc = |size: usize| -> GpuBuffer {
+            GpuBuffer::alloc(size).expect("alloc decode graph buffer")
+        };
+
+        let buffers = DecodeBuffers {
+            x: alloc(hidden * es),
+            normed: alloc(hidden * es),
+            attn_out: alloc(num_heads * head_dim * es),
+            attn_merged: alloc(hidden * es),
+            o_proj: alloc(hidden * es),
+            normed2: alloc(hidden * es),
+            sum_out: alloc(hidden * es),
+            down: alloc(hidden * es),
+
+            q_proj: alloc(num_heads * head_dim * es),
+            k_proj: alloc(num_kv_heads * head_dim * es),
+            v_proj: alloc(num_kv_heads * head_dim * es),
+
+            q_reshaped: alloc(num_heads * head_dim * es),
+            k_reshaped: alloc(num_kv_heads * head_dim * es),
+            v_reshaped: alloc(num_kv_heads * head_dim * es),
+
+            q_normed: alloc(num_heads * head_dim * es),
+            k_normed: alloc(num_kv_heads * head_dim * es),
+
+            q_rope: alloc(num_heads * head_dim * es),
+            k_rope: alloc(num_kv_heads * head_dim * es),
+
+            q_final: alloc(num_heads * head_dim * es),
+            k_final: alloc(num_kv_heads * head_dim * es),
+
+            gate: alloc(intermediate * es),
+            up: alloc(intermediate * es),
+            silu_out: alloc(intermediate * es),
+
+            fp32_hidden: alloc(
+                gemv_scratch_elems(hidden, hidden).max(gemv_scratch_elems(intermediate, hidden))
+                    * 4,
+            ),
+            fp32_q: alloc(gemv_scratch_elems(hidden, num_heads * head_dim) * 4),
+            fp32_kv: alloc(gemv_scratch_elems(hidden, num_kv_heads * head_dim) * 4),
+            fp32_intermediate: alloc(gemv_scratch_elems(hidden, intermediate) * 4),
+            fp32_vocab: alloc(gemv_scratch_elems(hidden, vocab_size) * 4),
+
+            token_id_gpu: alloc(4),
+            position_gpu: alloc(4),
+
+            logits: alloc(vocab_size * es),
+        };
+
+        let pre_attn_graphs = (0..num_layers).map(|_| CudaGraph::new()).collect();
+        let post_attn_graphs = (0..num_layers).map(|_| CudaGraph::new()).collect();
+
+        Self {
+            stream,
+            buffers,
+            pre_attn_graphs,
+            post_attn_graphs,
+            final_graph: CudaGraph::new(),
+            captured: false,
+            hidden,
+            num_heads,
+            num_kv_heads,
+            head_dim,
+            intermediate,
+            vocab_size,
+            num_layers,
+            eps,
+        }
+    }
+
+    pub fn is_captured(&self) -> bool {
+        self.captured
+    }
+
+    /// Capture all per-layer graphs. Called once after the first decode step.
+    pub fn capture(
+        &mut self,
+        layers: &[LayerWeightPtrs],
+        norm_weight: *const c_void,
+        lm_head_wt: *const c_void,
+        _embed_table: *const c_void,
+        rope_cos: *const c_void,
+        rope_sin: *const c_void,
+    ) {
+        let s = self.stream.as_raw();
+        let h = self.hidden as i32;
+        let nh = self.num_heads as i32;
+        let nkv = self.num_kv_heads as i32;
+        let hd = self.head_dim as i32;
+        let inter = self.intermediate as i32;
+        let vocab = self.vocab_size as i32;
+        let eps = self.eps;
+
+        let cublas = cublas_handle();
+
+        // Set cuBLAS to use our stream
+        unsafe {
+            dispatch::set_cublas_stream(cublas, s);
+        }
+
+        for (l, lw) in layers.iter().enumerate() {
+            // === Pre-attention graph ===
+            self.pre_attn_graphs[l]
+                .begin_capture(&self.stream)
+                .expect("begin pre-attn capture");
+            unsafe {
+                // RMSNorm
+                dispatch::rmsnorm_bf16(
+                    self.buffers.x.as_ptr() as _,
+                    lw.input_norm,
+                    self.buffers.normed.as_mut_ptr() as _,
+                    1,
+                    h,
+                    eps,
+                    s,
+                );
+
+                // Q projection (GEMV)
+                dispatch::gemv_bf16(
+                    self.buffers.normed.as_ptr() as _,
+                    lw.q_proj_wt,
+                    self.buffers.q_proj.as_mut_ptr() as _,
+                    self.buffers.fp32_q.as_mut_ptr() as _,
+                    h,
+                    nh * hd,
+                    s,
+                );
+
+                // K projection (GEMV)
+                dispatch::gemv_bf16(
+                    self.buffers.normed.as_ptr() as _,
+                    lw.k_proj_wt,
+                    self.buffers.k_proj.as_mut_ptr() as _,
+                    self.buffers.fp32_kv.as_mut_ptr() as _,
+                    h,
+                    nkv * hd,
+                    s,
+                );
+
+                // V projection (GEMV)
+                dispatch::gemv_bf16(
+                    self.buffers.normed.as_ptr() as _,
+                    lw.v_proj_wt,
+                    self.buffers.v_proj.as_mut_ptr() as _,
+                    self.buffers.fp32_kv.as_mut_ptr() as _,
+                    h,
+                    nkv * hd,
+                    s,
+                );
+
+                // Reshape heads: [1, H*D] -> [1, H, 1, D]
+                dispatch::reshape_heads_bf16(
+                    self.buffers.q_proj.as_ptr() as _,
+                    self.buffers.q_reshaped.as_mut_ptr() as _,
+                    1,
+                    nh,
+                    hd,
+                    s,
+                );
+                dispatch::reshape_heads_bf16(
+                    self.buffers.k_proj.as_ptr() as _,
+                    self.buffers.k_reshaped.as_mut_ptr() as _,
+                    1,
+                    nkv,
+                    hd,
+                    s,
+                );
+                dispatch::reshape_heads_bf16(
+                    self.buffers.v_proj.as_ptr() as _,
+                    self.buffers.v_reshaped.as_mut_ptr() as _,
+                    1,
+                    nkv,
+                    hd,
+                    s,
+                );
+
+                // QK norm (head-level rmsnorm: treat [1,H,1,D] as [H, D])
+                dispatch::rmsnorm_bf16(
+                    self.buffers.q_reshaped.as_ptr() as _,
+                    lw.q_norm,
+                    self.buffers.q_normed.as_mut_ptr() as _,
+                    nh,
+                    hd,
+                    eps,
+                    s,
+                );
+                dispatch::rmsnorm_bf16(
+                    self.buffers.k_reshaped.as_ptr() as _,
+                    lw.k_norm,
+                    self.buffers.k_normed.as_mut_ptr() as _,
+                    nkv,
+                    hd,
+                    eps,
+                    s,
+                );
+
+                // Transpose for RoPE: [1,H,1,D] -> [1,H,D]
+                dispatch::transpose_hsd_to_shd_bf16(
+                    self.buffers.q_normed.as_ptr() as _,
+                    self.buffers.q_rope.as_mut_ptr() as _,
+                    1,
+                    nh,
+                    hd,
+                    s,
+                );
+                dispatch::transpose_hsd_to_shd_bf16(
+                    self.buffers.k_normed.as_ptr() as _,
+                    self.buffers.k_rope.as_mut_ptr() as _,
+                    1,
+                    nkv,
+                    hd,
+                    s,
+                );
+
+                // RoPE (in-place, reads position_gpu)
+                dispatch::rope_bf16(
+                    self.buffers.q_rope.as_mut_ptr() as _,
+                    rope_cos,
+                    rope_sin,
+                    self.buffers.position_gpu.as_ptr() as _,
+                    1,
+                    nh,
+                    hd,
+                    s,
+                );
+                dispatch::rope_bf16(
+                    self.buffers.k_rope.as_mut_ptr() as _,
+                    rope_cos,
+                    rope_sin,
+                    self.buffers.position_gpu.as_ptr() as _,
+                    1,
+                    nkv,
+                    hd,
+                    s,
+                );
+
+                // Transpose back: [1,H,D] -> [1,H,1,D]
+                dispatch::transpose_shd_to_hsd_bf16(
+                    self.buffers.q_rope.as_ptr() as _,
+                    self.buffers.q_final.as_mut_ptr() as _,
+                    1,
+                    nh,
+                    hd,
+                    s,
+                );
+                dispatch::transpose_shd_to_hsd_bf16(
+                    self.buffers.k_rope.as_ptr() as _,
+                    self.buffers.k_final.as_mut_ptr() as _,
+                    1,
+                    nkv,
+                    hd,
+                    s,
+                );
+            }
+            self.pre_attn_graphs[l]
+                .end_capture(&self.stream)
+                .expect("end pre-attn capture");
+
+            // === Post-attention graph ===
+            self.post_attn_graphs[l]
+                .begin_capture(&self.stream)
+                .expect("begin post-attn capture");
+            unsafe {
+                // Merge heads: [1,H,1,D] -> [1, hidden]
+                // attn_out is written by ungraphed attention
+                dispatch::merge_heads_bf16(
+                    self.buffers.attn_out.as_ptr() as _,
+                    self.buffers.attn_merged.as_mut_ptr() as _,
+                    1,
+                    nh,
+                    hd,
+                    s,
+                );
+
+                // O projection
+                dispatch::gemv_bf16(
+                    self.buffers.attn_merged.as_ptr() as _,
+                    lw.o_proj_wt,
+                    self.buffers.o_proj.as_mut_ptr() as _,
+                    self.buffers.fp32_hidden.as_mut_ptr() as _,
+                    nh * hd,
+                    h,
+                    s,
+                );
+
+                // Fused Add+RMSNorm: normed2 = rmsnorm(o_proj + x), sum_out = o_proj + x
+                dispatch::add_rmsnorm_bf16(
+                    self.buffers.o_proj.as_ptr() as _,
+                    self.buffers.x.as_ptr() as _,
+                    lw.post_norm,
+                    self.buffers.normed2.as_mut_ptr() as _,
+                    self.buffers.sum_out.as_mut_ptr() as _,
+                    1,
+                    h,
+                    eps,
+                    s,
+                );
+
+                // Gate projection
+                dispatch::gemv_bf16(
+                    self.buffers.normed2.as_ptr() as _,
+                    lw.gate_proj_wt,
+                    self.buffers.gate.as_mut_ptr() as _,
+                    self.buffers.fp32_intermediate.as_mut_ptr() as _,
+                    h,
+                    inter,
+                    s,
+                );
+
+                // Up projection
+                dispatch::gemv_bf16(
+                    self.buffers.normed2.as_ptr() as _,
+                    lw.up_proj_wt,
+                    self.buffers.up.as_mut_ptr() as _,
+                    self.buffers.fp32_intermediate.as_mut_ptr() as _,
+                    h,
+                    inter,
+                    s,
+                );
+
+                // Fused SiLU x Mul
+                dispatch::silu_mul_bf16(
+                    self.buffers.gate.as_ptr() as _,
+                    self.buffers.up.as_ptr() as _,
+                    self.buffers.silu_out.as_mut_ptr() as _,
+                    inter,
+                    s,
+                );
+
+                // Down projection
+                dispatch::gemv_bf16(
+                    self.buffers.silu_out.as_ptr() as _,
+                    lw.down_proj_wt,
+                    self.buffers.down.as_mut_ptr() as _,
+                    self.buffers.fp32_hidden.as_mut_ptr() as _,
+                    inter,
+                    h,
+                    s,
+                );
+
+                // x = sum_out + down (residual connection for next layer)
+                dispatch::add_bf16(
+                    self.buffers.sum_out.as_ptr() as _,
+                    self.buffers.down.as_ptr() as _,
+                    self.buffers.x.as_mut_ptr() as _,
+                    h,
+                    s,
+                );
+            }
+            self.post_attn_graphs[l]
+                .end_capture(&self.stream)
+                .expect("end post-attn capture");
+        }
+
+        // === Final graph: norm + lm_head ===
+        self.final_graph
+            .begin_capture(&self.stream)
+            .expect("begin final capture");
+        unsafe {
+            dispatch::rmsnorm_bf16(
+                self.buffers.x.as_ptr() as _,
+                norm_weight,
+                self.buffers.normed.as_mut_ptr() as _,
+                1,
+                h,
+                eps,
+                s,
+            );
+            dispatch::gemv_bf16(
+                self.buffers.normed.as_ptr() as _,
+                lm_head_wt,
+                self.buffers.logits.as_mut_ptr() as _,
+                self.buffers.fp32_vocab.as_mut_ptr() as _,
+                h,
+                vocab,
+                s,
+            );
+        }
+        self.final_graph
+            .end_capture(&self.stream)
+            .expect("end final capture");
+
+        // Reset cuBLAS back to null stream
+        unsafe {
+            dispatch::set_cublas_stream(cublas, std::ptr::null_mut());
+        }
+
+        self.captured = true;
+    }
+
+    /// Execute a single decode step using captured graphs.
+    pub fn execute(
+        &mut self,
+        token_id: u32,
+        position: u32,
+        cache: &mut GpuKVCache,
+        _layers: &[LayerWeightPtrs],
+        embed_table: *const c_void,
+        vocab_size: i32,
+        hidden_size: i32,
+    ) {
+        assert!(self.captured, "must call capture() before execute()");
+        let s = self.stream.as_raw();
+        let nkv = self.num_kv_heads;
+        let nh = self.num_heads;
+        let hd = self.head_dim;
+        let es = 2usize; // BF16
+
+        // Upload token ID and position to fixed GPU buffers
+        self.buffers
+            .token_id_gpu
+            .copy_from_host(&token_id.to_le_bytes())
+            .unwrap();
+        self.buffers
+            .position_gpu
+            .copy_from_host(&position.to_le_bytes())
+            .unwrap();
+
+        // Embedding (outside graph since token_id changes each step)
+        unsafe {
+            dispatch::embedding_bf16(
+                embed_table,
+                self.buffers.token_id_gpu.as_ptr() as _,
+                self.buffers.x.as_mut_ptr() as _,
+                1,
+                hidden_size,
+                vocab_size,
+                s,
+            );
+        }
+
+        for l in 0..self.num_layers {
+            // Pre-attention graph (norm + QKV + reshape + QK-norm + RoPE)
+            self.pre_attn_graphs[l]
+                .launch(&self.stream)
+                .expect("launch pre-attn graph");
+
+            // Ungraphed: KV cache append
+            // k_final shape: [1, num_kv_heads, 1, head_dim] (after RoPE pipeline)
+            // v_reshaped shape: [1, num_kv_heads, 1, head_dim] (V skips RoPE)
+            let pos = position as usize;
+
+            let k_buf_size = nkv * hd * es;
+            let v_buf_size = nkv * hd * es;
+            let shape = [1usize, nkv, 1, hd];
+
+            // Synchronize before accessing buffers for KV cache append
+            self.stream.synchronize().expect("sync before kv cache");
+
+            let k_view = unsafe {
+                crate::kv_cache::tensor_from_gpu_buffer_pub(
+                    GpuBuffer::borrow_raw(self.buffers.k_final.as_mut_ptr(), k_buf_size),
+                    &shape,
+                    xserv_tensor::DType::BF16,
+                    0,
+                )
+            };
+            let v_view = unsafe {
+                crate::kv_cache::tensor_from_gpu_buffer_pub(
+                    GpuBuffer::borrow_raw(self.buffers.v_reshaped.as_mut_ptr(), v_buf_size),
+                    &shape,
+                    xserv_tensor::DType::BF16,
+                    0,
+                )
+            };
+            cache.append(l, &k_view, &v_view, 1, pos);
+
+            // Ungraphed: get full KV cache and run decode attention
+            let (k_full, v_full) = cache.get_kv_len(l, pos + 1);
+            let kv_len = (pos + 1) as i32;
+            let scale = 1.0 / (hd as f32).sqrt();
+
+            // Attention output written to attn_out (separate from q_final)
+            unsafe {
+                dispatch::decode_attention_bf16(
+                    self.buffers.q_final.as_ptr() as _,
+                    k_full.data_ptr() as _,
+                    v_full.data_ptr() as _,
+                    self.buffers.attn_out.as_mut_ptr() as _,
+                    1,
+                    nh as i32,
+                    nkv as i32,
+                    kv_len,
+                    hd as i32,
+                    scale,
+                    s,
+                );
+            }
+
+            // Synchronize before post-attention graph reads attn_out
+            self.stream.synchronize().expect("sync before post-attn");
+
+            // Post-attention graph (merge + O-proj + add_rmsnorm + FFN + residual)
+            self.post_attn_graphs[l]
+                .launch(&self.stream)
+                .expect("launch post-attn graph");
+        }
+
+        // Final graph (norm + lm_head)
+        self.final_graph
+            .launch(&self.stream)
+            .expect("launch final graph");
+
+        // Sync to ensure logits are ready
+        self.stream.synchronize().expect("sync after decode");
+    }
+
+    /// Get the logits buffer (for reading results after execute).
+    pub fn logits_buffer(&self) -> &GpuBuffer {
+        &self.buffers.logits
+    }
+
+    /// Invalidate captured graphs (e.g. when switching sequences).
+    pub fn invalidate(&mut self) {
+        self.captured = false;
+        self.pre_attn_graphs = (0..self.num_layers).map(|_| CudaGraph::new()).collect();
+        self.post_attn_graphs = (0..self.num_layers).map(|_| CudaGraph::new()).collect();
+        self.final_graph = CudaGraph::new();
+    }
+}
+
+unsafe impl Send for DecodeGraphState {}
+
+/// Lightweight struct holding raw pointers to a layer's weight tensors.
+/// Used to avoid passing the full model struct into the graph capture code.
+pub struct LayerWeightPtrs {
+    pub input_norm: *const c_void,
+    pub q_proj_wt: *const c_void,
+    pub k_proj_wt: *const c_void,
+    pub v_proj_wt: *const c_void,
+    pub o_proj_wt: *const c_void,
+    pub q_norm: *const c_void,
+    pub k_norm: *const c_void,
+    pub post_norm: *const c_void,
+    pub gate_proj_wt: *const c_void,
+    pub up_proj_wt: *const c_void,
+    pub down_proj_wt: *const c_void,
+}
+
+unsafe impl Send for LayerWeightPtrs {}
+unsafe impl Sync for LayerWeightPtrs {}

crates/xserv-model/src/eagle3.rs

@@ -0,0 +1,425 @@
+//! EAGLE3 speculative draft head for Qwen3-8B (Phase 25).
+//!
+//! Loads the AngelSlim/Qwen3-8B_eagle3 pytorch_model.bin and provides a
+//! single-step forward pass that takes 3 target hidden states + the previous
+//! token and returns a draft token in the target vocabulary.
+//!
+//! Architecture (from weights):
+//! - fc:           [hidden, 3*hidden] → fuse 3 target hidden states
+//! - midlayer:     1 decoder layer (attn input dim = 2*hidden)
+//! - norm + lm_head: → [draft_vocab_size=32000]
+//! - d2t:          draft_id → target_id offset mapping
+
+use std::collections::HashMap;
+use std::path::Path;
+use xserv_kernels::*;
+use xserv_tensor::{DType, Device, Tensor};
+
+/// Target layers to hook for EAGLE3 auxiliary hidden states, for Qwen3-8B
+/// (36 layers). Value comes from AngelSlim/vLLM speculators training config
+/// `dflash_qwen3_8b_sharegpt_online_5k.sh` which specifies target_layer_ids
+/// = "2 18 33". Must match training-time selection or EAGLE outputs are wrong.
+pub const EAGLE_HOOK_LAYERS: [usize; 3] = [2, 18, 33];
+const DRAFT_VOCAB_SIZE: usize = 32000;
+
+fn matmul_2d(a: &Tensor, b: &Tensor) -> Tensor {
+    assert_eq!(a.ndim(), 2);
+    assert_eq!(b.ndim(), 2);
+    matmul(a, b, GemmBackend::CuBlas)
+}
+
+pub struct Eagle3Head {
+    fc_wt: Tensor,                    // [hidden, 3*hidden] transposed for matmul
+    hidden_norm: Tensor,              // [hidden]
+    input_layernorm: Tensor,          // [hidden]
+    q_proj_wt: Tensor,                // [num_heads*head_dim, 2*hidden]
+    k_proj_wt: Tensor,                // [num_kv_heads*head_dim, 2*hidden]
+    v_proj_wt: Tensor,                // [num_kv_heads*head_dim, 2*hidden]
+    o_proj_wt: Tensor,                // [hidden, num_heads*head_dim]
+    gate_proj_wt: Tensor,             // [intermediate, hidden]
+    up_proj_wt: Tensor,               // [intermediate, hidden]
+    down_proj_wt: Tensor,             // [hidden, intermediate]
+    post_attention_layernorm: Tensor, // [hidden]
+    norm: Tensor,                     // [hidden] final
+    lm_head_wt: Tensor,               // [draft_vocab, hidden]
+    d2t: Vec<i64>,                    // [draft_vocab] offset mapping
+    /// t2d[target_id] = true iff target_id has a corresponding draft-vocab id
+    /// (i.e. can potentially be produced by EAGLE). Used to measure the
+    /// coverage cap on acceptance.
+    t2d: Vec<bool>,
+    hidden_size: usize,
+    num_heads: usize,
+    num_kv_heads: usize,
+    head_dim: usize,
+    max_seq_len: usize,
+    rope_cache: RopeCache,
+    // Stateful 1-layer KV cache: [1, num_kv_heads, max_seq_len, head_dim] BF16.
+    // We slice `..current_len` for attention. The head is tiny (~64 KB per
+    // 1000 tokens) so pre-allocating max_seq_len wastes negligible memory.
+    k_cache: Tensor,
+    v_cache: Tensor,
+    current_len: usize,
+}
+
+impl Eagle3Head {
+    pub fn load(dir: &Path, device: u32) -> Self {
+        let (weights, d2t, t2d) = load_eagle3_weights(dir, device);
+        let hidden_size = 4096;
+        let num_heads = 32;
+        let num_kv_heads = 8;
+        let head_dim = 128;
+        let intermediate_size = 12288;
+        let max_seq_len = 2048;
+        let rope_theta = 1_000_000.0f32;
+
+        let get = |name: &str| -> Tensor {
+            weights
+                .get(name)
+                .unwrap_or_else(|| panic!("missing eagle3 weight: {name}"))
+                .clone()
+        };
+
+        let fc_wt = get("fc.weight").transpose(0, 1).contiguous();
+        let q_proj_wt = get("midlayer.self_attn.q_proj.weight")
+            .transpose(0, 1)
+            .contiguous();
+        let k_proj_wt = get("midlayer.self_attn.k_proj.weight")
+            .transpose(0, 1)
+            .contiguous();
+        let v_proj_wt = get("midlayer.self_attn.v_proj.weight")
+            .transpose(0, 1)
+            .contiguous();
+        let o_proj_wt = get("midlayer.self_attn.o_proj.weight")
+            .transpose(0, 1)
+            .contiguous();
+        let gate_proj_wt = get("midlayer.mlp.gate_proj.weight")
+            .transpose(0, 1)
+            .contiguous();
+        let up_proj_wt = get("midlayer.mlp.up_proj.weight")
+            .transpose(0, 1)
+            .contiguous();
+        let down_proj_wt = get("midlayer.mlp.down_proj.weight")
+            .transpose(0, 1)
+            .contiguous();
+        let hidden_norm = get("midlayer.hidden_norm.weight");
+        let input_layernorm = get("midlayer.input_layernorm.weight");
+        let post_attention_layernorm = get("midlayer.post_attention_layernorm.weight");
+        let norm = get("norm.weight");
+        let lm_head_wt = get("lm_head.weight").transpose(0, 1).contiguous();
+
+        assert_eq!(d2t.len(), DRAFT_VOCAB_SIZE);
+
+        let rope_cache = RopeCache::new(max_seq_len, head_dim, rope_theta);
+
+        let k_cache = Tensor::zeros(
+            &[1, num_kv_heads, max_seq_len, head_dim],
+            DType::BF16,
+            Device::Cuda(device),
+        );
+        let v_cache = Tensor::zeros(
+            &[1, num_kv_heads, max_seq_len, head_dim],
+            DType::BF16,
+            Device::Cuda(device),
+        );
+
+        Self {
+            fc_wt,
+            hidden_norm,
+            input_layernorm,
+            q_proj_wt,
+            k_proj_wt,
+            v_proj_wt,
+            o_proj_wt,
+            gate_proj_wt,
+            up_proj_wt,
+            down_proj_wt,
+            post_attention_layernorm,
+            norm,
+            lm_head_wt,
+            d2t,
+            t2d,
+            hidden_size,
+            num_heads,
+            num_kv_heads,
+            head_dim,
+            max_seq_len,
+            rope_cache,
+            k_cache,
+            v_cache,
+            current_len: 0,
+        }
+    }
+
+    /// Reset the internal KV cache for a fresh sequence.
+    pub fn reset(&mut self) {
+        self.current_len = 0;
+    }
+
+    /// Truncate the internal KV cache to `new_len` entries. Used to discard
+    /// K/V of rejected drafts after a speculative round.
+    pub fn truncate_to(&mut self, new_len: usize) {
+        assert!(new_len <= self.current_len);
+        self.current_len = new_len;
+    }
+
+    /// Current number of committed K/V entries in the internal EAGLE cache.
+    pub fn current_len(&self) -> usize {
+        self.current_len
+    }
+
+    /// One draft step: produce a token in target vocabulary space.
+    ///
+    /// - `target_hidden`: 3 tensors [1, hidden_size] from target hook layers
+    /// - `embed_table`: the target model's embed_tokens (shared, not copied)
+    /// - `prev_token`: the previous committed token
+    /// - `position`: the decode position for RoPE
+    ///
+    /// Returns (draft_token_in_target_vocab, draft_logits_tensor).
+    pub fn step(
+        &mut self,
+        target_hidden: &[Tensor; 3],
+        embed_table: &Tensor,
+        prev_token: u32,
+        position: usize,
+    ) -> (u32, Tensor) {
+        let (id, logits, _) = self.step_with_aux(target_hidden, embed_table, prev_token, position);
+        (id, logits)
+    }
+
+    /// Like `step`, but also returns the final hidden state (aux) usable as
+    /// the fused_h for a subsequent recursive draft step via `step_recursive`.
+    pub fn step_with_aux(
+        &mut self,
+        target_hidden: &[Tensor; 3],
+        embed_table: &Tensor,
+        prev_token: u32,
+        position: usize,
+    ) -> (u32, Tensor, Tensor) {
+        // Fuse 3 target hidden states into fused_h via fc.
+        let h_cat = concat_hidden(target_hidden);
+        let fused_h = matmul_2d(&h_cat, &self.fc_wt);
+        self.forward_from_fused(fused_h, embed_table, prev_token, position)
+    }
+
+    /// Recursive draft step: reuses the previous EAGLE step's aux as fused_h,
+    /// bypassing the fc+3-hidden fusion. Used for γ≥2 chained drafts.
+    pub fn step_recursive(
+        &mut self,
+        fused_h: Tensor,
+        embed_table: &Tensor,
+        prev_token: u32,
+        position: usize,
+    ) -> (u32, Tensor, Tensor) {
+        self.forward_from_fused(fused_h, embed_table, prev_token, position)
+    }
+
+    fn forward_from_fused(
+        &mut self,
+        fused_h: Tensor,
+        embed_table: &Tensor,
+        prev_token: u32,
+        position: usize,
+    ) -> (u32, Tensor, Tensor) {
+        let eps = 1e-6f32;
+        assert!(
+            self.current_len < self.max_seq_len,
+            "EAGLE KV cache overflow: {} >= {}",
+            self.current_len,
+            self.max_seq_len
+        );
+
+        let emb = embedding(embed_table, &[prev_token]);
+        let residual = fused_h.clone();
+        let emb_normed = rmsnorm(&emb, &self.input_layernorm, eps);
+        let h_normed = rmsnorm(&fused_h, &self.hidden_norm, eps);
+        let attn_in = concat_last_dim(&emb_normed, &h_normed);
+
+        let q = matmul_2d(&attn_in, &self.q_proj_wt);
+        let k = matmul_2d(&attn_in, &self.k_proj_wt);
+        let v = matmul_2d(&attn_in, &self.v_proj_wt);
+
+        let q_3d = q.reshape(&[1, self.num_heads, self.head_dim]);
+        let k_3d = k.reshape(&[1, self.num_kv_heads, self.head_dim]);
+        let positions = [position as u32];
+        rope_inplace(&q_3d, &self.rope_cache, &positions);
+        rope_inplace(&k_3d, &self.rope_cache, &positions);
+
+        let v_3d = v.reshape(&[1, self.num_kv_heads, self.head_dim]);
+        self.append_to_kv_cache(&k_3d, &v_3d);
+        self.current_len += 1;
+        let kv_len = self.current_len;
+        let k_view = self.k_cache.narrow(2, 0, kv_len).contiguous();
+        let v_view = self.v_cache.narrow(2, 0, kv_len).contiguous();
+
+        let q_4d = q_3d.reshape(&[1, self.num_heads, 1, self.head_dim]);
+        let attn_out = decode_attention(&q_4d, &k_view, &v_view);
+
+        let attn_merged = attn_out.reshape(&[1, self.num_heads * self.head_dim]);
+        let attn_proj = matmul_2d(&attn_merged, &self.o_proj_wt);
+
+        let (mlp_in, residual) =
+            add_rmsnorm(&attn_proj, &residual, &self.post_attention_layernorm, eps);
+
+        let gate = matmul_2d(&mlp_in, &self.gate_proj_wt);
+        let up = matmul_2d(&mlp_in, &self.up_proj_wt);
+        let hidden = silu_mul(&gate, &up);
+        let down = matmul_2d(&hidden, &self.down_proj_wt);
+
+        let (x, prenorm) = add_rmsnorm(&down, &residual, &self.norm, eps);
+        let logits = matmul_2d(&x, &self.lm_head_wt);
+
+        let draft_id = argmax_bf16_single(&logits);
+        let target_id = (draft_id as i64 + self.d2t[draft_id as usize]) as u32;
+        // aux for recursive drafting = PRE-norm hidden (default norm_output=False
+        // in vllm/llama_eagle3.py). Feeding the pre-norm state matches training.
+        (target_id, logits, prenorm)
+    }
+
+    /// Write new K/V rows (shape [1, num_kv_heads, head_dim]) at position
+    /// `current_len` inside the [1, num_kv_heads, max_seq_len, head_dim] cache.
+    fn append_to_kv_cache(&mut self, new_k: &Tensor, new_v: &Tensor) {
+        let head_bytes = self.head_dim * self.k_cache.dtype().size_bytes();
+        for h in 0..self.num_kv_heads {
+            for (cache, src) in [(&self.k_cache, new_k), (&self.v_cache, new_v)] {
+                let dst = unsafe {
+                    (cache.data_ptr() as *mut u8)
+                        .add(((h * self.max_seq_len) + self.current_len) * head_bytes)
+                };
+                let s = unsafe { (src.data_ptr() as *const u8).add(h * head_bytes) };
+                d2d(dst, s, head_bytes);
+            }
+        }
+    }
+
+    /// Map a draft-vocab token id to the full target-vocab id via d2t.
+    pub fn map_draft_to_target(&self, draft_id: u32) -> u32 {
+        (draft_id as i64 + self.d2t[draft_id as usize]) as u32
+    }
+
+    /// Returns true iff `target_id` is representable in the draft vocabulary
+    /// (i.e., EAGLE could in principle produce it).
+    pub fn target_id_in_draft_vocab(&self, target_id: u32) -> bool {
+        self.t2d.get(target_id as usize).copied().unwrap_or(false)
+    }
+}
+
+fn d2d(dst: *mut u8, src: *const u8, bytes: usize) {
+    unsafe {
+        xserv_cuda::ffi::cudaMemcpy(dst, src, bytes, xserv_cuda::ffi::CUDA_MEMCPY_D2D);
+    }
+}
+
+fn concat_hidden(hidden: &[Tensor; 3]) -> Tensor {
+    let h = hidden[0].shape()[1];
+    let dtype = hidden[0].dtype();
+    let device = hidden[0].device();
+    let elem_bytes = dtype.size_bytes();
+    let out = Tensor::empty(&[1, 3 * h], dtype, device);
+    for (i, t) in hidden.iter().enumerate() {
+        assert!(t.is_contiguous());
+        let dst = unsafe { (out.data_ptr() as *mut u8).add(i * h * elem_bytes) };
+        d2d(dst, t.data_ptr() as *const u8, h * elem_bytes);
+    }
+    out
+}
+
+fn concat_last_dim(a: &Tensor, b: &Tensor) -> Tensor {
+    let da = a.shape()[1];
+    let db = b.shape()[1];
+    let dtype = a.dtype();
+    let device = a.device();
+    let elem_bytes = dtype.size_bytes();
+    let out = Tensor::empty(&[1, da + db], dtype, device);
+    d2d(
+        out.data_ptr() as *mut u8,
+        a.data_ptr() as *const u8,
+        da * elem_bytes,
+    );
+    let dst = unsafe { (out.data_ptr() as *mut u8).add(da * elem_bytes) };
+    d2d(dst, b.data_ptr() as *const u8, db * elem_bytes);
+    out
+}
+
+fn repeat_kv_for_single_token(kv: &Tensor, repeats: usize) -> Tensor {
+    if repeats == 1 {
+        return kv.clone();
+    }
+    let nkv = kv.shape()[1];
+    let d = kv.shape()[2];
+    let dtype = kv.dtype();
+    let device = kv.device();
+    let head_bytes = d * dtype.size_bytes();
+    let out = Tensor::empty(&[1, nkv * repeats, d], dtype, device);
+    for h in 0..nkv {
+        let src = unsafe { (kv.data_ptr() as *const u8).add(h * head_bytes) };
+        for r in 0..repeats {
+            let dst = unsafe { (out.data_ptr() as *mut u8).add((h * repeats + r) * head_bytes) };
+            d2d(dst, src, head_bytes);
+        }
+    }
+    out
+}
+
+/// Load EAGLE3 weights from safetensors, handling int64 d2t + bool t2d specially.
+fn load_eagle3_weights(dir: &Path, device: u32) -> (HashMap<String, Tensor>, Vec<i64>, Vec<bool>) {
+    let st_path = dir.join("model.safetensors");
+    assert!(
+        st_path.exists(),
+        "Eagle3 model.safetensors not found in {}. Convert with:\n\
+         python3 -c \"import torch; from safetensors.torch import save_file; \
+         sd=torch.load('pytorch_model.bin', map_location='cpu', weights_only=False); \
+         save_file(sd, 'model.safetensors')\"",
+        dir.display()
+    );
+
+    let data = std::fs::read(&st_path)
+        .unwrap_or_else(|e| panic!("failed to read {}: {e}", st_path.display()));
+    let st = safetensors::SafeTensors::deserialize(&data)
+        .unwrap_or_else(|e| panic!("failed to parse {}: {e}", st_path.display()));
+
+    let mut tensors = HashMap::new();
+    let mut d2t_vec: Vec<i64> = Vec::new();
+    let mut t2d_vec: Vec<bool> = Vec::new();
+
+    for (name, view) in st.tensors() {
+        if name == "t2d" {
+            let raw = view.data();
+            assert_eq!(view.dtype(), safetensors::Dtype::BOOL);
+            t2d_vec = raw.iter().map(|&b| b != 0).collect();
+            continue;
+        }
+        if name == "d2t" {
+            let raw = view.data();
+            assert_eq!(view.dtype(), safetensors::Dtype::I64);
+            let n = raw.len() / 8;
+            d2t_vec = (0..n)
+                .map(|i| i64::from_le_bytes(raw[i * 8..(i + 1) * 8].try_into().unwrap()))
+                .collect();
+            continue;
+        }
+        let dtype = match view.dtype() {
+            safetensors::Dtype::BF16 => DType::BF16,
+            safetensors::Dtype::F32 => DType::F32,
+            safetensors::Dtype::F16 => DType::F16,
+            other => {
+                eprintln!("eagle3: skipping {name} with unsupported dtype {other:?}");
+                continue;
+            }
+        };
+        let shape: Vec<usize> = view.shape().to_vec();
+        let raw = view.data();
+        let t = crate::loader::make_tensor(raw, &shape, dtype);
+        let t = t.to_device(Device::Cuda(device));
+        tensors.insert(name.to_string(), t);
+    }
+
+    assert!(
+        !d2t_vec.is_empty(),
+        "d2t tensor not found in eagle3 weights"
+    );
+    assert!(
+        !t2d_vec.is_empty(),
+        "t2d tensor not found in eagle3 weights"
+    );
+    (tensors, d2t_vec, t2d_vec)
+}

crates/xserv-model/src/gpt2.rs

@@ -0,0 +1,437 @@
+use std::collections::HashMap;
+use xserv_kernels::*;
+use xserv_tensor::{DType, Device, Tensor};
+
+use crate::config::ModelConfig;
+
+pub struct GPT2 {
+    pub config: ModelConfig,
+    wte: Tensor,
+    wpe: Tensor,
+    layers: Vec<GPT2Block>,
+    ln_f_g: Tensor,
+    ln_f_b: Tensor,
+    lm_head: Tensor, // precomputed wte^T
+}
+
+struct GPT2Block {
+    ln_1_g: Tensor,
+    ln_1_b: Tensor,
+    attn_qkv_w: Tensor,
+    attn_qkv_b: Tensor,
+    attn_out_w: Tensor,
+    attn_out_b: Tensor,
+    ln_2_g: Tensor,
+    ln_2_b: Tensor,
+    mlp_fc_w: Tensor,
+    mlp_fc_b: Tensor,
+    mlp_proj_w: Tensor,
+    mlp_proj_b: Tensor,
+}
+
+pub struct KVCache {
+    // Per layer, per head: raw bytes (works for both f32 and bf16)
+    k: Vec<Vec<Vec<u8>>>, // [num_layers][num_heads][seq_len * head_dim * elem_size]
+    v: Vec<Vec<Vec<u8>>>,
+    len: usize,
+    num_heads: usize,
+    head_dim: usize,
+    elem_size: usize,
+    dtype: DType,
+    device: Device,
+}
+
+impl KVCache {
+    pub fn new(
+        num_layers: usize,
+        num_heads: usize,
+        head_dim: usize,
+        dtype: DType,
+        device: Device,
+    ) -> Self {
+        Self {
+            k: (0..num_layers).map(|_| vec![vec![]; num_heads]).collect(),
+            v: (0..num_layers).map(|_| vec![vec![]; num_heads]).collect(),
+            len: 0,
+            num_heads,
+            head_dim,
+            elem_size: dtype.size_bytes(),
+            dtype,
+            device,
+        }
+    }
+
+    pub fn seq_len(&self) -> usize {
+        self.len
+    }
+
+    /// Append from a CPU tensor with shape [1, H, new_tokens, D].
+    pub fn append_kv_tensor(
+        &mut self,
+        layer: usize,
+        k_cpu: &Tensor,
+        v_cpu: &Tensor,
+        new_tokens: usize,
+    ) {
+        let hd = self.head_dim;
+        let es = self.elem_size;
+        let k_bytes = k_cpu.storage().as_cpu_bytes();
+        let v_bytes = v_cpu.storage().as_cpu_bytes();
+        let chunk = new_tokens * hd * es;
+        for h in 0..self.num_heads {
+            let off = h * chunk;
+            self.k[layer][h].extend_from_slice(&k_bytes[off..off + chunk]);
+            self.v[layer][h].extend_from_slice(&v_bytes[off..off + chunk]);
+        }
+        if layer == 0 {
+            self.len += new_tokens;
+        }
+    }
+
+    /// Reconstruct [1, H, seq_len, D] tensors.
+    pub fn get_kv_tensors(&self, layer: usize) -> (Tensor, Tensor) {
+        let sl = self.len;
+        let hd = self.head_dim;
+        let nh = self.num_heads;
+        let es = self.elem_size;
+        let head_bytes = sl * hd * es;
+        let total = nh * head_bytes;
+        let mut k_data = vec![0u8; total];
+        let mut v_data = vec![0u8; total];
+        for h in 0..nh {
+            let off = h * head_bytes;
+            k_data[off..off + head_bytes].copy_from_slice(&self.k[layer][h]);
+            v_data[off..off + head_bytes].copy_from_slice(&self.v[layer][h]);
+        }
+        let shape = &[1, nh, sl, hd];
+        let k = tensor_from_raw_bytes(&k_data, shape, self.dtype).to_device(self.device);
+        let v = tensor_from_raw_bytes(&v_data, shape, self.dtype).to_device(self.device);
+        (k, v)
+    }
+}
+
+fn tensor_from_raw_bytes(bytes: &[u8], shape: &[usize], dtype: DType) -> Tensor {
+    match dtype {
+        DType::F32 => {
+            let data: &[f32] = unsafe {
+                std::slice::from_raw_parts(bytes.as_ptr() as *const f32, bytes.len() / 4)
+            };
+            Tensor::from_slice(data, shape)
+        }
+        DType::BF16 => {
+            let data: &[half::bf16] = unsafe {
+                std::slice::from_raw_parts(bytes.as_ptr() as *const half::bf16, bytes.len() / 2)
+            };
+            Tensor::from_slice(data, shape)
+        }
+        _ => panic!("unsupported dtype for KV cache"),
+    }
+}
+
+impl GPT2 {
+    pub fn from_weights(config: ModelConfig, mut w: HashMap<String, Tensor>) -> Self {
+        crate::init_kernels();
+        let take = |w: &mut HashMap<String, Tensor>, name: &str| -> Tensor {
+            w.remove(name)
+                .unwrap_or_else(|| panic!("missing weight: {name}"))
+        };
+
+        let wte = take(&mut w, "wte.weight");
+        let wpe = take(&mut w, "wpe.weight");
+        let ln_f_g = take(&mut w, "ln_f.weight");
+        let ln_f_b = take(&mut w, "ln_f.bias");
+        let lm_head = wte.transpose(0, 1).contiguous();
+
+        let num_layers = config.num_layers();
+        let mut layers = Vec::with_capacity(num_layers);
+        for i in 0..num_layers {
+            let p = format!("h.{i}");
+            layers.push(GPT2Block {
+                ln_1_g: take(&mut w, &format!("{p}.ln_1.weight")),
+                ln_1_b: take(&mut w, &format!("{p}.ln_1.bias")),
+                attn_qkv_w: take(&mut w, &format!("{p}.attn.c_attn.weight")),
+                attn_qkv_b: take(&mut w, &format!("{p}.attn.c_attn.bias")),
+                attn_out_w: take(&mut w, &format!("{p}.attn.c_proj.weight")),
+                attn_out_b: take(&mut w, &format!("{p}.attn.c_proj.bias")),
+                ln_2_g: take(&mut w, &format!("{p}.ln_2.weight")),
+                ln_2_b: take(&mut w, &format!("{p}.ln_2.bias")),
+                mlp_fc_w: take(&mut w, &format!("{p}.mlp.c_fc.weight")),
+                mlp_fc_b: take(&mut w, &format!("{p}.mlp.c_fc.bias")),
+                mlp_proj_w: take(&mut w, &format!("{p}.mlp.c_proj.weight")),
+                mlp_proj_b: take(&mut w, &format!("{p}.mlp.c_proj.bias")),
+            });
+        }
+
+        Self {
+            config,
+            wte,
+            wpe,
+            layers,
+            ln_f_g,
+            ln_f_b,
+            lm_head,
+        }
+    }
+
+    /// Full forward pass without KV cache (for testing / correctness comparison).
+    pub fn forward(&self, token_ids: &[u32]) -> Tensor {
+        let seq_len = token_ids.len();
+        let hidden = self.config.hidden();
+        let num_heads = self.config.num_heads();
+        let head_dim = self.config.head_dim();
+
+        let tok_emb = embedding(&self.wte, token_ids);
+        let pos_ids: Vec<u32> = (0..seq_len as u32).collect();
+        let pos_emb = embedding(&self.wpe, &pos_ids);
+        let mut x = add_tensors(&tok_emb, &pos_emb);
+
+        for layer in &self.layers {
+            x = self.transformer_block(layer, &x, None, 0, seq_len, num_heads, head_dim, hidden);
+        }
+
+        let x = layernorm(&x, &self.ln_f_g, &self.ln_f_b, self.config.ln_eps());
+        matmul_2d(&x, &self.lm_head)
+    }
+
+    /// Forward pass with KV cache. First call = prefill, subsequent = decode.
+    pub fn forward_with_cache(&self, token_ids: &[u32], cache: &mut KVCache) -> Tensor {
+        let new_tokens = token_ids.len();
+        let pos_offset = cache.seq_len();
+        let hidden = self.config.hidden();
+        let num_heads = self.config.num_heads();
+        let head_dim = self.config.head_dim();
+
+        let tok_emb = embedding(&self.wte, token_ids);
+        let pos_ids: Vec<u32> = (pos_offset..pos_offset + new_tokens)
+            .map(|p| p as u32)
+            .collect();
+        let pos_emb = embedding(&self.wpe, &pos_ids);
+        let mut x = add_tensors(&tok_emb, &pos_emb);
+
+        for (layer_idx, layer) in self.layers.iter().enumerate() {
+            x = self.transformer_block(
+                layer,
+                &x,
+                Some((cache, layer_idx)),
+                pos_offset,
+                new_tokens,
+                num_heads,
+                head_dim,
+                hidden,
+            );
+        }
+
+        let x = layernorm(&x, &self.ln_f_g, &self.ln_f_b, self.config.ln_eps());
+        matmul_2d(&x, &self.lm_head)
+    }
+
+    fn transformer_block(
+        &self,
+        layer: &GPT2Block,
+        x: &Tensor,
+        cache: Option<(&mut KVCache, usize)>,
+        _pos_offset: usize,
+        new_tokens: usize,
+        num_heads: usize,
+        head_dim: usize,
+        hidden: usize,
+    ) -> Tensor {
+        let residual = x.clone();
+        let normed = layernorm(x, &layer.ln_1_g, &layer.ln_1_b, self.config.ln_eps());
+
+        let qkv = linear(&normed, &layer.attn_qkv_w, Some(&layer.attn_qkv_b));
+        let (q, k_new, v_new) = split_qkv(&qkv, num_heads, head_dim, new_tokens);
+
+        let (k_full, v_full) = if let Some((cache, layer_idx)) = cache {
+            let k_cpu = k_new.to_device(Device::Cpu);
+            let v_cpu = v_new.to_device(Device::Cpu);
+            cache.append_kv_tensor(layer_idx, &k_cpu, &v_cpu, new_tokens);
+            cache.get_kv_tensors(layer_idx)
+        } else {
+            (k_new, v_new)
+        };
+
+        let attn_out = attention(&q, &k_full, &v_full, true);
+        let attn_out = merge_heads(&attn_out, new_tokens, hidden);
+        let attn_out = linear(&attn_out, &layer.attn_out_w, Some(&layer.attn_out_b));
+        let x = add_tensors(&residual, &attn_out);
+
+        let residual = x.clone();
+        let normed = layernorm(&x, &layer.ln_2_g, &layer.ln_2_b, self.config.ln_eps());
+        let fc = linear(&normed, &layer.mlp_fc_w, Some(&layer.mlp_fc_b));
+        let activated = gelu(&fc);
+        let proj = linear(&activated, &layer.mlp_proj_w, Some(&layer.mlp_proj_b));
+        add_tensors(&residual, &proj)
+    }
+}
+
+// --- Helper ops (unchanged) ---
+
+fn linear(x: &Tensor, weight: &Tensor, bias: Option<&Tensor>) -> Tensor {
+    let out = matmul_2d(x, weight);
+    if let Some(b) = bias {
+        add_bias(&out, b)
+    } else {
+        out
+    }
+}
+
+fn matmul_2d(a: &Tensor, b: &Tensor) -> Tensor {
+    assert_eq!(a.ndim(), 2);
+    assert_eq!(b.ndim(), 2);
+    matmul(a, b, GemmBackend::CuBlas)
+}
+
+fn add_tensors(a: &Tensor, b: &Tensor) -> Tensor {
+    xserv_kernels::add(a, b)
+}
+
+fn add_bias(x: &Tensor, bias: &Tensor) -> Tensor {
+    // bias: [N], x: [S, N] — broadcast add via reshape
+    assert_eq!(x.ndim(), 2);
+    assert_eq!(bias.ndim(), 1);
+    let n = bias.shape()[0];
+    assert_eq!(x.shape()[1], n);
+    let rows = x.shape()[0];
+    // Broadcast: tile bias to [S, N] on CPU, then GPU add
+    let b_cpu = bias.to_device(Device::Cpu);
+    match x.dtype() {
+        DType::F32 => {
+            let bd = b_cpu.as_slice::<f32>();
+            let tiled: Vec<f32> = (0..rows).flat_map(|_| bd.iter().copied()).collect();
+            let b_full = Tensor::from_slice(&tiled, x.shape()).to_device(x.device());
+            xserv_kernels::add(x, &b_full)
+        }
+        DType::BF16 => {
+            let bd = b_cpu.as_slice::<half::bf16>();
+            let tiled: Vec<half::bf16> = (0..rows).flat_map(|_| bd.iter().copied()).collect();
+            let b_full = Tensor::from_slice(&tiled, x.shape()).to_device(x.device());
+            xserv_kernels::add(x, &b_full)
+        }
+        _ => panic!("unsupported dtype"),
+    }
+}
+
+fn split_qkv(
+    qkv: &Tensor,
+    num_heads: usize,
+    head_dim: usize,
+    seq_len: usize,
+) -> (Tensor, Tensor, Tensor) {
+    let hidden = num_heads * head_dim;
+    let qkv_cpu = qkv.to_device(Device::Cpu);
+    let device = qkv.device();
+    let dtype = qkv.dtype();
+
+    match dtype {
+        DType::F32 => {
+            let data = qkv_cpu.as_slice::<f32>();
+            let mut q_data = vec![0.0f32; num_heads * seq_len * head_dim];
+            let mut k_data = vec![0.0f32; num_heads * seq_len * head_dim];
+            let mut v_data = vec![0.0f32; num_heads * seq_len * head_dim];
+            for s in 0..seq_len {
+                let row = &data[s * 3 * hidden..(s + 1) * 3 * hidden];
+                for h in 0..num_heads {
+                    let src_off = h * head_dim;
+                    let dst_off = (h * seq_len + s) * head_dim;
+                    q_data[dst_off..dst_off + head_dim]
+                        .copy_from_slice(&row[src_off..src_off + head_dim]);
+                    k_data[dst_off..dst_off + head_dim]
+                        .copy_from_slice(&row[hidden + src_off..hidden + src_off + head_dim]);
+                    v_data[dst_off..dst_off + head_dim].copy_from_slice(
+                        &row[2 * hidden + src_off..2 * hidden + src_off + head_dim],
+                    );
+                }
+            }
+            let q =
+                Tensor::from_slice(&q_data, &[1, num_heads, seq_len, head_dim]).to_device(device);
+            let k =
+                Tensor::from_slice(&k_data, &[1, num_heads, seq_len, head_dim]).to_device(device);
+            let v =
+                Tensor::from_slice(&v_data, &[1, num_heads, seq_len, head_dim]).to_device(device);
+            (q, k, v)
+        }
+        DType::BF16 => {
+            let data = qkv_cpu.as_slice::<half::bf16>();
+            let mut q_data = vec![half::bf16::ZERO; num_heads * seq_len * head_dim];
+            let mut k_data = vec![half::bf16::ZERO; num_heads * seq_len * head_dim];
+            let mut v_data = vec![half::bf16::ZERO; num_heads * seq_len * head_dim];
+            for s in 0..seq_len {
+                let row = &data[s * 3 * hidden..(s + 1) * 3 * hidden];
+                for h in 0..num_heads {
+                    let src_off = h * head_dim;
+                    let dst_off = (h * seq_len + s) * head_dim;
+                    q_data[dst_off..dst_off + head_dim]
+                        .copy_from_slice(&row[src_off..src_off + head_dim]);
+                    k_data[dst_off..dst_off + head_dim]
+                        .copy_from_slice(&row[hidden + src_off..hidden + src_off + head_dim]);
+                    v_data[dst_off..dst_off + head_dim].copy_from_slice(
+                        &row[2 * hidden + src_off..2 * hidden + src_off + head_dim],
+                    );
+                }
+            }
+            let q =
+                Tensor::from_slice(&q_data, &[1, num_heads, seq_len, head_dim]).to_device(device);
+            let k =
+                Tensor::from_slice(&k_data, &[1, num_heads, seq_len, head_dim]).to_device(device);
+            let v =
+                Tensor::from_slice(&v_data, &[1, num_heads, seq_len, head_dim]).to_device(device);
+            (q, k, v)
+        }
+        _ => panic!("unsupported dtype {:?} in split_qkv", dtype),
+    }
+}
+
+fn merge_heads(x: &Tensor, seq_len: usize, hidden: usize) -> Tensor {
+    let num_heads = x.shape()[1];
+    let head_dim = x.shape()[3];
+    let x_cpu = x.to_device(Device::Cpu);
+    let device = x.device();
+    let dtype = x.dtype();
+
+    match dtype {
+        DType::F32 => {
+            let src = x_cpu.as_slice::<f32>();
+            let mut out = vec![0.0f32; seq_len * hidden];
+            for s in 0..seq_len {
+                for h in 0..num_heads {
+                    let src_off = (h * seq_len + s) * head_dim;
+                    let dst_off = s * hidden + h * head_dim;
+                    out[dst_off..dst_off + head_dim]
+                        .copy_from_slice(&src[src_off..src_off + head_dim]);
+                }
+            }
+            Tensor::from_slice(&out, &[seq_len, hidden]).to_device(device)
+        }
+        DType::BF16 => {
+            let src = x_cpu.as_slice::<half::bf16>();
+            let mut out = vec![half::bf16::ZERO; seq_len * hidden];
+            for s in 0..seq_len {
+                for h in 0..num_heads {
+                    let src_off = (h * seq_len + s) * head_dim;
+                    let dst_off = s * hidden + h * head_dim;
+                    out[dst_off..dst_off + head_dim]
+                        .copy_from_slice(&src[src_off..src_off + head_dim]);
+                }
+            }
+            Tensor::from_slice(&out, &[seq_len, hidden]).to_device(device)
+        }
+        _ => panic!("unsupported dtype {:?} in merge_heads", dtype),
+    }
+}
+
+/// Greedy sampling: return the argmax token ID from the last position's logits.
+pub fn sample_greedy(logits: &Tensor) -> u32 {
+    assert_eq!(logits.ndim(), 2);
+    let logits_cpu = logits.to_device(Device::Cpu);
+    let data = logits_cpu.as_slice::<f32>();
+    let vocab_size = logits.shape()[1];
+    let seq_len = logits.shape()[0];
+    let last_row = &data[(seq_len - 1) * vocab_size..seq_len * vocab_size];
+    last_row
+        .iter()
+        .enumerate()
+        .max_by(|a, b| a.1.partial_cmp(b.1).unwrap())
+        .map(|(idx, _)| idx as u32)
+        .unwrap()
+}

crates/xserv-model/src/gpt_oss_graph.rs

@@ -0,0 +1,195 @@
+//! CUDA-graph replay for gpt-oss batch=1 decode (Phase 21).
+//!
+//! A decode step launches ~200 kernels; with sparse MoE the GPU work is only
+//! a few ms, so launch overhead dominates TPOT. The whole step (embedding →
+//! 24 layers → logits) is captured ONCE into a CUDA graph and replayed per
+//! token with a single `cudaGraphLaunch`.
+//!
+//! Why the existing forward is capturable as-is:
+//! - Every per-step variable input lives in a stable-address device buffer
+//!   whose CONTENTS are updated outside the captured region: token id and
+//!   position (persistent buffers owned here), block table and context lens
+//!   (PagedKVCache GPU buffers, refreshed by `decode_prepare`). The KV scatter
+//!   and paged attention kernels read their write/read positions from those
+//!   buffers, and the sparse-MoE GEMVs read expert ids from `topk_ids` written
+//!   earlier in the same graph — all data-dependent, no host branching.
+//! - Kernel launches go through the thread-local launch stream
+//!   (`xserv_cuda::stream::push_stream`), so the capture stream sees them.
+//! - Intermediate tensors come from the caching allocator. Blocks freed while
+//!   capturing are quarantined (`allocator::begin_retain`) for the graph's
+//!   lifetime so no later allocation can take ownership of memory the graph
+//!   still references on every replay.
+//!
+//! Capture preconditions: at least one EAGER decode step must have run first,
+//! so the allocator pool already holds every bucket size the step needs
+//! (a pool-miss inside capture would call cudaMalloc — illegal while
+//! capturing) and cuBLAS has finished its one-time per-shape setup.
+
+use std::ffi::c_void;
+
+use xserv_cuda::allocator::{self, RetainedBlocks};
+use xserv_cuda::{CudaGraph, CudaStream, GpuBuffer};
+use xserv_tensor::Tensor;
+
+use crate::gpt_oss::GptOss;
+use crate::paged_kv_cache::PagedKVCache;
+
+pub struct GptOssDecodeGraph {
+    stream: CudaStream,
+    graph: CudaGraph,
+    ids_buf: GpuBuffer, // [1] u32, persistent graph input
+    pos_buf: GpuBuffer, // [1] u32, persistent graph input
+    logits: Tensor,     // graph output; rewritten in place by every replay
+    _arena: RetainedBlocks,
+}
+
+impl GptOssDecodeGraph {
+    /// Capture one batch=1 decode step and replay it once (capture records
+    /// without executing, so the replay performs this token's computation).
+    pub fn capture(
+        model: &GptOss,
+        token: u32,
+        position: usize,
+        slot: usize,
+        cache: &mut PagedKVCache,
+    ) -> Self {
+        let stream = CudaStream::new().expect("create capture stream");
+        let mut ids_buf = allocator::cached_alloc(4).expect("alloc ids buf");
+        let mut pos_buf = allocator::cached_alloc(4).expect("alloc pos buf");
+
+        model.decode_prepare(&[position], &[slot], cache);
+        ids_buf.copy_from_host(&token.to_le_bytes()).unwrap();
+        pos_buf
+            .copy_from_host(&(position as u32).to_le_bytes())
+            .unwrap();
+
+        // Retained warmup: run the exact step once eagerly with the quarantine
+        // ON. Freed intermediates are held back instead of recycled, so the
+        // pool ends up stocked with a dedicated block for EVERY allocation the
+        // step performs. The capture below repeats the same allocation
+        // sequence and therefore never misses the pool — a pool miss would
+        // call cudaMalloc, which is illegal while a stream is capturing (this
+        // is also why one block per bucket is not enough: the capture's own
+        // quarantine keeps freed blocks out of reuse). Re-running the step is
+        // idempotent: the KV scatter rewrites the same cache position.
+        allocator::begin_retain();
+        {
+            let _guard = xserv_cuda::push_stream(&stream);
+            let _ = model.decode_core(
+                ids_buf.as_ptr() as *const c_void,
+                pos_buf.as_ptr() as *const c_void,
+                1,
+                cache,
+            );
+        }
+        drop(allocator::end_retain()); // release the warmup blocks to the pool
+        stream.synchronize().expect("warmup sync");
+
+        allocator::begin_retain();
+        let mut graph = CudaGraph::new();
+        let logits;
+        {
+            let _guard = xserv_cuda::stream::push_stream(&stream);
+            graph
+                .begin_capture(&stream)
+                .expect("begin decode-graph capture");
+            logits = model.decode_core(
+                ids_buf.as_ptr() as *const c_void,
+                pos_buf.as_ptr() as *const c_void,
+                1,
+                cache,
+            );
+            graph
+                .end_capture(&stream)
+                .expect("end decode-graph capture");
+        }
+        let arena = allocator::end_retain();
+
+        graph.launch(&stream).expect("first decode-graph replay");
+        cache.advance_seq_len(slot, 1);
+
+        Self {
+            stream,
+            graph,
+            ids_buf,
+            pos_buf,
+            logits,
+            _arena: arena,
+        }
+    }
+
+    /// Run one decode step by replaying the captured graph.
+    pub fn step(
+        &mut self,
+        model: &GptOss,
+        token: u32,
+        position: usize,
+        slot: usize,
+        cache: &mut PagedKVCache,
+    ) -> Tensor {
+        model.decode_prepare(&[position], &[slot], cache);
+        self.ids_buf.copy_from_host(&token.to_le_bytes()).unwrap();
+        self.pos_buf
+            .copy_from_host(&(position as u32).to_le_bytes())
+            .unwrap();
+        self.graph
+            .launch(&self.stream)
+            .expect("decode-graph replay");
+        cache.advance_seq_len(slot, 1);
+        // Shallow clone: the caller reads these logits before the next replay
+        // rewrites the underlying buffer.
+        self.logits.clone()
+    }
+}
+
+/// Lazy capture policy: first decode step of the process runs eager (warms the
+/// allocator pool + cuBLAS so capture performs no "unsafe" CUDA calls), the
+/// second is captured, the rest replay. Batch>1 always falls back to eager.
+/// Disable with XSERV_DECODE_GRAPH=0.
+pub struct GraphedGptOssDecoder {
+    graph: Option<GptOssDecodeGraph>,
+    eager_steps: u32,
+    enabled: bool,
+}
+
+impl GraphedGptOssDecoder {
+    pub fn new() -> Self {
+        let enabled = std::env::var("XSERV_DECODE_GRAPH")
+            .map(|v| v != "0")
+            .unwrap_or(true);
+        Self {
+            graph: None,
+            eager_steps: 0,
+            enabled,
+        }
+    }
+
+    pub fn decode(
+        &mut self,
+        model: &GptOss,
+        tokens: &[u32],
+        positions: &[usize],
+        slots: &[usize],
+        cache: &mut PagedKVCache,
+    ) -> Tensor {
+        if self.enabled && tokens.len() == 1 {
+            if let Some(g) = self.graph.as_mut() {
+                return g.step(model, tokens[0], positions[0], slots[0], cache);
+            }
+            if self.eager_steps >= 1 {
+                let g = GptOssDecodeGraph::capture(model, tokens[0], positions[0], slots[0], cache);
+                let logits = g.logits.clone();
+                self.graph = Some(g);
+                return logits;
+            }
+        }
+        self.eager_steps += 1;
+        model.forward_decode_paged(tokens, positions, slots, cache)
+    }
+}
+
+impl Default for GraphedGptOssDecoder {
+    fn default() -> Self {
+        Self::new()
+    }
+}

crates/xserv-model/src/kv_cache.rs

@@ -0,0 +1,214 @@
+use crate::config::ModelConfig;
+use xserv_cuda::GpuBuffer;
+use xserv_tensor::{DType, Tensor};
+
+/// GPU-resident KV cache. Pre-allocates max_seq_len on GPU,
+/// appends new K/V via D2D copy at offset (no CPU round-trip).
+pub struct GpuKVCache {
+    // Per layer: contiguous GPU buffer for K and V
+    // Layout: [num_kv_heads, max_seq_len, head_dim] — contiguous per head
+    k_bufs: Vec<GpuBuffer>,
+    v_bufs: Vec<GpuBuffer>,
+    // Per layer: pre-allocated staging buffers for get_kv_len output.
+    // Size: num_kv_heads * max_seq_len * head_dim * elem_size (max possible output).
+    // Avoids cudaMalloc/cudaFree on every get_kv_len call.
+    k_staging: Vec<GpuBuffer>,
+    v_staging: Vec<GpuBuffer>,
+    seq_len: usize,
+    max_seq_len: usize,
+    num_kv_heads: usize,
+    head_dim: usize,
+    elem_size: usize,
+    dtype: DType,
+    device: u32,
+}
+
+impl GpuKVCache {
+    pub fn new(config: &ModelConfig, max_seq_len: usize, dtype: DType, device: u32) -> Self {
+        let num_layers = config.num_layers();
+        let num_kv_heads = config.num_kv_heads();
+        let head_dim = config.head_dim();
+        let elem_size = dtype.size_bytes();
+        let buf_size = num_kv_heads * max_seq_len * head_dim * elem_size;
+
+        let mut k_bufs = Vec::with_capacity(num_layers);
+        let mut v_bufs = Vec::with_capacity(num_layers);
+        let mut k_staging = Vec::with_capacity(num_layers);
+        let mut v_staging = Vec::with_capacity(num_layers);
+        for _ in 0..num_layers {
+            let mut k = GpuBuffer::alloc(buf_size).expect("alloc KV cache K");
+            let mut v = GpuBuffer::alloc(buf_size).expect("alloc KV cache V");
+            k.zero().unwrap();
+            v.zero().unwrap();
+            k_bufs.push(k);
+            v_bufs.push(v);
+            k_staging.push(GpuBuffer::alloc(buf_size).expect("alloc KV staging K"));
+            v_staging.push(GpuBuffer::alloc(buf_size).expect("alloc KV staging V"));
+        }
+
+        Self {
+            k_bufs,
+            v_bufs,
+            k_staging,
+            v_staging,
+            seq_len: 0,
+            max_seq_len,
+            num_kv_heads,
+            head_dim,
+            elem_size,
+            dtype,
+            device,
+        }
+    }
+
+    pub fn seq_len(&self) -> usize {
+        self.seq_len
+    }
+    pub fn max_seq_len(&self) -> usize {
+        self.max_seq_len
+    }
+
+    /// Append new K/V tensors for a given layer.
+    /// k_new, v_new: [1, num_kv_heads, new_tokens, head_dim] on GPU, contiguous.
+    /// `write_pos` is the sequence position to write at (caller manages this).
+    pub fn append(
+        &mut self,
+        layer: usize,
+        k_new: &Tensor,
+        v_new: &Tensor,
+        new_tokens: usize,
+        write_pos: usize,
+    ) {
+        assert!(
+            write_pos + new_tokens <= self.max_seq_len,
+            "KV cache overflow"
+        );
+        let es = self.elem_size;
+        let hd = self.head_dim;
+        let max_s = self.max_seq_len;
+        let nh = self.num_kv_heads;
+
+        let k_src = k_new.storage().gpu_buffer();
+        let v_src = v_new.storage().gpu_buffer();
+
+        for h in 0..nh {
+            let src_off = h * new_tokens * hd * es;
+            let dst_off = (h * max_s + write_pos) * hd * es;
+            let count = new_tokens * hd * es;
+            self.k_bufs[layer]
+                .copy_from_device_at(k_src, src_off, dst_off, count)
+                .unwrap();
+            self.v_bufs[layer]
+                .copy_from_device_at(v_src, src_off, dst_off, count)
+                .unwrap();
+        }
+    }
+
+    pub fn advance_seq_len(&mut self, new_tokens: usize) {
+        self.seq_len += new_tokens;
+        assert!(
+            self.seq_len <= self.max_seq_len,
+            "KV cache seq_len ({}) exceeds max_seq_len ({})",
+            self.seq_len,
+            self.max_seq_len
+        );
+    }
+
+    /// Get K/V cache tensors for a layer up to `seq_len` tokens: [1, num_kv_heads, seq_len, head_dim]
+    pub fn get_kv(&mut self, layer: usize) -> (Tensor, Tensor) {
+        let sl = self.seq_len;
+        self.get_kv_len(layer, sl)
+    }
+
+    pub fn get_kv_len(&mut self, layer: usize, sl: usize) -> (Tensor, Tensor) {
+        assert!(
+            sl <= self.max_seq_len,
+            "get_kv_len: sl ({sl}) exceeds max_seq_len ({})",
+            self.max_seq_len
+        );
+        let hd = self.head_dim;
+        let nh = self.num_kv_heads;
+        let es = self.elem_size;
+        let max_s = self.max_seq_len;
+
+        // Copy each head's valid portion into pre-allocated staging buffers.
+        // Split borrows: staging (mut) vs cache (shared) are separate struct fields,
+        // so the borrow checker allows simultaneous &mut staging + &cache.
+        let out_size = nh * sl * hd * es;
+        let k_stg = &mut self.k_staging[layer];
+        let k_buf = &self.k_bufs[layer];
+        let v_stg = &mut self.v_staging[layer];
+        let v_buf = &self.v_bufs[layer];
+        for h in 0..nh {
+            let src_off = (h * max_s) * hd * es;
+            let dst_off = (h * sl) * hd * es;
+            let count = sl * hd * es;
+            k_stg
+                .copy_from_device_at(k_buf, src_off, dst_off, count)
+                .unwrap();
+            v_stg
+                .copy_from_device_at(v_buf, src_off, dst_off, count)
+                .unwrap();
+        }
+        // Grab raw pointers before dropping the mutable borrows
+        let k_ptr = k_stg.as_mut_ptr();
+        let v_ptr = v_stg.as_mut_ptr();
+
+        // Create Tensors that borrow from the staging buffers (no cudaMalloc/cudaFree).
+        // Safety: staging buffers are owned by GpuKVCache and outlive the returned Tensors
+        // in practice (Tensors are consumed within the same forward pass before the next
+        // get_kv_len call overwrites the staging buffer).
+        let shape = &[1usize, nh, sl, hd];
+        let k = unsafe {
+            tensor_from_gpu_buffer(
+                GpuBuffer::borrow_raw(k_ptr, out_size),
+                shape,
+                self.dtype,
+                self.device,
+            )
+        };
+        let v = unsafe {
+            tensor_from_gpu_buffer(
+                GpuBuffer::borrow_raw(v_ptr, out_size),
+                shape,
+                self.dtype,
+                self.device,
+            )
+        };
+        (k, v)
+    }
+}
+
+/// Create a Tensor from a GpuBuffer (takes ownership).
+unsafe fn tensor_from_gpu_buffer(
+    buf: GpuBuffer,
+    shape: &[usize],
+    dtype: DType,
+    device: u32,
+) -> Tensor {
+    use smallvec::SmallVec;
+    use xserv_tensor::shape::contiguous_strides;
+    use xserv_tensor::storage::Storage;
+
+    let storage = Storage::cuda(buf, device);
+    Tensor::from_storage(
+        storage,
+        SmallVec::from_slice(shape),
+        contiguous_strides(shape),
+        0,
+        dtype,
+    )
+}
+
+/// Public version for use by other modules (e.g., batched decode concat).
+///
+/// # Safety
+/// `buf` must be a valid GPU allocation with at least `product(shape) * dtype.size_bytes()` bytes.
+pub unsafe fn tensor_from_gpu_buffer_pub(
+    buf: GpuBuffer,
+    shape: &[usize],
+    dtype: DType,
+    device: u32,
+) -> Tensor {
+    tensor_from_gpu_buffer(buf, shape, dtype, device)
+}

crates/xserv-model/src/lib.rs

@@ -0,0 +1,28 @@
+pub mod config;
+pub mod decode_graph;
+pub mod eagle3;
+pub mod gpt2;
+pub mod gpt_oss;
+pub mod gpt_oss_graph;
+pub mod kv_cache;
+pub mod loader;
+pub mod paged_kv_cache;
+pub mod qwen3;
+pub mod qwen3_graph;
+pub mod sampling;
+
+pub use config::ModelConfig;
+pub use decode_graph::{DecodeGraphState, LayerWeightPtrs};
+pub use gpt_oss::GptOss;
+pub use gpt_oss_graph::{GptOssDecodeGraph, GraphedGptOssDecoder};
+pub use gpt2::{GPT2, KVCache};
+pub use kv_cache::GpuKVCache;
+pub use paged_kv_cache::{BLOCK_SIZE, BlockAllocator, Location, PagedKVCache};
+pub use qwen3::Qwen3;
+pub use sampling::{SamplingParams, sample, sample_greedy_penalized};
+
+/// Initialize GPU kernel hooks. Called automatically by model constructors,
+/// but safe to call multiple times (idempotent via OnceLock).
+pub fn init_kernels() {
+    xserv_kernels::init();
+}

crates/xserv-model/src/loader.rs

@@ -0,0 +1,93 @@
+use half::{bf16, f16};
+use safetensors::SafeTensors;
+use std::collections::HashMap;
+use std::path::Path;
+use xserv_tensor::{DType, Device, Tensor};
+
+pub fn load_safetensors(path: &Path, device: Device) -> HashMap<String, Tensor> {
+    let data =
+        std::fs::read(path).unwrap_or_else(|e| panic!("failed to read {}: {e}", path.display()));
+    let st = SafeTensors::deserialize(&data)
+        .unwrap_or_else(|e| panic!("failed to parse safetensors {}: {e}", path.display()));
+
+    let mut tensors = HashMap::new();
+
+    for (name, view) in st.tensors() {
+        let shape: Vec<usize> = view.shape().to_vec();
+        let raw_bytes = view.data();
+        let dtype = match view.dtype() {
+            safetensors::Dtype::F32 => DType::F32,
+            safetensors::Dtype::F16 => DType::F16,
+            safetensors::Dtype::BF16 => DType::BF16,
+            safetensors::Dtype::F8_E4M3 => DType::FP8E4M3,
+            other => {
+                eprintln!("skipping tensor {name}: unsupported dtype {other:?}");
+                continue;
+            }
+        };
+
+        let tensor = make_tensor(raw_bytes, &shape, dtype);
+        let tensor = tensor.to_device(device);
+        tensors.insert(name.to_string(), tensor);
+    }
+
+    tensors
+}
+
+/// Load from a directory containing model.safetensors (or sharded files) + config.json.
+pub fn load_model_dir(dir: &Path, device: Device) -> HashMap<String, Tensor> {
+    let single = dir.join("model.safetensors");
+    if single.exists() {
+        return load_safetensors(&single, device);
+    }
+
+    // Try sharded: model-00001-of-NNNNN.safetensors
+    let mut all_tensors = HashMap::new();
+    let mut entries: Vec<_> = std::fs::read_dir(dir)
+        .unwrap()
+        .filter_map(|e| e.ok())
+        .filter(|e| {
+            e.path()
+                .file_name()
+                .map(|f| f.to_string_lossy().ends_with(".safetensors"))
+                .unwrap_or(false)
+        })
+        .collect();
+    entries.sort_by_key(|e| e.file_name());
+
+    for entry in entries {
+        let tensors = load_safetensors(&entry.path(), device);
+        all_tensors.extend(tensors);
+    }
+
+    assert!(
+        !all_tensors.is_empty(),
+        "no safetensors files found in {}",
+        dir.display()
+    );
+    all_tensors
+}
+
+pub(crate) fn make_tensor(raw_bytes: &[u8], shape: &[usize], dtype: DType) -> Tensor {
+    match dtype {
+        DType::F32 => {
+            let floats: &[f32] = unsafe {
+                std::slice::from_raw_parts(raw_bytes.as_ptr() as *const f32, raw_bytes.len() / 4)
+            };
+            Tensor::from_slice(floats, shape)
+        }
+        DType::F16 => {
+            let halfs: &[f16] = unsafe {
+                std::slice::from_raw_parts(raw_bytes.as_ptr() as *const f16, raw_bytes.len() / 2)
+            };
+            Tensor::from_slice(halfs, shape)
+        }
+        DType::BF16 => {
+            let bfs: &[bf16] = unsafe {
+                std::slice::from_raw_parts(raw_bytes.as_ptr() as *const bf16, raw_bytes.len() / 2)
+            };
+            Tensor::from_slice(bfs, shape)
+        }
+        DType::FP8E4M3 => Tensor::from_raw_bytes(raw_bytes, shape, DType::FP8E4M3),
+    }
+}

crates/xserv-model/src/paged_kv_cache.rs

@@ -0,0 +1,908 @@
+//! Paged KV cache: vLLM-style block-based KV cache with O(1) allocation
+//! and indirection via per-sequence block tables.
+//!
+//! Physical layout per layer:
+//!   K pool: [total_blocks, num_kv_heads, BLOCK_SIZE, head_dim] BF16
+//!   V pool: same
+//!
+//! Logical view per sequence: a list of physical block ids. Token at logical
+//! position p lives in block_ids[p / BLOCK_SIZE] at slot (p % BLOCK_SIZE).
+
+use crate::config::ModelConfig;
+use xserv_cuda::{GpuBuffer, PinnedBuffer};
+use xserv_tensor::{DType, Tensor};
+
+pub const BLOCK_SIZE: usize = 16;
+
+/// Stack-based block allocator: O(1) alloc/free.
+pub struct BlockAllocator {
+    free_stack: Vec<u32>,
+    total: usize,
+}
+
+impl BlockAllocator {
+    pub fn new(total_blocks: usize) -> Self {
+        // Reserve block 0 as a sentinel "null" block (never allocated).
+        // Free list contains [total-1, total-2, ..., 1] so pop returns 1 first.
+        // total_blocks==0 means "disabled" (e.g. swap off): empty free list.
+        let mut free_stack = Vec::with_capacity(total_blocks.saturating_sub(1));
+        for b in (1..total_blocks).rev() {
+            free_stack.push(b as u32);
+        }
+        Self {
+            free_stack,
+            total: total_blocks,
+        }
+    }
+
+    pub fn alloc(&mut self) -> Option<u32> {
+        self.free_stack.pop()
+    }
+
+    pub fn free(&mut self, block: u32) {
+        debug_assert!((block as usize) < self.total && block != 0);
+        self.free_stack.push(block);
+    }
+
+    pub fn free_count(&self) -> usize {
+        self.free_stack.len()
+    }
+
+    pub fn total(&self) -> usize {
+        self.total
+    }
+
+    pub fn can_alloc(&self, n: usize) -> bool {
+        self.free_stack.len() >= n
+    }
+}
+
+/// Where a sequence's KV blocks currently live.
+#[derive(Clone, Copy, PartialEq, Eq, Debug)]
+pub enum Location {
+    Gpu,
+    Cpu,
+}
+
+/// Per-sequence state held in the cache.
+#[derive(Clone)]
+pub struct SeqState {
+    /// Block ids into the GPU pool when `location == Gpu`, or into the CPU
+    /// (pinned host) pool when `location == Cpu`.
+    pub block_ids: Vec<u32>,
+    pub seq_len: usize,
+    pub location: Location,
+}
+
+pub struct PagedKVCache {
+    // [layer]: GpuBuffer of size total_blocks * nkv * BLOCK_SIZE * hd * elem_size
+    k_pools: Vec<GpuBuffer>,
+    v_pools: Vec<GpuBuffer>,
+
+    // CPU (pinned host) swap pools, same per-layer layout as the GPU pools but
+    // sized for `cpu_total_blocks`. Empty when swap is disabled.
+    cpu_k_pools: Vec<PinnedBuffer>,
+    cpu_v_pools: Vec<PinnedBuffer>,
+    cpu_allocator: BlockAllocator,
+
+    // Bytes occupied by one block within a single layer pool:
+    //   num_kv_heads * BLOCK_SIZE * head_dim * elem_size.
+    block_bytes: usize,
+
+    allocator: BlockAllocator,
+    seq_states: Vec<Option<SeqState>>,
+
+    // GPU-resident per-sequence metadata. Uploaded each step via sync_to_gpu().
+    // block_table_gpu: i32 [max_seqs, max_blocks_per_seq]
+    // context_lens_gpu: i32 [max_seqs]
+    block_table_gpu: GpuBuffer,
+    context_lens_gpu: GpuBuffer,
+    // Host-side staging mirroring the GPU buffers above.
+    block_table_host: Vec<i32>,
+    context_lens_host: Vec<i32>,
+
+    // Config
+    num_layers: usize,
+    num_kv_heads: usize,
+    head_dim: usize,
+    elem_size: usize,
+    dtype: DType,
+    device: u32,
+    max_seqs: usize,
+    max_blocks_per_seq: usize,
+}
+
+impl PagedKVCache {
+    /// Bytes occupied by all KV blocks for ONE physical block across the whole
+    /// model (both K and V, all layers). Use this to size pools against VRAM.
+    pub fn bytes_per_block(config: &ModelConfig, dtype: DType) -> usize {
+        2 * config.num_layers()
+            * config.num_kv_heads()
+            * BLOCK_SIZE
+            * config.head_dim()
+            * dtype.size_bytes()
+    }
+
+    /// Create a new paged cache.
+    /// - `total_blocks`: total number of physical GPU blocks across all sequences.
+    /// - `cpu_total_blocks`: physical blocks in the pinned-host swap pool (0 = swap off).
+    /// - `max_seqs`: max number of concurrent sequences (slots), incl. swapped.
+    /// - `max_blocks_per_seq`: capacity of the block table per slot
+    ///   (must be >= ceil(max_seq_len / BLOCK_SIZE)).
+    pub fn new(
+        config: &ModelConfig,
+        total_blocks: usize,
+        cpu_total_blocks: usize,
+        max_seqs: usize,
+        max_blocks_per_seq: usize,
+        dtype: DType,
+        device: u32,
+    ) -> Self {
+        Self::new_tp(
+            config,
+            config.num_kv_heads(),
+            total_blocks,
+            cpu_total_blocks,
+            max_seqs,
+            max_blocks_per_seq,
+            dtype,
+            device,
+        )
+    }
+
+    /// Like `new`, but with an explicit `num_kv_heads` — under tensor parallelism
+    /// each rank only stores its `num_kv_heads / world` heads, so the pool is
+    /// sized for the local head count, not the model's full count.
+    #[allow(clippy::too_many_arguments)]
+    pub fn new_tp(
+        config: &ModelConfig,
+        num_kv_heads: usize,
+        total_blocks: usize,
+        cpu_total_blocks: usize,
+        max_seqs: usize,
+        max_blocks_per_seq: usize,
+        dtype: DType,
+        device: u32,
+    ) -> Self {
+        assert!(
+            total_blocks >= 2,
+            "need at least 2 blocks (one is sentinel)"
+        );
+        let num_layers = config.num_layers();
+        let head_dim = config.head_dim();
+        let elem_size = dtype.size_bytes();
+        let block_bytes = num_kv_heads * BLOCK_SIZE * head_dim * elem_size;
+        let pool_bytes = total_blocks * block_bytes;
+
+        let mut k_pools = Vec::with_capacity(num_layers);
+        let mut v_pools = Vec::with_capacity(num_layers);
+        for _ in 0..num_layers {
+            let mut k = GpuBuffer::alloc(pool_bytes).expect("alloc paged K pool");
+            let mut v = GpuBuffer::alloc(pool_bytes).expect("alloc paged V pool");
+            k.zero().unwrap();
+            v.zero().unwrap();
+            k_pools.push(k);
+            v_pools.push(v);
+        }
+
+        // Pinned-host swap pools (one per layer, mirroring the GPU layout).
+        let mut cpu_k_pools = Vec::new();
+        let mut cpu_v_pools = Vec::new();
+        if cpu_total_blocks >= 2 {
+            let cpu_pool_bytes = cpu_total_blocks * block_bytes;
+            for _ in 0..num_layers {
+                cpu_k_pools
+                    .push(PinnedBuffer::alloc(cpu_pool_bytes).expect("alloc CPU K swap pool"));
+                cpu_v_pools
+                    .push(PinnedBuffer::alloc(cpu_pool_bytes).expect("alloc CPU V swap pool"));
+            }
+        }
+        let cpu_allocator = BlockAllocator::new(if cpu_total_blocks >= 2 {
+            cpu_total_blocks
+        } else {
+            0
+        });
+
+        let block_table_gpu =
+            GpuBuffer::alloc(max_seqs * max_blocks_per_seq * std::mem::size_of::<i32>())
+                .expect("alloc block table");
+        let context_lens_gpu =
+            GpuBuffer::alloc(max_seqs * std::mem::size_of::<i32>()).expect("alloc context lens");
+
+        let block_table_host = vec![0i32; max_seqs * max_blocks_per_seq];
+        let context_lens_host = vec![0i32; max_seqs];
+
+        let seq_states = (0..max_seqs).map(|_| None).collect();
+
+        Self {
+            k_pools,
+            v_pools,
+            cpu_k_pools,
+            cpu_v_pools,
+            cpu_allocator,
+            block_bytes,
+            allocator: BlockAllocator::new(total_blocks),
+            seq_states,
+            block_table_gpu,
+            context_lens_gpu,
+            block_table_host,
+            context_lens_host,
+            num_layers,
+            num_kv_heads,
+            head_dim,
+            elem_size,
+            dtype,
+            device,
+            max_seqs,
+            max_blocks_per_seq,
+        }
+    }
+
+    pub fn num_layers(&self) -> usize {
+        self.num_layers
+    }
+    pub fn num_kv_heads(&self) -> usize {
+        self.num_kv_heads
+    }
+    pub fn head_dim(&self) -> usize {
+        self.head_dim
+    }
+    pub fn dtype(&self) -> DType {
+        self.dtype
+    }
+    pub fn max_seqs(&self) -> usize {
+        self.max_seqs
+    }
+    pub fn max_blocks_per_seq(&self) -> usize {
+        self.max_blocks_per_seq
+    }
+    pub fn free_blocks(&self) -> usize {
+        self.allocator.free_count()
+    }
+    pub fn total_blocks(&self) -> usize {
+        self.allocator.total()
+    }
+
+    pub fn k_pool(&self, layer: usize) -> &GpuBuffer {
+        &self.k_pools[layer]
+    }
+    pub fn v_pool(&self, layer: usize) -> &GpuBuffer {
+        &self.v_pools[layer]
+    }
+    pub fn block_table_gpu(&self) -> &GpuBuffer {
+        &self.block_table_gpu
+    }
+    pub fn context_lens_gpu(&self) -> &GpuBuffer {
+        &self.context_lens_gpu
+    }
+
+    pub fn seq_len(&self, slot: usize) -> usize {
+        self.seq_states[slot]
+            .as_ref()
+            .map(|s| s.seq_len)
+            .unwrap_or(0)
+    }
+
+    pub fn is_slot_free(&self, slot: usize) -> bool {
+        self.seq_states[slot].is_none()
+    }
+
+    /// Register a new sequence at `slot`. Allocates the first block.
+    /// Returns Err(()) if no slot or no blocks are available.
+    pub fn register_sequence(&mut self, slot: usize) -> Result<(), &'static str> {
+        if slot >= self.max_seqs {
+            return Err("slot out of range");
+        }
+        if self.seq_states[slot].is_some() {
+            return Err("slot already in use");
+        }
+        let block = self.allocator.alloc().ok_or("out of blocks")?;
+        self.seq_states[slot] = Some(SeqState {
+            block_ids: vec![block],
+            seq_len: 0,
+            location: Location::Gpu,
+        });
+        Ok(())
+    }
+
+    /// Free all blocks for `slot` and clear the slot. Frees from whichever pool
+    /// (GPU or CPU) the sequence currently lives in.
+    pub fn free_sequence(&mut self, slot: usize) {
+        if let Some(state) = self.seq_states[slot].take() {
+            let alloc = match state.location {
+                Location::Gpu => &mut self.allocator,
+                Location::Cpu => &mut self.cpu_allocator,
+            };
+            for b in state.block_ids {
+                alloc.free(b);
+            }
+        }
+    }
+
+    /// Number of blocks needed to hold `seq_len + new_tokens` tokens, beyond
+    /// what is currently allocated for `slot`.
+    pub fn additional_blocks_needed(&self, slot: usize, new_tokens: usize) -> usize {
+        let state = self.seq_states[slot].as_ref().expect("unregistered slot");
+        let cur = state.block_ids.len();
+        let needed_total = (state.seq_len + new_tokens + BLOCK_SIZE - 1) / BLOCK_SIZE;
+        if needed_total > cur {
+            needed_total - cur
+        } else {
+            0
+        }
+    }
+
+    /// Pre-allocate enough physical blocks in `slot` to cover positions
+    /// `[0, end_pos)`. Call once before the per-layer append loop so that
+    /// every layer's append uses the same block table.
+    pub fn ensure_capacity(&mut self, slot: usize, end_pos: usize) {
+        let state = self.seq_states[slot].as_mut().expect("unregistered slot");
+        let needed_total = (end_pos + BLOCK_SIZE - 1) / BLOCK_SIZE;
+        while state.block_ids.len() < needed_total {
+            let b = self
+                .allocator
+                .alloc()
+                .expect("out of blocks (caller must check)");
+            assert!(
+                state.block_ids.len() < self.max_blocks_per_seq,
+                "block table overflow"
+            );
+            state.block_ids.push(b);
+        }
+    }
+
+    /// Append `num_tokens` of K/V into the paged pool for `slot` at logical
+    /// position `start_pos`. Caller must have called `ensure_capacity(slot, start_pos + num_tokens)`
+    /// first (or accept that this method may also extend block list).
+    /// Does NOT touch `seq_len`. Call `advance_seq_len(slot, num_tokens)` after
+    /// every layer has been written.
+    ///
+    /// `k_new`, `v_new`: GPU tensors with logical shape
+    ///   [1, num_kv_heads, num_tokens, head_dim]
+    /// stored contiguously (head-major, then tokens, then dim).
+    ///
+    /// Implementation: a single `reshape_and_cache` kernel per call. The
+    /// previous Rust loop fired `num_tokens * num_kv_heads` cudaMemcpys per
+    /// layer (≈290k for a 1024-token Qwen3 prefill across 36 layers).
+    pub fn append_tokens(
+        &mut self,
+        slot: usize,
+        layer: usize,
+        k_new: &Tensor,
+        v_new: &Tensor,
+        num_tokens: usize,
+        start_pos: usize,
+    ) {
+        if num_tokens == 0 {
+            return;
+        }
+        // Make sure blocks exist for the target range.
+        self.ensure_capacity(slot, start_pos + num_tokens);
+
+        let nkv = self.num_kv_heads;
+        let hd = self.head_dim;
+        let bs = BLOCK_SIZE;
+
+        // Stage block_ids on the GPU. Pool-allocated so this is essentially
+        // free after the first call (same bucket every step).
+        let block_ids: Vec<i32> = self.seq_states[slot]
+            .as_ref()
+            .unwrap()
+            .block_ids
+            .iter()
+            .map(|&b| b as i32)
+            .collect();
+        let bytes = block_ids.len() * std::mem::size_of::<i32>();
+        let mut block_ids_gpu =
+            xserv_cuda::allocator::cached_alloc(bytes).expect("alloc append block_ids");
+        let block_ids_bytes =
+            unsafe { std::slice::from_raw_parts(block_ids.as_ptr() as *const u8, bytes) };
+        block_ids_gpu
+            .copy_from_host(block_ids_bytes)
+            .expect("upload block_ids");
+
+        let k_src = k_new.data_ptr() as *const std::ffi::c_void;
+        let v_src = v_new.data_ptr() as *const std::ffi::c_void;
+        let k_pool_ptr = self.k_pools[layer].as_mut_ptr() as *mut std::ffi::c_void;
+        let v_pool_ptr = self.v_pools[layer].as_mut_ptr() as *mut std::ffi::c_void;
+
+        unsafe {
+            xserv_kernels::reshape_and_cache_bf16(
+                k_src,
+                v_src,
+                k_pool_ptr,
+                v_pool_ptr,
+                block_ids_gpu.as_ptr() as *const i32,
+                num_tokens,
+                nkv,
+                hd,
+                start_pos,
+                bs,
+                xserv_cuda::current_stream_raw(),
+            );
+        }
+        // block_ids_gpu drops here; the launch on the null stream will have
+        // finished consuming it before any subsequent op alloc()s the same
+        // bucket (null stream is sequential).
+    }
+
+    /// Batched append for the multi-sequence decode step: writes one new
+    /// K/V token per active sequence into `layer`'s pool, using
+    /// `block_table_gpu` and `context_lens_gpu` directly. Caller must have
+    /// just run `sync_active_batch_with_lens(slots, kv_lens)` so that:
+    ///   - row `i` of block_table_gpu holds the block ids for `slots[i]`
+    ///   - context_lens_gpu[i] == seq_len(slots[i]) + 1 (the kv_len **after**
+    ///     this step — i.e., the new token will be written at index kv_len-1)
+    ///
+    /// `k_new`, `v_new`: GPU tensors, contiguous, BF16, shape
+    ///   `[batch, num_kv_heads, head_dim]`.
+    ///
+    /// Like `append_tokens`, this does **not** touch `seq_len`. Call
+    /// `advance_seq_len(slot, 1)` for each slot after every layer has been
+    /// written.
+    pub fn append_tokens_batched(
+        &mut self,
+        layer: usize,
+        k_new: &Tensor,
+        v_new: &Tensor,
+        batch: usize,
+    ) {
+        if batch == 0 {
+            return;
+        }
+        let nkv = self.num_kv_heads;
+        let hd = self.head_dim;
+        debug_assert_eq!(k_new.shape(), &[batch, nkv, hd]);
+        debug_assert_eq!(v_new.shape(), &[batch, nkv, hd]);
+
+        let k_src = k_new.data_ptr() as *const std::ffi::c_void;
+        let v_src = v_new.data_ptr() as *const std::ffi::c_void;
+        let k_pool_ptr = self.k_pools[layer].as_mut_ptr() as *mut std::ffi::c_void;
+        let v_pool_ptr = self.v_pools[layer].as_mut_ptr() as *mut std::ffi::c_void;
+        let bt_ptr = self.block_table_gpu.as_ptr() as *const i32;
+        let cl_ptr = self.context_lens_gpu.as_ptr() as *const i32;
+
+        unsafe {
+            xserv_kernels::reshape_and_cache_batched_bf16(
+                k_src,
+                v_src,
+                k_pool_ptr,
+                v_pool_ptr,
+                bt_ptr,
+                cl_ptr,
+                batch,
+                nkv,
+                hd,
+                BLOCK_SIZE,
+                self.max_blocks_per_seq,
+                xserv_cuda::current_stream_raw(),
+            );
+        }
+    }
+
+    /// Advance the logical seq_len after append_tokens for ALL layers has completed.
+    pub fn advance_seq_len(&mut self, slot: usize, num_tokens: usize) {
+        let state = self.seq_states[slot].as_mut().expect("unregistered slot");
+        state.seq_len += num_tokens;
+    }
+
+    /// Roll a registered sequence back to `new_len` tokens.
+    ///
+    /// This only changes cache metadata and frees whole physical blocks that are
+    /// no longer reachable. Bytes inside retained blocks are left untouched; the
+    /// logical `seq_len` prevents attention from reading them, and later writes
+    /// to the same positions overwrite them.
+    pub fn truncate_sequence(&mut self, slot: usize, new_len: usize) -> Result<(), &'static str> {
+        if slot >= self.max_seqs {
+            return Err("truncate_sequence: slot out of range");
+        }
+        let state = self.seq_states[slot]
+            .as_mut()
+            .ok_or("truncate_sequence: empty slot")?;
+        if new_len > state.seq_len {
+            return Err("truncate_sequence: cannot extend");
+        }
+
+        let needed_blocks = ((new_len + BLOCK_SIZE - 1) / BLOCK_SIZE).max(1);
+        while state.block_ids.len() > needed_blocks {
+            let block = state.block_ids.pop().expect("checked len");
+            match state.location {
+                Location::Gpu => self.allocator.free(block),
+                Location::Cpu => self.cpu_allocator.free(block),
+            }
+        }
+        state.seq_len = new_len;
+        Ok(())
+    }
+
+    /// Copy K/V data from `src_pos` to `dst_pos` within the same slot, across
+    /// all layers. Used by tree speculative decoding to remap an accepted
+    /// sibling's K/V to the canonical sequential position after acceptance.
+    ///
+    /// Requires: both positions within the currently-allocated block range.
+    pub fn copy_kv_position(&self, slot: usize, src_pos: usize, dst_pos: usize) {
+        let state = self.seq_states[slot]
+            .as_ref()
+            .expect("copy_kv_position: slot not registered");
+        assert!(
+            src_pos < state.seq_len && dst_pos < state.seq_len,
+            "copy_kv_position: positions must be within seq_len"
+        );
+        // Upload this sequence's block_ids to a small GPU buffer.
+        let block_ids_host: Vec<i32> = state.block_ids.iter().map(|&b| b as i32).collect();
+        let bytes: &[u8] = unsafe {
+            std::slice::from_raw_parts(
+                block_ids_host.as_ptr() as *const u8,
+                block_ids_host.len() * 4,
+            )
+        };
+        let mut ids_buf =
+            xserv_cuda::allocator::cached_alloc(bytes.len()).expect("alloc block_ids for copy");
+        ids_buf.copy_from_host(bytes).unwrap();
+        let ids_ptr = ids_buf.as_ptr() as *const i32;
+
+        let stream = xserv_cuda::current_stream_raw();
+        let num_layers = self.k_pools.len();
+        for layer in 0..num_layers {
+            unsafe {
+                xserv_kernels::copy_kv_position(
+                    self.k_pools[layer].as_ptr() as *mut std::ffi::c_void,
+                    self.v_pools[layer].as_ptr() as *mut std::ffi::c_void,
+                    ids_ptr,
+                    src_pos,
+                    dst_pos,
+                    self.num_kv_heads,
+                    self.head_dim,
+                    BLOCK_SIZE,
+                    stream,
+                );
+            }
+        }
+    }
+
+    /// Refresh the host-side block table + context lens from `seq_states`,
+    /// then upload to GPU. Call once per decode step before the paged kernel.
+    pub fn sync_to_gpu(&mut self) {
+        let stride = self.max_blocks_per_seq;
+        for slot in 0..self.max_seqs {
+            let row = &mut self.block_table_host[slot * stride..(slot + 1) * stride];
+            row.fill(0);
+            let len = match &self.seq_states[slot] {
+                Some(s) => {
+                    for (i, b) in s.block_ids.iter().enumerate() {
+                        row[i] = *b as i32;
+                    }
+                    s.seq_len as i32
+                }
+                None => 0,
+            };
+            self.context_lens_host[slot] = len;
+        }
+
+        self.upload_metadata();
+    }
+
+    /// Pack the given active slots into rows 0..slots.len() of block_table_gpu
+    /// and context_lens_gpu, then upload. Used by paged decode where the kernel
+    /// iterates over `batch` active sequences in order.
+    pub fn sync_active_batch_to_gpu(&mut self, slots: &[usize]) {
+        let lens: Vec<i32> = slots
+            .iter()
+            .map(|&s| self.seq_states[s].as_ref().unwrap().seq_len as i32)
+            .collect();
+        self.sync_active_batch_with_lens(slots, &lens);
+    }
+
+    /// Like sync_active_batch_to_gpu but uses caller-supplied kv_lens (number
+    /// of valid K/V tokens to attend over per active row). Useful when the
+    /// kv_len for the current step differs from the cached seq_len (e.g.
+    /// before advance_seq_len has run).
+    pub fn sync_active_batch_with_lens(&mut self, slots: &[usize], kv_lens: &[i32]) {
+        assert_eq!(slots.len(), kv_lens.len());
+        assert!(
+            slots.len() <= self.max_seqs,
+            "active batch exceeds max_seqs"
+        );
+        let stride = self.max_blocks_per_seq;
+        for row in &mut self.block_table_host {
+            *row = 0;
+        }
+        for cl in &mut self.context_lens_host {
+            *cl = 0;
+        }
+        for (i, &slot) in slots.iter().enumerate() {
+            let s = self.seq_states[slot]
+                .as_ref()
+                .expect("unregistered slot in active batch");
+            let row = &mut self.block_table_host[i * stride..(i + 1) * stride];
+            for (j, b) in s.block_ids.iter().enumerate() {
+                row[j] = *b as i32;
+            }
+            self.context_lens_host[i] = kv_lens[i];
+        }
+        self.upload_metadata();
+    }
+
+    fn upload_metadata(&mut self) {
+        let bt_bytes = unsafe {
+            std::slice::from_raw_parts(
+                self.block_table_host.as_ptr() as *const u8,
+                self.block_table_host.len() * std::mem::size_of::<i32>(),
+            )
+        };
+        self.block_table_gpu.copy_from_host(bt_bytes).unwrap();
+
+        let cl_bytes = unsafe {
+            std::slice::from_raw_parts(
+                self.context_lens_host.as_ptr() as *const u8,
+                self.context_lens_host.len() * std::mem::size_of::<i32>(),
+            )
+        };
+        self.context_lens_gpu.copy_from_host(cl_bytes).unwrap();
+    }
+
+    /// Materialize a contiguous K/V tensor for a sequence at `layer`, shaped
+    /// [1, num_kv_heads, seq_len, head_dim]. Used for prefill, where Flash
+    /// Attention 2 expects contiguous K/V.
+    ///
+    /// Allocates from the cached allocator; the returned Tensors own their storage.
+    pub fn gather_kv_contiguous(&self, slot: usize, layer: usize) -> (Tensor, Tensor) {
+        let state = self.seq_states[slot].as_ref().expect("unregistered slot");
+        let sl = state.seq_len;
+        let nkv = self.num_kv_heads;
+        let hd = self.head_dim;
+        let es = self.elem_size;
+        let bs = BLOCK_SIZE;
+
+        let out_bytes = nkv * sl * hd * es;
+        let mut k_dst = xserv_cuda::allocator::cached_alloc(out_bytes).expect("alloc gather K");
+        let mut v_dst = xserv_cuda::allocator::cached_alloc(out_bytes).expect("alloc gather V");
+
+        let k_pool = &self.k_pools[layer];
+        let v_pool = &self.v_pools[layer];
+
+        let mut p = 0usize;
+        while p < sl {
+            let logical_blk = p / bs;
+            let slot_in_blk = p % bs;
+            let chunk = (bs - slot_in_blk).min(sl - p);
+            let phys = state.block_ids[logical_blk] as usize;
+
+            for h in 0..nkv {
+                let src_off = ((phys * nkv + h) * bs + slot_in_blk) * hd * es;
+                let dst_off = (h * sl + p) * hd * es;
+                let count = chunk * hd * es;
+                k_dst
+                    .copy_from_device_at(k_pool, src_off, dst_off, count)
+                    .unwrap();
+                v_dst
+                    .copy_from_device_at(v_pool, src_off, dst_off, count)
+                    .unwrap();
+            }
+            p += chunk;
+        }
+
+        let shape = &[1usize, nkv, sl, hd];
+        let k = unsafe { tensor_from_owned_buf(k_dst, shape, self.dtype, self.device) };
+        let v = unsafe { tensor_from_owned_buf(v_dst, shape, self.dtype, self.device) };
+        (k, v)
+    }
+
+    // ----- Swapping (vLLM-style preemption to pinned host memory) -----
+
+    pub fn free_cpu_blocks(&self) -> usize {
+        self.cpu_allocator.free_count()
+    }
+    pub fn swap_enabled(&self) -> bool {
+        !self.cpu_k_pools.is_empty()
+    }
+
+    pub fn is_swapped(&self, slot: usize) -> bool {
+        matches!(
+            self.seq_states[slot].as_ref().map(|s| s.location),
+            Some(Location::Cpu)
+        )
+    }
+
+    /// Number of physical blocks currently held by `slot` (in either pool).
+    pub fn block_count(&self, slot: usize) -> usize {
+        self.seq_states[slot]
+            .as_ref()
+            .map(|s| s.block_ids.len())
+            .unwrap_or(0)
+    }
+
+    /// Whether a swapped sequence at `slot` can be brought back (enough free GPU blocks).
+    pub fn can_swap_in(&self, slot: usize) -> bool {
+        self.allocator.can_alloc(self.block_count(slot))
+    }
+
+    /// Whether the GPU sequence at `slot` can be evicted (enough free CPU blocks).
+    pub fn can_swap_out(&self, slot: usize) -> bool {
+        self.cpu_allocator.can_alloc(self.block_count(slot))
+    }
+
+    /// Evict `slot`'s KV from GPU to pinned host memory and free its GPU blocks.
+    /// The slot stays registered (location = Cpu); the sequence is paused.
+    pub fn swap_out(&mut self, slot: usize) -> Result<(), &'static str> {
+        let state = self.seq_states[slot]
+            .as_ref()
+            .ok_or("swap_out: empty slot")?;
+        if state.location == Location::Cpu {
+            return Ok(());
+        }
+        let gpu_ids = state.block_ids.clone();
+        let n = gpu_ids.len();
+        if !self.cpu_allocator.can_alloc(n) {
+            return Err("swap_out: CPU pool full");
+        }
+
+        let cpu_ids: Vec<u32> = (0..n)
+            .map(|_| self.cpu_allocator.alloc().expect("checked can_alloc"))
+            .collect();
+
+        let bb = self.block_bytes;
+        for layer in 0..self.num_layers {
+            for i in 0..n {
+                let g_off = gpu_ids[i] as usize * bb;
+                let c_off = cpu_ids[i] as usize * bb;
+                self.k_pools[layer]
+                    .copy_to_host_at(
+                        &mut self.cpu_k_pools[layer].as_mut_slice()[c_off..c_off + bb],
+                        g_off,
+                        bb,
+                    )
+                    .unwrap();
+                self.v_pools[layer]
+                    .copy_to_host_at(
+                        &mut self.cpu_v_pools[layer].as_mut_slice()[c_off..c_off + bb],
+                        g_off,
+                        bb,
+                    )
+                    .unwrap();
+            }
+        }
+
+        for b in gpu_ids {
+            self.allocator.free(b);
+        }
+        let state = self.seq_states[slot].as_mut().unwrap();
+        state.block_ids = cpu_ids;
+        state.location = Location::Cpu;
+        Ok(())
+    }
+
+    /// Bring `slot`'s KV back from host to GPU and free its CPU blocks.
+    pub fn swap_in(&mut self, slot: usize) -> Result<(), &'static str> {
+        let state = self.seq_states[slot]
+            .as_ref()
+            .ok_or("swap_in: empty slot")?;
+        if state.location == Location::Gpu {
+            return Ok(());
+        }
+        let cpu_ids = state.block_ids.clone();
+        let n = cpu_ids.len();
+        if !self.allocator.can_alloc(n) {
+            return Err("swap_in: GPU pool full");
+        }
+
+        let gpu_ids: Vec<u32> = (0..n)
+            .map(|_| self.allocator.alloc().expect("checked can_alloc"))
+            .collect();
+
+        let bb = self.block_bytes;
+        for layer in 0..self.num_layers {
+            for i in 0..n {
+                let g_off = gpu_ids[i] as usize * bb;
+                let c_off = cpu_ids[i] as usize * bb;
+                self.k_pools[layer]
+                    .copy_from_host_at(
+                        &self.cpu_k_pools[layer].as_slice()[c_off..c_off + bb],
+                        g_off,
+                        bb,
+                    )
+                    .unwrap();
+                self.v_pools[layer]
+                    .copy_from_host_at(
+                        &self.cpu_v_pools[layer].as_slice()[c_off..c_off + bb],
+                        g_off,
+                        bb,
+                    )
+                    .unwrap();
+            }
+        }
+
+        for b in cpu_ids {
+            self.cpu_allocator.free(b);
+        }
+        let state = self.seq_states[slot].as_mut().unwrap();
+        state.block_ids = gpu_ids;
+        state.location = Location::Gpu;
+        Ok(())
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn tiny_config() -> ModelConfig {
+        serde_json::from_value(serde_json::json!({
+            "model_type": "qwen3",
+            "hidden_size": 8,
+            "intermediate_size": 16,
+            "num_attention_heads": 1,
+            "num_key_value_heads": 1,
+            "num_hidden_layers": 1,
+            "vocab_size": 32,
+            "max_position_embeddings": 64
+        }))
+        .unwrap()
+    }
+
+    #[test]
+    fn truncate_sequence_frees_whole_blocks_and_keeps_slot_registered() {
+        if xserv_cuda::device::set_device(0).is_err() {
+            eprintln!("skipping CUDA-backed PagedKVCache test: device 0 unavailable");
+            return;
+        }
+
+        let config = tiny_config();
+        let mut cache = PagedKVCache::new(&config, 5, 0, 1, 4, DType::BF16, 0);
+
+        assert_eq!(
+            cache.truncate_sequence(1, 0),
+            Err("truncate_sequence: slot out of range")
+        );
+        assert_eq!(
+            cache.truncate_sequence(0, 0),
+            Err("truncate_sequence: empty slot")
+        );
+
+        cache.register_sequence(0).unwrap();
+        cache.ensure_capacity(0, BLOCK_SIZE * 3 + 1);
+        cache.advance_seq_len(0, BLOCK_SIZE * 3 + 1);
+        assert_eq!(cache.seq_len(0), BLOCK_SIZE * 3 + 1);
+        assert_eq!(cache.block_count(0), 4);
+        assert_eq!(cache.free_blocks(), 0);
+
+        cache.truncate_sequence(0, BLOCK_SIZE + 1).unwrap();
+        assert_eq!(cache.seq_len(0), BLOCK_SIZE + 1);
+        assert_eq!(cache.block_count(0), 2);
+        assert_eq!(cache.free_blocks(), 2);
+
+        cache.truncate_sequence(0, BLOCK_SIZE).unwrap();
+        assert_eq!(cache.seq_len(0), BLOCK_SIZE);
+        assert_eq!(cache.block_count(0), 1);
+        assert_eq!(cache.free_blocks(), 3);
+
+        cache.truncate_sequence(0, 0).unwrap();
+        assert_eq!(cache.seq_len(0), 0);
+        assert_eq!(cache.block_count(0), 1);
+        assert_eq!(cache.free_blocks(), 3);
+        assert_eq!(
+            cache.truncate_sequence(0, 1),
+            Err("truncate_sequence: cannot extend")
+        );
+    }
+}
+
+unsafe fn tensor_from_owned_buf(
+    buf: GpuBuffer,
+    shape: &[usize],
+    dtype: DType,
+    device: u32,
+) -> Tensor {
+    use smallvec::SmallVec;
+    use xserv_tensor::shape::contiguous_strides;
+    use xserv_tensor::storage::Storage;
+
+    let storage = Storage::cuda(buf, device);
+    Tensor::from_storage(
+        storage,
+        SmallVec::from_slice(shape),
+        contiguous_strides(shape),
+        0,
+        dtype,
+    )
+}

crates/xserv-model/src/qwen3_graph.rs

@@ -0,0 +1,185 @@
+//! CUDA-graph replay for Qwen3 batch=1 decode (Phase 24 / speculative draft).
+//!
+//! Same pattern as `gpt_oss_graph.rs`, but for the Qwen3 dense decode path used
+//! by speculative decoding's draft model. A Qwen3-0.6B decode step is ~140
+//! kernel launches; wrapping the whole step into one `cudaGraphLaunch` cuts
+//! the ~4× γ draft cost per speculative round.
+//!
+//! See `gpt_oss_graph.rs` for the design commentary; the capture preconditions,
+//! retained-warmup mechanism, and quarantine lifetime are all identical here.
+
+use std::ffi::c_void;
+
+use xserv_cuda::allocator::{self, RetainedBlocks};
+use xserv_cuda::{CudaGraph, CudaStream, GpuBuffer};
+use xserv_tensor::Tensor;
+
+use crate::paged_kv_cache::PagedKVCache;
+use crate::qwen3::Qwen3;
+
+pub struct Qwen3DecodeGraph {
+    stream: CudaStream,
+    graph: CudaGraph,
+    ids_buf: GpuBuffer, // [1] u32, persistent graph input
+    pos_buf: GpuBuffer, // [1] u32, persistent graph input
+    logits: Tensor,     // graph output; rewritten in place by every replay
+    _arena: RetainedBlocks,
+}
+
+impl Qwen3DecodeGraph {
+    /// Capture one batch=1 decode step and replay it once.
+    pub fn capture(
+        model: &Qwen3,
+        token: u32,
+        position: usize,
+        slot: usize,
+        cache: &mut PagedKVCache,
+    ) -> Self {
+        let stream = CudaStream::new().expect("create capture stream");
+        let mut ids_buf = allocator::cached_alloc(4).expect("alloc ids buf");
+        let mut pos_buf = allocator::cached_alloc(4).expect("alloc pos buf");
+
+        model.decode_prepare(&[position], &[slot], cache);
+        ids_buf.copy_from_host(&token.to_le_bytes()).unwrap();
+        pos_buf
+            .copy_from_host(&(position as u32).to_le_bytes())
+            .unwrap();
+
+        // Retained warmup: run the exact step once eagerly with the quarantine
+        // ON to stock the pool. See gpt_oss_graph.rs:66-86 for the full
+        // rationale. Re-running the step is idempotent: the KV scatter
+        // overwrites the same cache position and advance_seq_len is *inside*
+        // decode_core, so we roll it back afterwards.
+        let seq_len_before = cache.seq_len(slot);
+        allocator::begin_retain();
+        {
+            let _guard = xserv_cuda::push_stream(&stream);
+            let _ = model.decode_core(
+                ids_buf.as_ptr() as *const c_void,
+                pos_buf.as_ptr() as *const c_void,
+                1,
+                &[slot],
+                cache,
+            );
+        }
+        drop(allocator::end_retain());
+        stream.synchronize().expect("warmup sync");
+        // decode_core advanced seq_len; roll back so capture starts from the
+        // same logical state as the eager warmup.
+        cache
+            .truncate_sequence(slot, seq_len_before)
+            .expect("rollback after warmup");
+
+        allocator::begin_retain();
+        let mut graph = CudaGraph::new();
+        let logits;
+        {
+            let _guard = xserv_cuda::stream::push_stream(&stream);
+            graph
+                .begin_capture(&stream)
+                .expect("begin decode-graph capture");
+            logits = model.decode_core(
+                ids_buf.as_ptr() as *const c_void,
+                pos_buf.as_ptr() as *const c_void,
+                1,
+                &[slot],
+                cache,
+            );
+            graph
+                .end_capture(&stream)
+                .expect("end decode-graph capture");
+        }
+        let arena = allocator::end_retain();
+
+        // The capture path called advance_seq_len (host-side) but the actual
+        // GPU compute has not yet run. Roll back and let the first replay
+        // advance it exactly once with real K/V writes.
+        cache
+            .truncate_sequence(slot, seq_len_before)
+            .expect("rollback after capture");
+
+        graph.launch(&stream).expect("first decode-graph replay");
+        cache.advance_seq_len(slot, 1);
+
+        Self {
+            stream,
+            graph,
+            ids_buf,
+            pos_buf,
+            logits,
+            _arena: arena,
+        }
+    }
+
+    /// Run one decode step by replaying the captured graph.
+    pub fn step(
+        &mut self,
+        model: &Qwen3,
+        token: u32,
+        position: usize,
+        slot: usize,
+        cache: &mut PagedKVCache,
+    ) -> Tensor {
+        model.decode_prepare(&[position], &[slot], cache);
+        self.ids_buf.copy_from_host(&token.to_le_bytes()).unwrap();
+        self.pos_buf
+            .copy_from_host(&(position as u32).to_le_bytes())
+            .unwrap();
+        self.graph
+            .launch(&self.stream)
+            .expect("decode-graph replay");
+        cache.advance_seq_len(slot, 1);
+        self.logits.clone()
+    }
+}
+
+/// Lazy capture policy: first decode step of the process runs eager, the
+/// second is captured, the rest replay. Batch>1 always falls back to eager.
+/// Disable with `XSERV_DECODE_GRAPH=0`.
+pub struct GraphedQwen3Decoder {
+    graph: Option<Qwen3DecodeGraph>,
+    eager_steps: u32,
+    enabled: bool,
+}
+
+impl GraphedQwen3Decoder {
+    pub fn new() -> Self {
+        let enabled = std::env::var("XSERV_DECODE_GRAPH")
+            .map(|v| v != "0")
+            .unwrap_or(true);
+        Self {
+            graph: None,
+            eager_steps: 0,
+            enabled,
+        }
+    }
+
+    pub fn decode(
+        &mut self,
+        model: &Qwen3,
+        tokens: &[u32],
+        positions: &[usize],
+        slots: &[usize],
+        cache: &mut PagedKVCache,
+    ) -> Tensor {
+        if self.enabled && tokens.len() == 1 {
+            if let Some(g) = self.graph.as_mut() {
+                return g.step(model, tokens[0], positions[0], slots[0], cache);
+            }
+            if self.eager_steps >= 1 {
+                let g = Qwen3DecodeGraph::capture(model, tokens[0], positions[0], slots[0], cache);
+                let logits = g.logits.clone();
+                self.graph = Some(g);
+                return logits;
+            }
+        }
+        self.eager_steps += 1;
+        model.forward_decode_paged(tokens, positions, slots, cache)
+    }
+}
+
+impl Default for GraphedQwen3Decoder {
+    fn default() -> Self {
+        Self::new()
+    }
+}

crates/xserv-model/src/sampling.rs

@@ -0,0 +1,197 @@
+use half::bf16;
+use rand::Rng;
+use xserv_tensor::{DType, Device, Tensor};
+
+#[derive(Clone)]
+pub struct SamplingParams {
+    pub temperature: f32,
+    pub top_k: usize,
+    pub top_p: f32,
+}
+
+impl Default for SamplingParams {
+    fn default() -> Self {
+        Self {
+            temperature: 0.0,
+            top_k: 0,
+            top_p: 1.0,
+        }
+    }
+}
+
+/// Sample a token from logits with shape [seq_len, vocab_size].
+/// Uses the last position's logits. Handles both F32 and BF16 dtypes.
+pub fn sample(logits: &Tensor, params: &SamplingParams) -> u32 {
+    assert_eq!(logits.ndim(), 2);
+    // Greedy fast path: GPU argmax + 4-byte D2H instead of copying the whole
+    // [seq, vocab] logits to the host and scanning it (~201k bf16/token).
+    // NaN logits lose every `>` comparison in the kernel, matching the
+    // NaN-safe host argmax below.
+    if params.temperature == 0.0
+        && logits.dtype() == DType::BF16
+        && matches!(logits.device(), Device::Cuda(_))
+        && logits.is_contiguous()
+    {
+        let ids = xserv_kernels::argmax_bf16_to_host(logits);
+        return *ids.last().unwrap();
+    }
+    let vocab_size = logits.shape()[1];
+    let seq_len = logits.shape()[0];
+    let logits_cpu = logits.to_device(Device::Cpu);
+
+    // Extract last row as f32
+    let mut last_row: Vec<f32> = match logits.dtype() {
+        DType::F32 => {
+            let data = logits_cpu.as_slice::<f32>();
+            data[(seq_len - 1) * vocab_size..seq_len * vocab_size].to_vec()
+        }
+        DType::BF16 => {
+            let data = logits_cpu.as_slice::<bf16>();
+            data[(seq_len - 1) * vocab_size..seq_len * vocab_size]
+                .iter()
+                .map(|v| v.to_f32())
+                .collect()
+        }
+        _ => panic!("unsupported dtype for sampling: {:?}", logits.dtype()),
+    };
+
+    // Greedy
+    if params.temperature == 0.0 {
+        return argmax(&last_row);
+    }
+
+    // NaN-safe: sampling path uses partial_cmp().unwrap() in top-k/top-p
+    // sorts and softmax; a single NaN logit would panic the engine thread.
+    // Replace NaN with -inf (equivalent to masking) instead.
+    let mut nan_seen = false;
+    for v in last_row.iter_mut() {
+        if v.is_nan() {
+            nan_seen = true;
+            *v = f32::NEG_INFINITY;
+        }
+    }
+    if nan_seen {
+        eprintln!("[sampling] WARNING: NaN logits encountered in sample()");
+    }
+
+    // Apply temperature
+    let mut logits_f32: Vec<f32> = last_row.iter().map(|v| v / params.temperature).collect();
+
+    // Top-k filtering
+    if params.top_k > 0 && params.top_k < vocab_size {
+        let mut indices: Vec<usize> = (0..vocab_size).collect();
+        indices.select_nth_unstable_by(params.top_k, |&a, &b| {
+            logits_f32[b].partial_cmp(&logits_f32[a]).unwrap()
+        });
+        // Everything after top_k should be masked
+        for &i in &indices[params.top_k..] {
+            logits_f32[i] = f32::NEG_INFINITY;
+        }
+    }
+
+    // Top-p (nucleus) filtering
+    if params.top_p < 1.0 {
+        // Sort indices by descending logit value
+        let mut indices: Vec<usize> = (0..vocab_size).collect();
+        indices.sort_unstable_by(|&a, &b| logits_f32[b].partial_cmp(&logits_f32[a]).unwrap());
+
+        // Compute softmax probabilities for the sorted order
+        let max_val = logits_f32[indices[0]];
+        let sorted_probs: Vec<f32> = indices
+            .iter()
+            .map(|&i| (logits_f32[i] - max_val).exp())
+            .collect();
+        let sum: f32 = sorted_probs.iter().sum();
+        let sorted_probs: Vec<f32> = sorted_probs.iter().map(|v| v / sum).collect();
+
+        // Cumulative sum, find cutoff
+        let mut cumsum = 0.0f32;
+        let mut cutoff = indices.len();
+        for (rank, &prob) in sorted_probs.iter().enumerate() {
+            cumsum += prob;
+            if cumsum > params.top_p {
+                cutoff = rank + 1; // keep at least this many
+                break;
+            }
+        }
+
+        // Mask everything beyond cutoff
+        for &i in &indices[cutoff..] {
+            logits_f32[i] = f32::NEG_INFINITY;
+        }
+    }
+
+    // Softmax
+    let max_val = logits_f32.iter().cloned().fold(f32::NEG_INFINITY, f32::max);
+    let exps: Vec<f32> = logits_f32.iter().map(|v| (v - max_val).exp()).collect();
+    let sum: f32 = exps.iter().sum();
+    let probs: Vec<f32> = exps.iter().map(|v| v / sum).collect();
+
+    // Weighted random sampling
+    let mut rng = rand::thread_rng();
+    let r: f32 = rng.r#gen();
+    let mut cumsum = 0.0f32;
+    for (i, &p) in probs.iter().enumerate() {
+        cumsum += p;
+        if cumsum > r {
+            return i as u32;
+        }
+    }
+
+    // Fallback (rounding edge case)
+    (vocab_size - 1) as u32
+}
+
+/// Greedy argmax with a repetition penalty applied to `recent` token ids
+/// (HF-style: divide positive logits by `penalty`, multiply negative by it).
+/// `penalty <= 1.0` is a no-op. Mitigates greedy repetition loops on reasoning
+/// models without changing the forward pass. NaN-safe.
+pub fn sample_greedy_penalized(logits: &Tensor, recent: &[u32], penalty: f32) -> u32 {
+    assert_eq!(logits.ndim(), 2);
+    let vocab_size = logits.shape()[1];
+    let seq_len = logits.shape()[0];
+    let logits_cpu = logits.to_device(Device::Cpu);
+    let mut last_row: Vec<f32> = match logits.dtype() {
+        DType::F32 => {
+            logits_cpu.as_slice::<f32>()[(seq_len - 1) * vocab_size..seq_len * vocab_size].to_vec()
+        }
+        DType::BF16 => logits_cpu.as_slice::<bf16>()
+            [(seq_len - 1) * vocab_size..seq_len * vocab_size]
+            .iter()
+            .map(|v| v.to_f32())
+            .collect(),
+        _ => panic!("unsupported dtype for sampling: {:?}", logits.dtype()),
+    };
+    if penalty > 1.0 {
+        for &id in recent {
+            let i = id as usize;
+            if i < last_row.len() {
+                let v = last_row[i];
+                last_row[i] = if v > 0.0 { v / penalty } else { v * penalty };
+            }
+        }
+    }
+    argmax(&last_row)
+}
+
+fn argmax(data: &[f32]) -> u32 {
+    // NaN-safe: a single NaN logit must not crash the engine thread (a
+    // partial_cmp().unwrap() panics on NaN). Skip NaNs; warn once if seen.
+    let mut best_i = 0usize;
+    let mut best = f32::NEG_INFINITY;
+    let mut nan_seen = false;
+    for (i, &v) in data.iter().enumerate() {
+        if v.is_nan() {
+            nan_seen = true;
+            continue;
+        }
+        if v > best {
+            best = v;
+            best_i = i;
+        }
+    }
+    if nan_seen {
+        eprintln!("[sampling] WARNING: NaN logits encountered in argmax");
+    }
+    best_i as u32
+}

crates/xserv-server/Cargo.toml

@@ -0,0 +1,24 @@
+[package]
+name = "xserv-server"
+version.workspace = true
+edition.workspace = true
+
+[[bin]]
+name = "xserv-server"
+path = "src/main.rs"
+
+[dependencies]
+xserv-cuda = { path = "../xserv-cuda" }
+xserv-tensor = { path = "../xserv-tensor" }
+xserv-kernels = { path = "../xserv-kernels" }
+xserv-model = { path = "../xserv-model" }
+xserv-tokenizer = { path = "../xserv-tokenizer" }
+xserv-distributed = { path = "../xserv-distributed" }
+half.workspace = true
+serde.workspace = true
+serde_json.workspace = true
+tokio.workspace = true
+axum.workspace = true
+uuid.workspace = true
+tokio-stream.workspace = true
+minijinja.workspace = true

crates/xserv-server/src/api.rs

@@ -0,0 +1,573 @@
+use axum::Extension;
+use axum::Json;
+use axum::http::StatusCode;
+use axum::response::sse::{Event, KeepAlive, Sse};
+use axum::response::{IntoResponse, Response};
+use serde::{Deserialize, Serialize};
+use std::convert::Infallible;
+use std::path::Path;
+use std::sync::Arc;
+use tokio_stream::StreamExt;
+use tokio_stream::wrappers::ReceiverStream;
+use uuid::Uuid;
+
+use crate::AppState;
+use crate::engine::{GenerateEvent, GenerateRequest};
+use xserv_model::SamplingParams;
+
+#[derive(Deserialize)]
+pub struct ChatRequest {
+    #[serde(default)]
+    pub model: Option<String>,
+    pub messages: Vec<Message>,
+    #[serde(default = "default_max_tokens")]
+    pub max_tokens: usize,
+    #[serde(default)]
+    pub stream: Option<bool>,
+    #[serde(default)]
+    pub temperature: Option<f32>,
+    #[serde(default)]
+    pub top_k: Option<usize>,
+    #[serde(default)]
+    pub top_p: Option<f32>,
+}
+
+#[derive(Deserialize, Serialize, Clone)]
+pub struct Message {
+    pub role: String,
+    pub content: String,
+}
+
+fn default_max_tokens() -> usize {
+    256
+}
+
+#[derive(Serialize)]
+pub struct ModelsResponse {
+    object: &'static str,
+    data: Vec<ModelInfo>,
+}
+
+#[derive(Serialize)]
+pub struct ModelInfo {
+    id: String,
+    object: &'static str,
+    owned_by: &'static str,
+}
+
+// ---------------------------------------------------------------------------
+// Chat Template: Jinja2 rendering via minijinja
+// ---------------------------------------------------------------------------
+
+pub struct ChatTemplate {
+    source: String,
+    model_type: String,
+}
+
+impl ChatTemplate {
+    pub fn load(model_dir: &Path, model_type: &str) -> Self {
+        // 1. Try standalone chat_template.jinja file
+        let jinja_path = model_dir.join("chat_template.jinja");
+        if jinja_path.exists() {
+            let source = std::fs::read_to_string(&jinja_path)
+                .unwrap_or_else(|e| panic!("failed to read {}: {e}", jinja_path.display()));
+            eprintln!("[chat-template] loaded from {}", jinja_path.display());
+            return Self {
+                source,
+                model_type: model_type.to_string(),
+            };
+        }
+
+        // 2. Try tokenizer_config.json → chat_template field
+        let tok_cfg_path = model_dir.join("tokenizer_config.json");
+        if tok_cfg_path.exists() {
+            if let Ok(data) = std::fs::read_to_string(&tok_cfg_path) {
+                if let Ok(v) = serde_json::from_str::<serde_json::Value>(&data) {
+                    if let Some(ct) = v.get("chat_template").and_then(|v| v.as_str()) {
+                        eprintln!("[chat-template] loaded from tokenizer_config.json");
+                        return Self {
+                            source: ct.to_string(),
+                            model_type: model_type.to_string(),
+                        };
+                    }
+                }
+            }
+        }
+
+        // 3. No template found — use empty source, will fall back to hardcoded
+        eprintln!("[chat-template] no Jinja template found, using hardcoded fallback");
+        Self {
+            source: String::new(),
+            model_type: model_type.to_string(),
+        }
+    }
+
+    pub fn render(&self, messages: &[Message]) -> String {
+        if self.source.is_empty() {
+            return build_prompt_hardcoded(messages, &self.model_type);
+        }
+
+        match self.render_jinja(messages) {
+            Ok(prompt) => prompt,
+            Err(e) => {
+                eprintln!("[chat-template] Jinja render error: {e}, falling back to hardcoded");
+                build_prompt_hardcoded(messages, &self.model_type)
+            }
+        }
+    }
+
+    fn render_jinja(&self, messages: &[Message]) -> Result<String, minijinja::Error> {
+        let mut env = minijinja::Environment::new();
+
+        // Register custom functions the template may call.
+        env.add_function("strftime_now", strftime_now);
+        env.add_function("raise_exception", raise_exception);
+
+        // Python str methods used by harmony/gpt-oss templates.
+        env.add_filter("startswith", |s: String, prefix: String| -> bool {
+            s.starts_with(&prefix)
+        });
+
+        env.add_template("chat", &self.source)?;
+        let tmpl = env.get_template("chat")?;
+
+        let ctx = minijinja::context! {
+            messages => minijinja::Value::from_serialize(messages),
+            add_generation_prompt => true,
+            bos_token => "",
+            eos_token => "",
+        };
+
+        tmpl.render(ctx)
+    }
+}
+
+fn strftime_now(fmt: String) -> String {
+    use std::time::SystemTime;
+    let now = SystemTime::now()
+        .duration_since(SystemTime::UNIX_EPOCH)
+        .unwrap()
+        .as_secs();
+    // Only support %Y-%m-%d (the only format used by known templates)
+    let days = now / 86400;
+    let (y, m, d) = days_to_ymd(days);
+    fmt.replace("%Y", &format!("{y:04}"))
+        .replace("%m", &format!("{m:02}"))
+        .replace("%d", &format!("{d:02}"))
+}
+
+fn days_to_ymd(days_since_epoch: u64) -> (u32, u32, u32) {
+    // Civil calendar from days since 1970-01-01 (Rata Die algorithm)
+    let z = days_since_epoch as i64 + 719468;
+    let era = (if z >= 0 { z } else { z - 146096 }) / 146097;
+    let doe = (z - era * 146097) as u32;
+    let yoe = (doe - doe / 1460 + doe / 36524 - doe / 146096) / 365;
+    let y = yoe as i64 + era * 400;
+    let doy = doe - (365 * yoe + yoe / 4 - yoe / 100);
+    let mp = (5 * doy + 2) / 153;
+    let d = doy - (153 * mp + 2) / 5 + 1;
+    let m = if mp < 10 { mp + 3 } else { mp - 9 };
+    let y = if m <= 2 { y + 1 } else { y };
+    (y as u32, m, d)
+}
+
+fn raise_exception(msg: String) -> Result<String, minijinja::Error> {
+    Err(minijinja::Error::new(
+        minijinja::ErrorKind::InvalidOperation,
+        msg,
+    ))
+}
+
+// ---------------------------------------------------------------------------
+// Hardcoded fallback templates (for models without a Jinja template)
+// ---------------------------------------------------------------------------
+
+fn build_prompt_hardcoded(messages: &[Message], model_type: &str) -> String {
+    if model_type == "gpt_oss" {
+        return build_prompt_gpt_oss(messages);
+    }
+    // Default: Qwen3 ChatML format
+    let mut prompt = String::new();
+    for msg in messages {
+        match msg.role.as_str() {
+            "system" | "user" | "assistant" => {
+                prompt.push_str("<|im_start|>");
+                prompt.push_str(&msg.role);
+                prompt.push('\n');
+                prompt.push_str(&msg.content);
+                prompt.push_str("<|im_end|>\n");
+            }
+            _ => {}
+        }
+    }
+    prompt.push_str("<|im_start|>assistant\n");
+    prompt.push_str("<think>\n\n</think>\n\n");
+    prompt
+}
+
+fn build_prompt_gpt_oss(messages: &[Message]) -> String {
+    let mut prompt = String::new();
+    // Canonical harmony system message (mirrors the model's chat_template.jinja
+    // build_system_message macro). A hand-rolled substitute puts gpt-oss out of
+    // distribution and destabilizes channel selection. This hardcoded builder is
+    // only a fallback for gpt-oss models that ship no Jinja template; the
+    // gpt-oss-20b release does ship one, so the template path is normally used.
+    prompt.push_str("<|start|>system<|message|>");
+    prompt.push_str("You are ChatGPT, a large language model trained by OpenAI.\n");
+    prompt.push_str("Knowledge cutoff: 2024-06\n");
+    prompt.push_str(&format!(
+        "Current date: {}\n\n",
+        strftime_now("%Y-%m-%d".to_string())
+    ));
+    prompt.push_str("Reasoning: low\n\n");
+    prompt.push_str("# Valid channels: analysis, commentary, final. Channel must be included for every message.");
+    prompt.push_str("<|end|>");
+    let dev_instructions: String = messages
+        .iter()
+        .filter(|m| m.role == "system")
+        .map(|m| m.content.as_str())
+        .collect::<Vec<_>>()
+        .join("\n\n");
+    if !dev_instructions.is_empty() {
+        prompt.push_str("<|start|>developer<|message|># Instructions\n\n");
+        prompt.push_str(&dev_instructions);
+        prompt.push_str("<|end|>");
+    }
+    for msg in messages {
+        match msg.role.as_str() {
+            "user" => {
+                prompt.push_str("<|start|>user<|message|>");
+                prompt.push_str(&msg.content);
+                prompt.push_str("<|end|>");
+            }
+            "assistant" => {
+                prompt.push_str("<|start|>assistant<|channel|>final<|message|>");
+                prompt.push_str(&msg.content);
+                prompt.push_str("<|end|>");
+            }
+            _ => {}
+        }
+    }
+    prompt.push_str("<|start|>assistant<|channel|>final<|message|>");
+    prompt
+}
+
+// ---------------------------------------------------------------------------
+// HTTP handlers
+// ---------------------------------------------------------------------------
+
+pub async fn health() -> &'static str {
+    "ok"
+}
+
+pub async fn list_models(Extension(state): Extension<Arc<AppState>>) -> Json<ModelsResponse> {
+    Json(ModelsResponse {
+        object: "list",
+        data: vec![ModelInfo {
+            id: state.model_name.clone(),
+            object: "model",
+            owned_by: "xserv",
+        }],
+    })
+}
+
+pub async fn chat_completions(
+    Extension(state): Extension<Arc<AppState>>,
+    Json(req): Json<ChatRequest>,
+) -> Response {
+    if req.stream == Some(true) {
+        chat_stream(state, req)
+    } else {
+        chat_non_stream(state, req).await
+    }
+}
+
+async fn chat_non_stream(state: Arc<AppState>, req: ChatRequest) -> Response {
+    let id = format!("chatcmpl-{}", Uuid::new_v4());
+    let model_name = state.model_name.clone();
+    let created = unix_timestamp();
+
+    if let Some(response) = validate_request(&req, &model_name) {
+        return response;
+    }
+
+    let prompt = state.chat_template.render(&req.messages);
+    let prompt_tokens = state.engine_tokenizer.lock().unwrap().encode(&prompt);
+    let prompt_token_count = prompt_tokens.len();
+
+    let max_seq_len = state.max_seq_len;
+    if prompt_token_count >= max_seq_len {
+        return bad_request(format!(
+            "prompt is {} tokens, exceeds max_seq_len {}",
+            prompt_token_count, max_seq_len
+        ));
+    }
+    let max_tokens = req.max_tokens.min(max_seq_len - prompt_token_count);
+
+    let (tx, mut rx) = tokio::sync::mpsc::channel::<GenerateEvent>(64);
+    let gen_req = GenerateRequest {
+        prompt_tokens,
+        max_tokens,
+        sampling: sampling_params(&req),
+        sender: tx,
+    };
+    if let Err(resp) = submit_to_engine(&state, gen_req) {
+        return resp;
+    }
+
+    let mut content = String::new();
+    let mut completion_token_count: usize = 0;
+    let mut finish_reason = "length".to_string();
+    while let Some(event) = rx.recv().await {
+        match event {
+            GenerateEvent::Token { text, .. } => {
+                completion_token_count += 1;
+                content.push_str(&text);
+            }
+            GenerateEvent::Done { finish_reason: fr } => {
+                finish_reason = fr;
+                break;
+            }
+        }
+    }
+
+    let fr_value = match normalize_finish_reason(&finish_reason) {
+        Some(s) => serde_json::Value::String(s.to_string()),
+        None => serde_json::Value::Null,
+    };
+    Json(serde_json::json!({
+        "id": id,
+        "object": "chat.completion",
+        "created": created,
+        "model": model_name,
+        "choices": [{
+            "index": 0,
+            "message": { "role": "assistant", "content": content },
+            "finish_reason": fr_value,
+        }],
+        "usage": {
+            "prompt_tokens": prompt_token_count,
+            "completion_tokens": completion_token_count,
+            "total_tokens": prompt_token_count + completion_token_count
+        }
+    }))
+    .into_response()
+}
+
+fn chat_stream(state: Arc<AppState>, req: ChatRequest) -> Response {
+    let id = format!("chatcmpl-{}", Uuid::new_v4());
+    let model_name = state.model_name.clone();
+    let created = unix_timestamp();
+
+    if let Some(response) = validate_request(&req, &model_name) {
+        return response;
+    }
+
+    let prompt = state.chat_template.render(&req.messages);
+    let prompt_tokens = state.engine_tokenizer.lock().unwrap().encode(&prompt);
+
+    let max_seq_len = state.max_seq_len;
+    if prompt_tokens.len() >= max_seq_len {
+        return bad_request(format!(
+            "prompt is {} tokens, exceeds max_seq_len {}",
+            prompt_tokens.len(),
+            max_seq_len
+        ));
+    }
+    let max_tokens = req.max_tokens.min(max_seq_len - prompt_tokens.len());
+
+    let (engine_tx, engine_rx) = tokio::sync::mpsc::channel::<GenerateEvent>(64);
+    let gen_req = GenerateRequest {
+        prompt_tokens,
+        max_tokens,
+        sampling: sampling_params(&req),
+        sender: engine_tx,
+    };
+    if let Err(resp) = submit_to_engine(&state, gen_req) {
+        return resp;
+    }
+
+    // SSE event channel: engine events -> SSE events
+    let (sse_tx, sse_rx) = tokio::sync::mpsc::channel::<Result<Event, Infallible>>(64);
+
+    tokio::spawn(async move {
+        let mut engine_stream = ReceiverStream::new(engine_rx);
+        let mut first = true;
+
+        while let Some(event) = engine_stream.next().await {
+            match event {
+                GenerateEvent::Token { text, .. } => {
+                    if first {
+                        // First chunk: role announcement
+                        let chunk =
+                            make_chunk(&id, &model_name, created, None, Some("assistant"), None);
+                        let _ = sse_tx.send(Ok(Event::default().data(chunk))).await;
+                        first = false;
+                    }
+                    let chunk = make_chunk(&id, &model_name, created, Some(&text), None, None);
+                    if sse_tx.send(Ok(Event::default().data(chunk))).await.is_err() {
+                        return; // client disconnected
+                    }
+                }
+                GenerateEvent::Done { finish_reason } => {
+                    if first {
+                        // Edge case: Done arrived with no tokens
+                        let chunk =
+                            make_chunk(&id, &model_name, created, None, Some("assistant"), None);
+                        let _ = sse_tx.send(Ok(Event::default().data(chunk))).await;
+                    }
+                    // Only "stop" and "length" are OpenAI-standard values. Internal
+                    // codes like "error" (client-stalled from tp/pp engine) map to
+                    // null so SDK clients see a clean stream close.
+                    let fr = normalize_finish_reason(&finish_reason);
+                    let chunk = make_chunk(&id, &model_name, created, None, None, fr);
+                    let _ = sse_tx.send(Ok(Event::default().data(chunk))).await;
+                    let _ = sse_tx
+                        .send(Ok(Event::default().data("[DONE]".to_string())))
+                        .await;
+                    return;
+                }
+            }
+        }
+    });
+
+    Sse::new(ReceiverStream::new(sse_rx))
+        .keep_alive(KeepAlive::default())
+        .into_response()
+}
+
+fn validate_request(req: &ChatRequest, model_name: &str) -> Option<Response> {
+    if let Some(model) = &req.model {
+        if model != model_name {
+            return Some(bad_request(format!(
+                "model '{model}' is not loaded; available model is '{model_name}'"
+            )));
+        }
+    }
+
+    if req.max_tokens == 0 {
+        return Some(bad_request("max_tokens must be greater than 0"));
+    }
+
+    if let Some(t) = req.temperature {
+        if !t.is_finite() || t < 0.0 {
+            return Some(bad_request("temperature must be a finite value >= 0"));
+        }
+    }
+    if let Some(p) = req.top_p {
+        if !p.is_finite() || !(0.0..=1.0).contains(&p) {
+            return Some(bad_request("top_p must be in [0, 1]"));
+        }
+    }
+    if let Some(k) = req.top_k {
+        if k > 1_000_000 {
+            return Some(bad_request("top_k must be <= 1_000_000"));
+        }
+    }
+
+    None
+}
+
+/// Hand a request to the engine thread. Poison-tolerant (recovers the lock if a
+/// prior handler panicked) and returns a clean 503 instead of panicking when the
+/// engine thread is gone, so one dead engine doesn't cascade into every request.
+fn submit_to_engine(state: &AppState, req: GenerateRequest) -> Result<(), Response> {
+    let sender = state
+        .engine_sender
+        .lock()
+        .unwrap_or_else(|e| e.into_inner());
+    sender.try_send(req).map_err(|err| match err {
+        std::sync::mpsc::TrySendError::Full(_) => {
+            service_unavailable("inference engine is busy, retry later")
+        }
+        std::sync::mpsc::TrySendError::Disconnected(_) => {
+            service_unavailable("inference engine is not available")
+        }
+    })
+}
+
+fn service_unavailable(message: impl Into<String>) -> Response {
+    (
+        StatusCode::SERVICE_UNAVAILABLE,
+        Json(serde_json::json!({
+            "error": { "message": message.into(), "type": "server_error" }
+        })),
+    )
+        .into_response()
+}
+
+fn bad_request(message: impl Into<String>) -> Response {
+    (
+        StatusCode::BAD_REQUEST,
+        Json(serde_json::json!({
+            "error": {
+                "message": message.into(),
+                "type": "invalid_request_error"
+            }
+        })),
+    )
+        .into_response()
+}
+
+fn make_chunk(
+    id: &str,
+    model: &str,
+    created: u64,
+    content: Option<&str>,
+    role: Option<&str>,
+    finish_reason: Option<&str>,
+) -> String {
+    let mut delta = serde_json::Map::new();
+    if let Some(r) = role {
+        delta.insert("role".into(), serde_json::Value::String(r.into()));
+        // Role chunk also includes empty content per OpenAI spec
+        delta.insert("content".into(), serde_json::Value::String(String::new()));
+    }
+    if let Some(c) = content {
+        delta.insert("content".into(), serde_json::Value::String(c.into()));
+    }
+
+    let fr = match finish_reason {
+        Some(r) => serde_json::Value::String(r.into()),
+        None => serde_json::Value::Null,
+    };
+
+    serde_json::json!({
+        "id": id,
+        "object": "chat.completion.chunk",
+        "created": created,
+        "model": model,
+        "choices": [{
+            "index": 0,
+            "delta": delta,
+            "finish_reason": fr,
+        }]
+    })
+    .to_string()
+}
+
+fn unix_timestamp() -> u64 {
+    std::time::SystemTime::now()
+        .duration_since(std::time::UNIX_EPOCH)
+        .unwrap()
+        .as_secs()
+}
+
+fn sampling_params(req: &ChatRequest) -> SamplingParams {
+    SamplingParams {
+        temperature: req.temperature.unwrap_or(0.0),
+        top_k: req.top_k.unwrap_or(0),
+        top_p: req.top_p.unwrap_or(1.0),
+    }
+}
+
+/// Map engine finish_reason strings to OpenAI-standard values. Any engine-internal
+/// code (e.g. "error" from tp/pp client-stall) collapses to None so SDK clients see
+/// a clean null instead of an unknown value.
+fn normalize_finish_reason(fr: &str) -> Option<&'static str> {
+    match fr {
+        "stop" => Some("stop"),
+        "length" => Some("length"),
+        _ => None,
+    }
+}

crates/xserv-server/src/engine.rs

@@ -0,0 +1,460 @@
+use std::collections::VecDeque;
+use std::path::Path;
+use std::sync::Once;
+use std::sync::mpsc;
+use std::time::Instant;
+use xserv_model::loader;
+use xserv_model::{BLOCK_SIZE, ModelConfig, PagedKVCache, Qwen3, SamplingParams, sample};
+use xserv_tensor::{DType, Device};
+use xserv_tokenizer::Tokenizer;
+
+pub struct Engine {
+    model: Qwen3,
+    config: ModelConfig,
+    tokenizer: Tokenizer,
+    max_batch_size: usize,
+    max_seq_len: usize,
+    paged_cache: PagedKVCache,
+}
+
+pub struct GenerateRequest {
+    pub prompt_tokens: Vec<u32>,
+    pub max_tokens: usize,
+    pub sampling: SamplingParams,
+    pub sender: tokio::sync::mpsc::Sender<GenerateEvent>,
+}
+
+pub enum GenerateEvent {
+    Token { id: u32, text: String },
+    Done { finish_reason: String },
+}
+
+struct Sequence {
+    id: u64,
+    prompt_tokens: Vec<u32>,
+    generated_tokens: Vec<u32>,
+    max_tokens: usize,
+    sampling: SamplingParams,
+    seq_slot: Option<usize>,
+    sender: tokio::sync::mpsc::Sender<GenerateEvent>,
+    prefilled: bool,
+    /// Set when a `try_send` failed (client too slow or gone). The scheduler
+    /// reaps the sequence next iteration instead of blocking the decode thread.
+    client_stalled: bool,
+    eos_token_id: Option<u32>,
+    decode_buffer: Vec<u8>,
+    created_at: Instant,
+}
+
+impl Engine {
+    pub fn load(model_dir: &Path, max_batch_size: usize, max_seq_len: usize) -> Self {
+        Self::load_with_swap(model_dir, max_batch_size, max_seq_len, 8)
+    }
+
+    pub fn load_with_swap(
+        model_dir: &Path,
+        max_batch_size: usize,
+        max_seq_len: usize,
+        swap_space_gb: usize,
+    ) -> Self {
+        xserv_cuda::device::set_device(0).unwrap();
+        let config = ModelConfig::from_file(&model_dir.join("config.json"));
+        eprintln!("[engine] Loading weights...");
+        let weights = loader::load_model_dir(model_dir, Device::Cuda(0));
+        eprintln!("[engine] Loaded {} tensors", weights.len());
+        let model = Qwen3::from_weights(config.clone(), weights);
+        let tokenizer = Tokenizer::from_file(&model_dir.join("tokenizer.json"));
+
+        // Tier-1 sizing: size the GPU block pool to *available VRAM* after the
+        // weights are resident, not to worst-case max_batch * max_ctx. This is
+        // what makes paged attention elastic — sequences share the pool on
+        // demand, and overflow is swapped to host (Tier-2) rather than reserved.
+        let bytes_per_block = PagedKVCache::bytes_per_block(&config, DType::BF16);
+        let info = xserv_cuda::device::device_info(0).expect("device info");
+        // Reserve headroom for activations, cuBLAS workspace and the [B, vocab]
+        // logits buffer; the transpose peak during load is already behind us.
+        const ACTIVATION_RESERVE: usize = 3 * 1024 * 1024 * 1024; // 3 GiB
+        let util_num = 90; // use 90% of remaining free memory for KV
+        let usable = info.free_memory.saturating_sub(ACTIVATION_RESERVE);
+        let mut total_blocks = (usable * util_num / 100) / bytes_per_block;
+        // Cap at a sane upper bound and ensure a floor.
+        total_blocks = total_blocks.max(256);
+        // Test hook: force a small GPU pool to exercise the swap path. Must stay
+        // >= max_blocks_per_seq so a single max-length sequence still fits.
+        if let Ok(v) = std::env::var("XSERV_MAX_KV_BLOCKS") {
+            if let Ok(n) = v.parse::<usize>() {
+                total_blocks = total_blocks.min(n);
+                eprintln!("[engine] XSERV_MAX_KV_BLOCKS override: gpu_blocks={total_blocks}");
+            }
+        }
+
+        let max_blocks_per_seq = (max_seq_len + BLOCK_SIZE - 1) / BLOCK_SIZE;
+        // Slots must cover running + swapped sequences, so be generous (cheap:
+        // each slot is just a block-table row of i32s).
+        let max_seqs_slots = (max_batch_size * 8).max(32);
+        // CPU swap pool: swap_space_gb of pinned host memory.
+        let cpu_total_blocks = (swap_space_gb * 1024 * 1024 * 1024) / bytes_per_block;
+
+        let paged_cache = PagedKVCache::new(
+            &config,
+            total_blocks,
+            cpu_total_blocks,
+            max_seqs_slots,
+            max_blocks_per_seq,
+            DType::BF16,
+            0,
+        );
+
+        eprintln!(
+            "[engine] Ready (max_batch={max_batch_size}, max_seq_len={max_seq_len}, \
+             gpu_blocks={total_blocks} ({:.1} GiB), swap_blocks={cpu_total_blocks} ({swap_space_gb} GiB), \
+             free_vram={:.1} GiB)",
+            (total_blocks * bytes_per_block) as f64 / 1e9,
+            info.free_memory as f64 / 1e9,
+        );
+        Self {
+            model,
+            config,
+            tokenizer,
+            max_batch_size,
+            max_seq_len,
+            paged_cache,
+        }
+    }
+
+    pub fn tokenizer(&self) -> &Tokenizer {
+        &self.tokenizer
+    }
+
+    pub fn max_seq_len(&self) -> usize {
+        self.max_seq_len
+    }
+
+    /// Main scheduler loop. Receives requests from channel, manages concurrent sequences.
+    ///
+    /// Sequences move between three sets:
+    ///   waiting  — admitted to the queue, no GPU slot yet
+    ///   running  — KV resident on GPU, actively prefilling/decoding
+    ///   swapped  — KV evicted to pinned host memory (preempted), paused
+    /// When running sequences grow past the GPU block pool, the newest are
+    /// swapped out to host (vLLM-style) and swapped back in when blocks free up.
+    pub fn run(&mut self, rx: mpsc::Receiver<GenerateRequest>) {
+        let mut waiting: VecDeque<Sequence> = VecDeque::new();
+        let mut running: Vec<Sequence> = Vec::new();
+        let mut swapped: Vec<Sequence> = Vec::new();
+        let mut next_id: u64 = 0;
+
+        eprintln!("[scheduler] Listening for requests...");
+
+        loop {
+            // Step 1: Remove finished sequences and return their slots.
+            let finished_slots: Vec<usize> = running
+                .iter()
+                .filter(|s| is_finished(s))
+                .filter_map(|s| s.seq_slot)
+                .collect();
+            for slot in finished_slots {
+                self.paged_cache.free_sequence(slot);
+            }
+            running.retain(|seq| !is_finished(seq));
+
+            // Step 2: Swap previously-evicted sequences back in when there is
+            // room (oldest first). They resume decoding from where they paused.
+            while running.len() < self.max_batch_size && !swapped.is_empty() {
+                let slot = swapped[0].seq_slot.expect("swapped slot");
+                if !self.paged_cache.can_swap_in(slot) {
+                    break;
+                }
+                self.paged_cache.swap_in(slot).expect("swap_in");
+                let seq = swapped.remove(0);
+                eprintln!(
+                    "[scheduler] swapped in seq {} ({} blocks)",
+                    seq.id,
+                    self.paged_cache.block_count(slot)
+                );
+                running.push(seq);
+            }
+
+            // Step 3: Admit new sequences (block-aware). Only admit if the GPU
+            // pool can hold the prompt AND leave one block of decode headroom
+            // per already-running sequence, so admission never starves decode.
+            {
+                let mut avail = self.paged_cache.free_blocks();
+                let decode_reserve = running.len();
+                while running.len() < self.max_batch_size {
+                    let Some(front) = waiting.front() else {
+                        break;
+                    };
+                    let prompt_blocks = front.prompt_tokens.len().div_ceil(BLOCK_SIZE).max(1);
+                    if avail < prompt_blocks + decode_reserve {
+                        break;
+                    }
+                    let free_slot = (0..self.paged_cache.max_seqs())
+                        .find(|&s| self.paged_cache.is_slot_free(s));
+                    let Some(slot) = free_slot else {
+                        break;
+                    };
+                    let mut seq = waiting.pop_front().unwrap();
+                    self.paged_cache
+                        .register_sequence(slot)
+                        .expect("register paged slot");
+                    seq.seq_slot = Some(slot);
+                    running.push(seq);
+                    avail -= prompt_blocks; // projected free after this seq prefills
+                }
+            }
+
+            // Step 4: If nothing to do, blocking wait for new request.
+            if running.is_empty() && waiting.is_empty() && swapped.is_empty() {
+                match rx.recv() {
+                    Ok(req) => {
+                        let seq = self.make_sequence(req, &mut next_id);
+                        waiting.push_back(seq);
+                        continue;
+                    }
+                    Err(_) => break, // channel closed
+                }
+            }
+            // Nothing runnable this iteration (e.g. all swapped, waiting on
+            // blocks to free): loop to retry swap-in/admission next iteration.
+            if running.is_empty() {
+                continue;
+            }
+
+            // Step 5a: Process prefills (one at a time — different prompt lengths).
+            // Admission guaranteed block headroom, so ensure_capacity won't starve.
+            let mut newly_prefilled = Vec::new();
+            for seq in running.iter_mut() {
+                if !seq.prefilled {
+                    let slot = seq.seq_slot.expect("slot");
+                    let logits = self.model.forward_prefill_paged(
+                        &seq.prompt_tokens,
+                        slot,
+                        &mut self.paged_cache,
+                    );
+                    let next = sample(&logits, &seq.sampling);
+                    seq.generated_tokens.push(next);
+                    seq.prefilled = true;
+                    emit_token(&self.tokenizer, seq, next);
+                    newly_prefilled.push(seq.id);
+                }
+            }
+
+            // Step 5b: Ensure block headroom for this decode step; preempt the
+            // newest running sequences to host if the pool can't cover it.
+            let mut needed = decode_block_need(&self.paged_cache, &running, &newly_prefilled);
+            while self.paged_cache.free_blocks() < needed {
+                // Victim: newest prefilled, decoding (not just-prefilled) sequence.
+                let victim = (0..running.len()).rev().find(|&p| {
+                    running[p].prefilled
+                        && !newly_prefilled.contains(&running[p].id)
+                        && running[p].seq_slot.is_some()
+                });
+                let Some(pos) = victim else {
+                    break;
+                };
+                let seq = running.remove(pos);
+                let slot = seq.seq_slot.unwrap();
+                if self.paged_cache.can_swap_out(slot) {
+                    let nblocks = self.paged_cache.block_count(slot);
+                    self.paged_cache.swap_out(slot).expect("swap_out");
+                    eprintln!(
+                        "[scheduler] preempt: swapped out seq {} ({nblocks} blocks) to host",
+                        seq.id
+                    );
+                    swapped.push(seq);
+                    needed = decode_block_need(&self.paged_cache, &running, &newly_prefilled);
+                } else {
+                    running.insert(pos, seq); // CPU pool full — can't evict further
+                    break;
+                }
+            }
+
+            // Step 5c: Batched paged decode for the surviving prefilled sequences.
+            let decode_indices: Vec<usize> = running
+                .iter()
+                .enumerate()
+                .filter(|(_, s)| s.prefilled && !newly_prefilled.contains(&s.id))
+                .map(|(i, _)| i)
+                .collect();
+
+            if !decode_indices.is_empty() {
+                static LOG_ONCE: Once = Once::new();
+                LOG_ONCE.call_once(|| {
+                    eprintln!("[scheduler] paged decode active");
+                });
+
+                let tokens: Vec<u32> = decode_indices
+                    .iter()
+                    .map(|&i| *running[i].generated_tokens.last().unwrap())
+                    .collect();
+                let positions: Vec<usize> = decode_indices
+                    .iter()
+                    .map(|&i| self.paged_cache.seq_len(running[i].seq_slot.unwrap()))
+                    .collect();
+                let slots: Vec<usize> = decode_indices
+                    .iter()
+                    .map(|&i| running[i].seq_slot.unwrap())
+                    .collect();
+
+                let logits = self.model.forward_decode_paged(
+                    &tokens,
+                    &positions,
+                    &slots,
+                    &mut self.paged_cache,
+                );
+
+                // Fast path: every active sequence is greedy → run argmax on
+                // the GPU and only D2H the chosen token ids (a few bytes per
+                // sequence) instead of the full [B, vocab_size] BF16 logits
+                // (~1.2 MB for B=4, Qwen3 vocab=152K).
+                let all_greedy = decode_indices
+                    .iter()
+                    .all(|&i| running[i].sampling.temperature == 0.0);
+                if all_greedy {
+                    let next_ids = xserv_kernels::argmax_bf16_to_host(&logits);
+                    for (j, &i) in decode_indices.iter().enumerate() {
+                        let next = next_ids[j];
+                        running[i].generated_tokens.push(next);
+                        emit_token(&self.tokenizer, &mut running[i], next);
+                    }
+                } else {
+                    // Mixed sampling: keep the CPU path for now (top-k/top-p
+                    // sampling still runs there). Only the rows that need it
+                    // get exercised; greedy rows could in principle reuse the
+                    // GPU argmax but the CPU pass is short for B<=4.
+                    let vocab_size = logits.shape()[1];
+                    let logits_cpu = logits.to_device(xserv_tensor::Device::Cpu);
+                    let data = logits_cpu.as_slice::<half::bf16>();
+                    for (j, &i) in decode_indices.iter().enumerate() {
+                        let row_start = j * vocab_size;
+                        let row_logits = &data[row_start..row_start + vocab_size];
+                        let next = if running[i].sampling.temperature == 0.0 {
+                            row_logits
+                                .iter()
+                                .enumerate()
+                                .max_by(|a, b| a.1.to_f32().partial_cmp(&b.1.to_f32()).unwrap())
+                                .map(|(idx, _)| idx as u32)
+                                .unwrap()
+                        } else {
+                            let row_tensor =
+                                xserv_tensor::Tensor::from_slice(row_logits, &[1, vocab_size]);
+                            sample(&row_tensor, &running[i].sampling)
+                        };
+                        running[i].generated_tokens.push(next);
+                        emit_token(&self.tokenizer, &mut running[i], next);
+                    }
+                }
+            }
+
+            // Step 6: Check for newly arrived requests (non-blocking)
+            loop {
+                match rx.try_recv() {
+                    Ok(req) => {
+                        let seq = self.make_sequence(req, &mut next_id);
+                        waiting.push_back(seq);
+                    }
+                    Err(mpsc::TryRecvError::Empty) => break,
+                    Err(mpsc::TryRecvError::Disconnected) => return,
+                }
+            }
+        }
+    }
+
+    fn make_sequence(&mut self, req: GenerateRequest, next_id: &mut u64) -> Sequence {
+        let id = *next_id;
+        *next_id += 1;
+        Sequence {
+            id,
+            prompt_tokens: req.prompt_tokens,
+            generated_tokens: Vec::new(),
+            max_tokens: req.max_tokens,
+            sampling: req.sampling,
+            seq_slot: None,
+            sender: req.sender,
+            prefilled: false,
+            client_stalled: false,
+            eos_token_id: self.tokenizer.eos_token_id(),
+            decode_buffer: Vec::new(),
+            created_at: Instant::now(),
+        }
+    }
+}
+
+/// Total additional GPU blocks the next decode step needs across all
+/// currently-decoding (prefilled, not just-prefilled) sequences.
+fn decode_block_need(paged: &PagedKVCache, running: &[Sequence], newly_prefilled: &[u64]) -> usize {
+    running
+        .iter()
+        .filter(|s| s.prefilled && !newly_prefilled.contains(&s.id))
+        .filter_map(|s| s.seq_slot)
+        .map(|slot| paged.additional_blocks_needed(slot, 1))
+        .sum()
+}
+
+fn emit_token(tokenizer: &Tokenizer, seq: &mut Sequence, token_id: u32) {
+    if tokenizer.eos_token_id() == Some(token_id) {
+        let tail = tokenizer.flush_decode_stream(&mut seq.decode_buffer);
+        send_token_if_nonempty(seq, tail);
+        try_send_event(
+            seq,
+            GenerateEvent::Done {
+                finish_reason: "stop".to_string(),
+            },
+        );
+        return;
+    }
+
+    let text = tokenizer.decode_token_stream(token_id, &mut seq.decode_buffer);
+    if seq.generated_tokens.len() >= seq.max_tokens {
+        let tail = tokenizer.flush_decode_stream(&mut seq.decode_buffer);
+        send_token_if_nonempty(seq, text);
+        send_token_if_nonempty(seq, tail);
+        try_send_event(
+            seq,
+            GenerateEvent::Done {
+                finish_reason: "length".to_string(),
+            },
+        );
+    } else {
+        send_token_if_nonempty(seq, text);
+    }
+}
+
+fn send_token_if_nonempty(seq: &mut Sequence, text: String) {
+    if !text.is_empty() {
+        let id = *seq.generated_tokens.last().unwrap_or(&0);
+        try_send_event(seq, GenerateEvent::Token { id, text });
+    }
+}
+
+/// Send an event without blocking the shared decode thread. If the client is
+/// too slow (channel full) or gone (closed), flag the sequence for eviction
+/// instead of blocking — one slow consumer must never stall the whole
+/// continuous-batching loop. When the sequence is reaped its `sender` drops,
+/// closing the channel so the client's receive loop ends rather than hanging.
+fn try_send_event(seq: &mut Sequence, event: GenerateEvent) {
+    if let Err(err) = seq.sender.try_send(event) {
+        seq.client_stalled = true;
+        if let tokio::sync::mpsc::error::TrySendError::Full(_) = err {
+            eprintln!(
+                "[scheduler] seq {}: client too slow (stream channel full), evicting",
+                seq.id
+            );
+        }
+    }
+}
+
+fn is_finished(seq: &Sequence) -> bool {
+    if seq.client_stalled {
+        return true;
+    }
+    if seq.generated_tokens.is_empty() {
+        return false;
+    }
+    let last = *seq.generated_tokens.last().unwrap();
+    if seq.generated_tokens.len() >= seq.max_tokens {
+        return true;
+    }
+    seq.sender.is_closed() || seq.eos_token_id == Some(last)
+}

crates/xserv-server/src/main.rs

@@ -0,0 +1,153 @@
+mod api;
+mod engine;
+mod pp_engine;
+mod tp_engine;
+
+use axum::{
+    Extension, Router,
+    extract::DefaultBodyLimit,
+    routing::{get, post},
+};
+use engine::GenerateRequest;
+use std::path::PathBuf;
+use std::sync::{Arc, Mutex, mpsc};
+use xserv_model::ModelConfig;
+
+pub struct AppState {
+    pub model_name: String,
+    pub chat_template: api::ChatTemplate,
+    pub engine_sender: Mutex<mpsc::SyncSender<GenerateRequest>>,
+    pub engine_tokenizer: Mutex<xserv_tokenizer::Tokenizer>,
+    pub max_seq_len: usize,
+}
+
+#[tokio::main]
+async fn main() {
+    let args: Vec<String> = std::env::args().collect();
+    if args.len() < 2 {
+        eprintln!(
+            "Usage: xserv-server <model-dir> [--port PORT] [--max-batch N] [--max-seq-len N] [--swap-space-gb N] [--tp N] [--pp N]"
+        );
+        std::process::exit(1);
+    }
+
+    let model_dir = PathBuf::from(&args[1]);
+    let port: u16 = args
+        .iter()
+        .position(|a| a == "--port")
+        .and_then(|i| args.get(i + 1))
+        .and_then(|s| s.parse().ok())
+        .unwrap_or(8080);
+    let max_batch: usize = args
+        .iter()
+        .position(|a| a == "--max-batch")
+        .and_then(|i| args.get(i + 1))
+        .and_then(|s| s.parse().ok())
+        .unwrap_or(4)
+        .max(1);
+    let requested_max_seq_len: usize = args
+        .iter()
+        .position(|a| a == "--max-seq-len")
+        .and_then(|i| args.get(i + 1))
+        .and_then(|s| s.parse().ok())
+        .unwrap_or(2048)
+        .max(1);
+    let swap_space_gb: usize = args
+        .iter()
+        .position(|a| a == "--swap-space-gb")
+        .and_then(|i| args.get(i + 1))
+        .and_then(|s| s.parse().ok())
+        .unwrap_or(8);
+    let tp: usize = args
+        .iter()
+        .position(|a| a == "--tp")
+        .and_then(|i| args.get(i + 1))
+        .and_then(|s| s.parse().ok())
+        .unwrap_or(1)
+        .max(1);
+    let pp: usize = args
+        .iter()
+        .position(|a| a == "--pp")
+        .and_then(|i| args.get(i + 1))
+        .and_then(|s| s.parse().ok())
+        .unwrap_or(1)
+        .max(1);
+    if tp > 1 && pp > 1 {
+        eprintln!("--tp and --pp cannot be combined yet (2D TP×PP is future work)");
+        std::process::exit(1);
+    }
+    let model_config = ModelConfig::from_file(&model_dir.join("config.json"));
+    // gpt-oss is only implemented in the TP engine; route it there even at
+    // tp=1 (single-rank world) so quantized models can serve on one GPU.
+    let is_gpt_oss = model_config.model_type.as_deref() == Some("gpt_oss");
+    if pp > 1 && is_gpt_oss {
+        eprintln!(
+            "gpt-oss is not supported by the pipeline-parallel engine (Qwen3 only); use --tp instead"
+        );
+        std::process::exit(1);
+    }
+    let model_max_seq_len = model_config.max_seq_len();
+    if model_max_seq_len == 0 {
+        eprintln!("model config has invalid max_seq_len=0");
+        std::process::exit(1);
+    }
+    let max_seq_len = requested_max_seq_len.min(model_max_seq_len);
+    if max_seq_len != requested_max_seq_len {
+        eprintln!(
+            "[server] --max-seq-len {requested_max_seq_len} exceeds model limit {model_max_seq_len}; using {max_seq_len}"
+        );
+    }
+
+    let model_name = model_dir
+        .file_name()
+        .map(|n| n.to_string_lossy().to_string())
+        .unwrap_or_else(|| "unknown".to_string());
+
+    let tokenizer = xserv_tokenizer::Tokenizer::from_file(&model_dir.join("tokenizer.json"));
+
+    // Bounded channel to backpressure incoming requests when the engine falls
+    // behind, instead of letting them pile up in RAM. try_send in the API
+    // handler surfaces this as 503 to the client.
+    let (tx, rx) = mpsc::sync_channel::<GenerateRequest>(256);
+
+    let model_dir_clone = model_dir.clone();
+    std::thread::spawn(move || {
+        if pp > 1 {
+            // Pipeline-parallel path: stage-0 coordinator + worker stage threads.
+            pp_engine::run_pp(&model_dir_clone, pp, max_seq_len, rx);
+        } else if tp <= 1 && !is_gpt_oss {
+            let mut engine = engine::Engine::load_with_swap(
+                &model_dir_clone,
+                max_batch,
+                max_seq_len,
+                swap_space_gb,
+            );
+            engine.run(rx);
+        } else {
+            // Tensor-parallel path: rank-0 coordinator + worker rank threads.
+            tp_engine::run_tp(&model_dir_clone, tp, max_seq_len, rx);
+        }
+    });
+
+    let model_type = model_config.model_type.clone().unwrap_or_default();
+    let chat_template = api::ChatTemplate::load(&model_dir, &model_type);
+    let state = Arc::new(AppState {
+        model_name,
+        chat_template,
+        engine_sender: Mutex::new(tx),
+        engine_tokenizer: Mutex::new(tokenizer),
+        max_seq_len,
+    });
+
+    let app = Router::new()
+        .route("/health", get(api::health))
+        .route("/v1/models", get(api::list_models))
+        .route("/v1/chat/completions", post(api::chat_completions))
+        .layer(DefaultBodyLimit::max(4 * 1024 * 1024))
+        .layer(Extension(state));
+
+    let addr = format!("0.0.0.0:{port}");
+    eprintln!("[server] Listening on {addr} (max_batch={max_batch}, max_seq_len={max_seq_len})");
+    let listener = tokio::net::TcpListener::bind(&addr).await.unwrap();
+    axum::serve(listener, app).await.unwrap();
+}

crates/xserv-server/src/pp_engine.rs

@@ -0,0 +1,338 @@
+//! Pipeline-parallel inference engine for the HTTP server (Phase 18).
+//!
+//! Layer-wise split: stage `s` holds layers `[s*L, (s+1)*L)`. Stage 0 owns the
+//! token embedding and acts as the coordinator (scheduler + tokenizer + response
+//! sender + stop logic); the last stage owns `norm`/`lm_head` and does sampling.
+//! Hidden states are handed off stage->stage via NCCL P2P (`PpContext`); the
+//! sampled token id (a single u32) is returned last-stage -> stage0 over an
+//! in-process channel (same process, so no NCCL needed for that).
+//!
+//! v1 is serial: one request at a time, one token per step, the pipeline is
+//! filled and drained each step (stage0's decode step t+1 depends on the token
+//! the last stage sampled at step t). This gives correctness + per-GPU memory
+//! savings; throughput via microbatch/1F1B overlap is future work
+//! (see docs/18-pipeline-parallelism.md).
+
+use std::ffi::c_void;
+use std::path::{Path, PathBuf};
+use std::sync::Arc;
+use std::sync::mpsc;
+use std::thread;
+
+use half::bf16;
+use xserv_distributed::{PpContext, UniqueId};
+use xserv_model::loader;
+use xserv_model::sampling::SamplingParams;
+use xserv_model::{BLOCK_SIZE, ModelConfig, PagedKVCache, Qwen3, sample};
+use xserv_tensor::{DType, Device, Tensor};
+use xserv_tokenizer::Tokenizer;
+
+use crate::engine::{GenerateEvent, GenerateRequest};
+
+/// Control messages from the coordinator (stage 0) to a worker stage. The heavy
+/// hidden-state tensors do NOT travel here — they go GPU->GPU over NCCL. Only
+/// tiny control info (slot ids, token count, sampling params) is sent.
+#[derive(Clone)]
+enum PpCommand {
+    Register(usize),
+    Free(usize),
+    /// Receive `[n_tokens, hidden]` from the previous stage, run this stage's
+    /// layers; if last stage, sample with `sampling` and return the token.
+    Prefill {
+        n_tokens: usize,
+        slot: usize,
+        sampling: SamplingParams,
+    },
+    /// Receive `[1, hidden]`, run this stage's layers; last stage samples.
+    Decode {
+        slot: usize,
+        sampling: SamplingParams,
+    },
+    Shutdown,
+}
+
+struct StageCtx {
+    model: Qwen3,
+    cache: PagedKVCache,
+    pp: Arc<PpContext>,
+    hidden: usize,
+    device: u32,
+}
+
+/// Build this stage: NCCL init, load + slice weights, size a per-stage KV pool
+/// for THIS stage's layers only (so per-GPU KV is ~1/P).
+fn build_stage(
+    model_dir: &Path,
+    config: &ModelConfig,
+    stage: usize,
+    world: usize,
+    device: u32,
+    max_seq_len: usize,
+    id: UniqueId,
+) -> StageCtx {
+    let pp = Arc::new(PpContext::init(stage, world, id, device));
+    let weights = loader::load_model_dir(model_dir, Device::Cpu);
+    let model = Qwen3::from_weights_pp(config.clone(), weights, stage, world, device);
+
+    // The KV cache only needs this stage's layers; build it from a config clone
+    // whose layer count is the per-stage count (heads are NOT split under PP).
+    let per_stage = config.num_layers() / world;
+    let mut stage_config = config.clone();
+    stage_config.num_hidden_layers = Some(per_stage);
+
+    let max_blocks_per_seq = max_seq_len.div_ceil(BLOCK_SIZE);
+    let total_blocks = max_blocks_per_seq + 8; // v1 serial: one active sequence
+    let cache = PagedKVCache::new(
+        &stage_config,
+        total_blocks,
+        0,
+        4,
+        max_blocks_per_seq,
+        DType::BF16,
+        device,
+    );
+    StageCtx {
+        model,
+        cache,
+        pp,
+        hidden: config.hidden(),
+        device,
+    }
+}
+
+/// Allocate a zeroed `[n, hidden]` device tensor and receive into it from `peer`.
+fn recv_hidden(sc: &StageCtx, n: usize, peer: usize) -> Tensor {
+    let zeros = vec![bf16::ZERO; n * sc.hidden];
+    let x = Tensor::from_slice(&zeros, &[n, sc.hidden]).to_device(Device::Cuda(sc.device));
+    let ptr = x.storage().gpu_buffer().as_ptr() as *mut c_void;
+    sc.pp.recv_bf16_ptr(ptr, n * sc.hidden, peer);
+    xserv_cuda::device::synchronize().unwrap();
+    x
+}
+
+/// Send the `[*, hidden]` hidden state to `peer`, then synchronize so NCCL has
+/// finished reading `x` before it is dropped/reused.
+fn send_hidden(sc: &StageCtx, x: &Tensor, peer: usize) {
+    let ptr = x.storage().gpu_buffer().as_ptr() as *const c_void;
+    sc.pp.send_bf16_ptr(ptr, x.numel(), peer);
+    xserv_cuda::device::synchronize().unwrap();
+}
+
+fn worker_loop(
+    stage: usize,
+    world: usize,
+    id: UniqueId,
+    model_dir: PathBuf,
+    config: ModelConfig,
+    max_seq_len: usize,
+    cmd_rx: mpsc::Receiver<PpCommand>,
+    ack_tx: mpsc::Sender<()>,
+    token_tx: mpsc::Sender<u32>,
+) {
+    let mut sc = build_stage(
+        &model_dir,
+        &config,
+        stage,
+        world,
+        stage as u32,
+        max_seq_len,
+        id,
+    );
+    let is_last = stage == world - 1;
+    let prev = stage - 1;
+    let next = stage + 1;
+
+    while let Ok(cmd) = cmd_rx.recv() {
+        match cmd {
+            PpCommand::Register(slot) => {
+                let _ = sc.cache.register_sequence(slot);
+                let _ = ack_tx.send(());
+            }
+            PpCommand::Free(slot) => {
+                sc.cache.free_sequence(slot);
+                let _ = ack_tx.send(());
+            }
+            PpCommand::Prefill {
+                n_tokens,
+                slot,
+                sampling,
+            } => {
+                let x = recv_hidden(&sc, n_tokens, prev);
+                let x = sc.model.forward_layers_prefill(x, slot, &mut sc.cache);
+                if is_last {
+                    let logits = sc.model.head(&x);
+                    let _ = token_tx.send(sample(&logits, &sampling));
+                } else {
+                    send_hidden(&sc, &x, next);
+                }
+            }
+            PpCommand::Decode { slot, sampling } => {
+                let x = recv_hidden(&sc, 1, prev);
+                let x = sc.model.forward_layers_decode(x, &[slot], &mut sc.cache);
+                if is_last {
+                    let logits = sc.model.head(&x);
+                    let _ = token_tx.send(sample(&logits, &sampling));
+                } else {
+                    send_hidden(&sc, &x, next);
+                }
+            }
+            PpCommand::Shutdown => {
+                let _ = ack_tx.send(());
+                break;
+            }
+        }
+    }
+}
+
+/// Run the PP coordinator (stage 0) on the calling thread. Spawns worker stages
+/// 1..world and consumes generation requests from `rx`.
+pub fn run_pp(
+    model_dir: &Path,
+    world: usize,
+    max_seq_len: usize,
+    rx: mpsc::Receiver<GenerateRequest>,
+) {
+    assert!(world >= 2, "run_pp requires world >= 2");
+    let config = ModelConfig::from_file(&model_dir.join("config.json"));
+    assert!(
+        config.num_layers() % world == 0,
+        "num_layers {} not divisible by pp {world}",
+        config.num_layers()
+    );
+    let tokenizer = Tokenizer::from_file(&model_dir.join("tokenizer.json"));
+    let id = xserv_distributed::get_unique_id();
+
+    // Worker stages 1..world. Each gets a control channel; all share one ack
+    // channel and one token channel (only the last stage actually sends tokens).
+    let (ack_tx, ack_rx) = mpsc::channel::<()>();
+    let (token_tx, token_rx) = mpsc::channel::<u32>();
+    let mut cmd_txs: Vec<mpsc::Sender<PpCommand>> = Vec::new();
+    for stage in 1..world {
+        let (ctx_tx, ctx_rx) = mpsc::channel::<PpCommand>();
+        cmd_txs.push(ctx_tx);
+        let ack_tx = ack_tx.clone();
+        let token_tx = token_tx.clone();
+        let model_dir = model_dir.to_path_buf();
+        let config = config.clone();
+        thread::spawn(move || {
+            worker_loop(
+                stage,
+                world,
+                id,
+                model_dir,
+                config,
+                max_seq_len,
+                ctx_rx,
+                ack_tx,
+                token_tx,
+            );
+        });
+    }
+
+    // Stage 0 (this thread): coordinator + embedding + first layers.
+    let mut sc = build_stage(model_dir, &config, 0, world, 0, max_seq_len, id);
+    eprintln!("[pp-engine] ready (pp={world}, max_seq_len={max_seq_len})");
+
+    let n_workers = world - 1;
+    let next_peer = 1usize;
+    let broadcast = |txs: &[mpsc::Sender<PpCommand>], cmd: PpCommand| {
+        for t in txs {
+            let _ = t.send(cmd.clone());
+        }
+    };
+    let wait_acks = |rx: &mpsc::Receiver<()>| {
+        for _ in 0..n_workers {
+            let _ = rx.recv();
+        }
+    };
+
+    let slot = 0usize;
+    while let Ok(req) = rx.recv() {
+        broadcast(&cmd_txs, PpCommand::Register(slot));
+        sc.cache.register_sequence(slot).expect("register slot");
+        wait_acks(&ack_rx);
+
+        // Prefill: embed prompt, run stage-0 layers, push hidden into the pipe.
+        broadcast(
+            &cmd_txs,
+            PpCommand::Prefill {
+                n_tokens: req.prompt_tokens.len(),
+                slot,
+                sampling: req.sampling.clone(),
+            },
+        );
+        let x = sc.model.embed(&req.prompt_tokens);
+        let x = sc.model.forward_layers_prefill(x, slot, &mut sc.cache);
+        send_hidden(&sc, &x, next_peer);
+        let mut next = token_rx.recv().expect("prefill token");
+
+        let mut decode_buf: Vec<u8> = Vec::new();
+        let mut generated = 1usize;
+        let mut stalled = !emit_text(&tokenizer, &req, next, &mut decode_buf);
+
+        let finish = loop {
+            if stalled {
+                break "error";
+            }
+            if tokenizer.is_eos(next) {
+                break "stop";
+            }
+            if generated >= req.max_tokens {
+                break "length";
+            }
+            broadcast(
+                &cmd_txs,
+                PpCommand::Decode {
+                    slot,
+                    sampling: req.sampling.clone(),
+                },
+            );
+            let x = sc.model.embed(&[next]);
+            let x = sc.model.forward_layers_decode(x, &[slot], &mut sc.cache);
+            send_hidden(&sc, &x, next_peer);
+            next = token_rx.recv().expect("decode token");
+            generated += 1;
+            stalled = !emit_text(&tokenizer, &req, next, &mut decode_buf);
+        };
+
+        let tail = tokenizer.flush_decode_stream(&mut decode_buf);
+        if !tail.is_empty() {
+            let _ = req.sender.try_send(GenerateEvent::Token {
+                id: next,
+                text: tail,
+            });
+        }
+        let _ = req.sender.try_send(GenerateEvent::Done {
+            finish_reason: finish.to_string(),
+        });
+
+        broadcast(&cmd_txs, PpCommand::Free(slot));
+        sc.cache.free_sequence(slot);
+        wait_acks(&ack_rx);
+    }
+
+    broadcast(&cmd_txs, PpCommand::Shutdown);
+}
+
+/// Stream a token's decoded text to the client (EOS contributes no text).
+/// Returns false if the send would block (client too slow) or the client is
+/// gone — the caller stops generating so the coordinator thread is free to
+/// admit the next request instead of blocking on one slow consumer.
+fn emit_text(
+    tokenizer: &Tokenizer,
+    req: &GenerateRequest,
+    token_id: u32,
+    buf: &mut Vec<u8>,
+) -> bool {
+    if tokenizer.is_eos(token_id) {
+        return true;
+    }
+    let text = tokenizer.decode_token_stream(token_id, buf);
+    if !text.is_empty() {
+        return req
+            .sender
+            .try_send(GenerateEvent::Token { id: token_id, text })
+            .is_ok();
+    }
+    true
+}

crates/xserv-server/src/tp_engine.rs

@@ -0,0 +1,366 @@
+//! Tensor-parallel inference engine for the HTTP server.
+//!
+//! Serial coordinator model: one rank-0 coordinator thread (the caller) drives
+//! generation and owns the scheduler; ranks 1..world are worker threads. For
+//! each step the coordinator broadcasts a command (Register/Prefill/Decode/Free)
+//! to the workers and runs the same op on its own shard; the per-layer NCCL
+//! AllReduces keep all ranks in lockstep. Only the coordinator samples — the
+//! chosen token is carried in the next Decode command, so this is correct for
+//! both greedy and stochastic sampling.
+//!
+//! Requests are processed one at a time (sufficient for the quality benchmark,
+//! which issues serial requests). Continuous batching across ranks is future
+//! work; the single-GPU `Engine` still handles TP=1.
+
+use std::path::{Path, PathBuf};
+use std::sync::Arc;
+use std::sync::mpsc;
+use std::thread;
+
+use xserv_distributed::{TpContext, UniqueId};
+use xserv_model::loader;
+use xserv_model::{
+    BLOCK_SIZE, GptOss, GraphedGptOssDecoder, ModelConfig, PagedKVCache, Qwen3, sample,
+    sample_greedy_penalized,
+};
+use xserv_tensor::{DType, Device, Tensor};
+use xserv_tokenizer::Tokenizer;
+
+use crate::engine::{GenerateEvent, GenerateRequest};
+
+#[derive(Clone)]
+enum TpCommand {
+    Register(usize),
+    Free(usize),
+    Prefill {
+        tokens: Vec<u32>,
+        slot: usize,
+    },
+    Decode {
+        tokens: Vec<u32>,
+        positions: Vec<usize>,
+        slots: Vec<usize>,
+    },
+    Shutdown,
+}
+
+enum TpModel {
+    Qwen3(Qwen3),
+    GptOss(GptOss),
+}
+
+impl TpModel {
+    fn forward_prefill_paged(
+        &self,
+        tokens: &[u32],
+        slot: usize,
+        cache: &mut PagedKVCache,
+    ) -> Tensor {
+        match self {
+            TpModel::Qwen3(m) => m.forward_prefill_paged(tokens, slot, cache),
+            TpModel::GptOss(m) => m.forward_prefill_paged(tokens, slot, cache),
+        }
+    }
+
+    fn forward_decode_paged(
+        &self,
+        tokens: &[u32],
+        positions: &[usize],
+        slots: &[usize],
+        cache: &mut PagedKVCache,
+    ) -> Tensor {
+        match self {
+            TpModel::Qwen3(m) => m.forward_decode_paged(tokens, positions, slots, cache),
+            TpModel::GptOss(m) => m.forward_decode_paged(tokens, positions, slots, cache),
+        }
+    }
+}
+
+struct RankCtx {
+    model: TpModel,
+    cache: PagedKVCache,
+    decoder: GraphedGptOssDecoder,
+}
+
+/// Decode one step: gpt-oss batch=1 goes through the CUDA-graph decoder
+/// (lazy capture, replay thereafter); everything else runs eager.
+fn rank_decode(rc: &mut RankCtx, tokens: &[u32], positions: &[usize], slots: &[usize]) -> Tensor {
+    match &rc.model {
+        TpModel::GptOss(m) => rc
+            .decoder
+            .decode(m, tokens, positions, slots, &mut rc.cache),
+        TpModel::Qwen3(_) => rc
+            .model
+            .forward_decode_paged(tokens, positions, slots, &mut rc.cache),
+    }
+}
+
+fn build_rank(
+    model_dir: &Path,
+    config: &ModelConfig,
+    rank: usize,
+    world: usize,
+    device: u32,
+    max_seq_len: usize,
+    tp: Option<Arc<TpContext>>,
+) -> RankCtx {
+    let weights = loader::load_model_dir(model_dir, Device::Cpu);
+    let model = if config.is_moe() {
+        TpModel::GptOss(GptOss::from_weights_tp(
+            config.clone(),
+            weights,
+            rank,
+            world,
+            device,
+            tp,
+        ))
+    } else {
+        TpModel::Qwen3(Qwen3::from_weights_tp(
+            config.clone(),
+            weights,
+            rank,
+            world,
+            device,
+            tp,
+        ))
+    };
+    let local_kv = config.num_kv_heads() / world;
+    let max_blocks_per_seq = (max_seq_len + BLOCK_SIZE - 1) / BLOCK_SIZE;
+    let total_blocks = max_blocks_per_seq + 8;
+    let cache = PagedKVCache::new_tp(
+        config,
+        local_kv,
+        total_blocks,
+        0,
+        4,
+        max_blocks_per_seq,
+        DType::BF16,
+        device,
+    );
+    RankCtx {
+        model,
+        cache,
+        decoder: GraphedGptOssDecoder::new(),
+    }
+}
+
+fn worker_loop(
+    rank: usize,
+    world: usize,
+    id: UniqueId,
+    model_dir: PathBuf,
+    config: ModelConfig,
+    max_seq_len: usize,
+    cmd_rx: mpsc::Receiver<TpCommand>,
+    ack_tx: mpsc::Sender<()>,
+) {
+    let tp = Arc::new(TpContext::init(rank, world, id, rank as u32));
+    let mut rc = build_rank(
+        &model_dir,
+        &config,
+        rank,
+        world,
+        rank as u32,
+        max_seq_len,
+        Some(tp),
+    );
+    while let Ok(cmd) = cmd_rx.recv() {
+        match cmd {
+            TpCommand::Register(slot) => {
+                let _ = rc.cache.register_sequence(slot);
+            }
+            TpCommand::Free(slot) => rc.cache.free_sequence(slot),
+            TpCommand::Prefill { tokens, slot } => {
+                let _ = rc.model.forward_prefill_paged(&tokens, slot, &mut rc.cache);
+            }
+            TpCommand::Decode {
+                tokens,
+                positions,
+                slots,
+            } => {
+                let _ = rank_decode(&mut rc, &tokens, &positions, &slots);
+            }
+            TpCommand::Shutdown => {
+                let _ = ack_tx.send(());
+                break;
+            }
+        }
+        let _ = ack_tx.send(());
+    }
+}
+
+/// Run the TP coordinator (rank 0) on the calling thread. Spawns worker ranks
+/// internally and consumes generation requests from `rx`.
+pub fn run_tp(
+    model_dir: &Path,
+    world: usize,
+    max_seq_len: usize,
+    rx: mpsc::Receiver<GenerateRequest>,
+) {
+    // world=1 is a valid single-rank configuration (gpt-oss has no
+    // single-GPU engine path; NCCL init and all_reduce no-op at world=1).
+    assert!(world >= 1, "run_tp requires world >= 1");
+    let config = ModelConfig::from_file(&model_dir.join("config.json"));
+    assert!(
+        config.num_kv_heads() % world == 0,
+        "num_kv_heads {} not divisible by tp {world}",
+        config.num_kv_heads()
+    );
+    let tokenizer = Tokenizer::from_file(&model_dir.join("tokenizer.json"));
+    let id = xserv_distributed::get_unique_id();
+
+    // Spawn worker ranks 1..world.
+    let (ack_tx, ack_rx) = mpsc::channel::<()>();
+    let mut cmd_txs: Vec<mpsc::Sender<TpCommand>> = Vec::new();
+    for rank in 1..world {
+        let (ctx_tx, ctx_rx) = mpsc::channel::<TpCommand>();
+        cmd_txs.push(ctx_tx);
+        let ack_tx = ack_tx.clone();
+        let model_dir = model_dir.to_path_buf();
+        let config = config.clone();
+        thread::spawn(move || {
+            worker_loop(
+                rank,
+                world,
+                id,
+                model_dir,
+                config,
+                max_seq_len,
+                ctx_rx,
+                ack_tx,
+            );
+        });
+    }
+
+    // Rank 0 (this thread).
+    let tp = Arc::new(TpContext::init(0, world, id, 0));
+    let mut rc = build_rank(model_dir, &config, 0, world, 0, max_seq_len, Some(tp));
+    eprintln!("[tp-engine] ready (tp={world}, max_seq_len={max_seq_len})");
+
+    // Optional repetition penalty to break greedy repetition loops (reasoning
+    // models loop under pure greedy when numerics diverge from the reference).
+    // Off by default; XSERV_REP_PENALTY>1 enables it over the last
+    // XSERV_REP_WINDOW generated tokens. Applied only on the greedy path.
+    let rep_penalty: f32 = std::env::var("XSERV_REP_PENALTY")
+        .ok()
+        .and_then(|s| s.parse().ok())
+        .unwrap_or(1.0);
+    let rep_window: usize = std::env::var("XSERV_REP_WINDOW")
+        .ok()
+        .and_then(|s| s.parse().ok())
+        .unwrap_or(128);
+    let pick = |logits: &Tensor, sp: &xserv_model::SamplingParams, history: &[u32]| -> u32 {
+        if rep_penalty > 1.0 && sp.temperature == 0.0 {
+            let start = history.len().saturating_sub(rep_window);
+            sample_greedy_penalized(logits, &history[start..], rep_penalty)
+        } else {
+            sample(logits, sp)
+        }
+    };
+
+    let n_workers = world - 1;
+    let broadcast = |txs: &[mpsc::Sender<TpCommand>], cmd: TpCommand| {
+        for t in txs {
+            let _ = t.send(cmd.clone());
+        }
+    };
+    let wait_acks = |rx: &mpsc::Receiver<()>| {
+        for _ in 0..n_workers {
+            let _ = rx.recv();
+        }
+    };
+
+    let slot = 0usize;
+    while let Ok(req) = rx.recv() {
+        broadcast(&cmd_txs, TpCommand::Register(slot));
+        rc.cache.register_sequence(slot).expect("register slot");
+        wait_acks(&ack_rx);
+
+        // Prefill.
+        broadcast(
+            &cmd_txs,
+            TpCommand::Prefill {
+                tokens: req.prompt_tokens.clone(),
+                slot,
+            },
+        );
+        let logits = rc
+            .model
+            .forward_prefill_paged(&req.prompt_tokens, slot, &mut rc.cache);
+        wait_acks(&ack_rx);
+        let mut gen_ids: Vec<u32> = Vec::new();
+        let mut next = pick(&logits, &req.sampling, &gen_ids);
+        gen_ids.push(next);
+
+        let mut decode_buf: Vec<u8> = Vec::new();
+        let mut generated = 1usize;
+        let mut stalled = !emit_text(&tokenizer, &req, next, &mut decode_buf);
+
+        let finish = loop {
+            if stalled {
+                break "error";
+            }
+            if tokenizer.is_eos(next) {
+                break "stop";
+            }
+            if generated >= req.max_tokens {
+                break "length";
+            }
+            let pos = rc.cache.seq_len(slot);
+            broadcast(
+                &cmd_txs,
+                TpCommand::Decode {
+                    tokens: vec![next],
+                    positions: vec![pos],
+                    slots: vec![slot],
+                },
+            );
+            let logits = rank_decode(&mut rc, &[next], &[pos], &[slot]);
+            wait_acks(&ack_rx);
+            next = pick(&logits, &req.sampling, &gen_ids);
+            gen_ids.push(next);
+            generated += 1;
+            stalled = !emit_text(&tokenizer, &req, next, &mut decode_buf);
+        };
+
+        let tail = tokenizer.flush_decode_stream(&mut decode_buf);
+        if !tail.is_empty() {
+            let _ = req.sender.try_send(GenerateEvent::Token {
+                id: next,
+                text: tail,
+            });
+        }
+        let _ = req.sender.try_send(GenerateEvent::Done {
+            finish_reason: finish.to_string(),
+        });
+
+        broadcast(&cmd_txs, TpCommand::Free(slot));
+        rc.cache.free_sequence(slot);
+        wait_acks(&ack_rx);
+    }
+
+    broadcast(&cmd_txs, TpCommand::Shutdown);
+}
+
+/// Stream a token's decoded text to the client (EOS contributes no text).
+/// Returns false if the send would block (client too slow) or the client is
+/// gone — the caller stops generating so the serial coordinator thread is free
+/// to admit the next request instead of blocking on one slow consumer.
+fn emit_text(
+    tokenizer: &Tokenizer,
+    req: &GenerateRequest,
+    token_id: u32,
+    buf: &mut Vec<u8>,
+) -> bool {
+    if tokenizer.is_eos(token_id) {
+        return true;
+    }
+    let text = tokenizer.decode_token_stream(token_id, buf);
+    if !text.is_empty() {
+        return req
+            .sender
+            .try_send(GenerateEvent::Token { id: token_id, text })
+            .is_ok();
+    }
+    true
+}

crates/xserv-tensor/src/dtype.rs

@@ -5,6 +5,7 @@ pub enum DType {
     F32,
     F16,
     BF16,
+    FP8E4M3,
 }
 
 impl DType {
@@ -13,6 +14,7 @@ impl DType {
             DType::F32 => 4,
             DType::F16 => 2,
             DType::BF16 => 2,
+            DType::FP8E4M3 => 1,
         }
     }
 
@@ -21,6 +23,7 @@ impl DType {
             DType::F32 => "f32",
             DType::F16 => "f16",
             DType::BF16 => "bf16",
+            DType::FP8E4M3 => "fp8e4m3",
         }
     }
 }
@@ -40,18 +43,30 @@ pub trait TensorDType: Copy + Send + Sync + 'static {
 
 impl TensorDType for f32 {
     const DTYPE: DType = DType::F32;
-    fn to_f64(self) -> f64 { self as f64 }
-    fn from_f64(v: f64) -> Self { v as f32 }
+    fn to_f64(self) -> f64 {
+        self as f64
+    }
+    fn from_f64(v: f64) -> Self {
+        v as f32
+    }
 }
 
 impl TensorDType for f16 {
     const DTYPE: DType = DType::F16;
-    fn to_f64(self) -> f64 { self.to_f32() as f64 }
-    fn from_f64(v: f64) -> Self { f16::from_f32(v as f32) }
+    fn to_f64(self) -> f64 {
+        self.to_f32() as f64
+    }
+    fn from_f64(v: f64) -> Self {
+        f16::from_f32(v as f32)
+    }
 }
 
 impl TensorDType for bf16 {
     const DTYPE: DType = DType::BF16;
-    fn to_f64(self) -> f64 { self.to_f32() as f64 }
-    fn from_f64(v: f64) -> Self { bf16::from_f32(v as f32) }
+    fn to_f64(self) -> f64 {
+        self.to_f32() as f64
+    }
+    fn from_f64(v: f64) -> Self {
+        bf16::from_f32(v as f32)
+    }
 }

crates/xserv-tensor/src/lib.rs

@@ -4,5 +4,6 @@ pub mod storage;
 pub mod tensor;
 
 pub use dtype::{DType, TensorDType};
-pub use storage::Device;
-pub use tensor::Tensor;
+pub use shape::Dims;
+pub use storage::{Device, Storage};
+pub use tensor::{Tensor, register_gpu_contiguous};

crates/xserv-tensor/src/shape.rs

@@ -18,12 +18,21 @@ pub fn contiguous_strides(shape: &[usize]) -> Dims {
 }
 
 /// Check if the given strides represent contiguous (row-major) layout for the shape.
+/// A stride mismatch on a dimension of size 1 is allowed because that
+/// dimension is never stepped.
 pub fn is_contiguous(shape: &[usize], strides: &[usize]) -> bool {
     if shape.is_empty() {
         return true;
     }
-    let expected = contiguous_strides(shape);
-    strides == expected.as_slice()
+    let ndim = shape.len();
+    let mut expected_stride = 1usize;
+    for d in (0..ndim).rev() {
+        if shape[d] != 1 && strides[d] != expected_stride {
+            return false;
+        }
+        expected_stride *= shape[d];
+    }
+    true
 }
 
 /// Total number of elements given a shape.
@@ -37,8 +46,16 @@ pub fn broadcast_shape(a: &[usize], b: &[usize]) -> Option<Dims> {
     let ndim = a.len().max(b.len());
     let mut result = SmallVec::with_capacity(ndim);
     for i in 0..ndim {
-        let da = if i < ndim - a.len() { 1 } else { a[i - (ndim - a.len())] };
-        let db = if i < ndim - b.len() { 1 } else { b[i - (ndim - b.len())] };
+        let da = if i < ndim - a.len() {
+            1
+        } else {
+            a[i - (ndim - a.len())]
+        };
+        let db = if i < ndim - b.len() {
+            1
+        } else {
+            b[i - (ndim - b.len())]
+        };
         if da == db {
             result.push(da);
         } else if da == 1 {
@@ -91,8 +108,14 @@ mod tests {
 
     #[test]
     fn test_broadcast_shape() {
-        assert_eq!(broadcast_shape(&[3, 1], &[1, 4]).unwrap().as_slice(), &[3, 4]);
-        assert_eq!(broadcast_shape(&[2, 3, 4], &[4]).unwrap().as_slice(), &[2, 3, 4]);
+        assert_eq!(
+            broadcast_shape(&[3, 1], &[1, 4]).unwrap().as_slice(),
+            &[3, 4]
+        );
+        assert_eq!(
+            broadcast_shape(&[2, 3, 4], &[4]).unwrap().as_slice(),
+            &[2, 3, 4]
+        );
         assert_eq!(broadcast_shape(&[1], &[5, 3]).unwrap().as_slice(), &[5, 3]);
         assert!(broadcast_shape(&[3], &[4]).is_none());
     }
@@ -100,6 +123,9 @@ mod tests {
     #[test]
     fn test_broadcast_strides() {
         // [3,1] with strides [1,1] broadcast to [3,4]
-        assert_eq!(broadcast_strides(&[3, 1], &[1, 1], &[3, 4]).as_slice(), &[1, 0]);
+        assert_eq!(
+            broadcast_strides(&[3, 1], &[1, 1], &[3, 4]).as_slice(),
+            &[1, 0]
+        );
     }
 }

crates/xserv-tensor/src/storage.rs

@@ -3,7 +3,7 @@ use xserv_cuda::{GpuBuffer, Result as CudaResult};
 
 enum StorageInner {
     Cpu { data: Vec<u8> },
-    Cuda { buffer: GpuBuffer },
+    Cuda { buffer: GpuBuffer, device: u32 },
 }
 
 /// Reference-counted storage for tensor data. Multiple tensors can share
@@ -31,21 +31,21 @@ impl Storage {
         Self(Arc::new(StorageInner::Cpu { data }))
     }
 
-    pub fn cuda(buffer: GpuBuffer) -> Self {
-        Self(Arc::new(StorageInner::Cuda { buffer }))
+    pub fn cuda(buffer: GpuBuffer, device: u32) -> Self {
+        Self(Arc::new(StorageInner::Cuda { buffer, device }))
     }
 
     pub fn device(&self) -> Device {
         match self.0.as_ref() {
             StorageInner::Cpu { .. } => Device::Cpu,
-            StorageInner::Cuda { .. } => Device::Cuda(0),
+            StorageInner::Cuda { device, .. } => Device::Cuda(*device),
         }
     }
 
     pub fn len_bytes(&self) -> usize {
         match self.0.as_ref() {
             StorageInner::Cpu { data } => data.len(),
-            StorageInner::Cuda { buffer } => buffer.len(),
+            StorageInner::Cuda { buffer, .. } => buffer.len(),
         }
     }
 
@@ -59,7 +59,7 @@ impl Storage {
 
     pub fn gpu_buffer(&self) -> &GpuBuffer {
         match self.0.as_ref() {
-            StorageInner::Cuda { buffer } => buffer,
+            StorageInner::Cuda { buffer, .. } => buffer,
             StorageInner::Cpu { .. } => panic!("cannot access CPU storage as GPU buffer"),
         }
     }
@@ -71,11 +71,11 @@ impl Storage {
             return Ok(self.clone());
         }
         match (current, target) {
-            (Device::Cpu, Device::Cuda(_dev)) => {
+            (Device::Cpu, Device::Cuda(dev)) => {
                 let cpu_data = self.as_cpu_bytes();
-                let mut buf = GpuBuffer::alloc(cpu_data.len())?;
+                let mut buf = xserv_cuda::allocator::cached_alloc(cpu_data.len())?;
                 buf.copy_from_host(cpu_data)?;
-                Ok(Storage::cuda(buf))
+                Ok(Storage::cuda(buf, dev))
             }
             (Device::Cuda(_), Device::Cpu) => {
                 let gpu_buf = self.gpu_buffer();
@@ -83,11 +83,11 @@ impl Storage {
                 gpu_buf.copy_to_host(&mut data)?;
                 Ok(Storage::cpu(data))
             }
-            (Device::Cuda(_), Device::Cuda(_)) => {
+            (Device::Cuda(_), Device::Cuda(dev)) => {
                 let src = self.gpu_buffer();
-                let mut dst = GpuBuffer::alloc(src.len())?;
+                let mut dst = xserv_cuda::allocator::cached_alloc(src.len())?;
                 dst.copy_from_device(src)?;
-                Ok(Storage::cuda(dst))
+                Ok(Storage::cuda(dst, dev))
             }
             _ => unreachable!(),
         }
@@ -97,10 +97,10 @@ impl Storage {
     pub fn deep_copy(&self) -> CudaResult<Self> {
         match self.0.as_ref() {
             StorageInner::Cpu { data } => Ok(Storage::cpu(data.clone())),
-            StorageInner::Cuda { buffer } => {
-                let mut dst = GpuBuffer::alloc(buffer.len())?;
+            StorageInner::Cuda { buffer, device } => {
+                let mut dst = xserv_cuda::allocator::cached_alloc(buffer.len())?;
                 dst.copy_from_device(buffer)?;
-                Ok(Storage::cuda(dst))
+                Ok(Storage::cuda(dst, *device))
             }
         }
     }
@@ -109,10 +109,24 @@ impl Storage {
     pub fn zeros(len_bytes: usize, device: Device) -> CudaResult<Self> {
         match device {
             Device::Cpu => Ok(Storage::cpu(vec![0u8; len_bytes])),
-            Device::Cuda(_) => {
-                let mut buf = GpuBuffer::alloc(len_bytes)?;
+            Device::Cuda(dev) => {
+                let mut buf = xserv_cuda::allocator::cached_alloc(len_bytes)?;
                 buf.zero()?;
-                Ok(Storage::cuda(buf))
+                Ok(Storage::cuda(buf, dev))
+            }
+        }
+    }
+
+    /// Allocate storage **without zeroing** on the given device.
+    /// The buffer may contain stale data from the caching allocator's pool.
+    /// Only use when the caller guarantees the kernel will fully overwrite
+    /// every element before any read.
+    pub fn empty(len_bytes: usize, device: Device) -> CudaResult<Self> {
+        match device {
+            Device::Cpu => Ok(Storage::cpu(vec![0u8; len_bytes])), // CPU still zeros (cheap)
+            Device::Cuda(dev) => {
+                let buf = xserv_cuda::allocator::cached_alloc(len_bytes)?;
+                Ok(Storage::cuda(buf, dev))
             }
         }
     }

crates/xserv-tensor/src/tensor.rs

@@ -1,7 +1,21 @@
+use std::sync::OnceLock;
+
 use crate::dtype::{DType, TensorDType};
 use crate::shape::{self, Dims};
 use crate::storage::{Device, Storage};
 
+/// Global hook for GPU strided-to-contiguous copy.
+/// Set by `xserv-kernels` (or any crate that provides a GPU kernel) via
+/// `register_gpu_contiguous`. When set, `contiguous()` on a non-contiguous
+/// GPU tensor calls this instead of doing a CPU round-trip.
+static GPU_CONTIGUOUS_FN: OnceLock<fn(&Tensor) -> Tensor> = OnceLock::new();
+
+/// Register a function that makes a non-contiguous GPU tensor contiguous.
+/// Intended to be called once by the kernel crate at startup.
+pub fn register_gpu_contiguous(f: fn(&Tensor) -> Tensor) {
+    let _ = GPU_CONTIGUOUS_FN.set(f);
+}
+
 /// Multi-dimensional array with CPU or GPU storage.
 ///
 /// Tensors support view semantics: transpose, slice, etc. share
@@ -18,6 +32,23 @@ pub struct Tensor {
 impl Tensor {
     // --- Creation ---
 
+    /// Create a tensor from raw components (for advanced use like GPU KV cache).
+    pub fn from_storage(
+        storage: Storage,
+        shape: Dims,
+        strides: Dims,
+        offset: usize,
+        dtype: DType,
+    ) -> Self {
+        Self {
+            storage,
+            shape,
+            strides,
+            offset,
+            dtype,
+        }
+    }
+
     pub fn from_slice<T: TensorDType>(data: &[T], shape: &[usize]) -> Self {
         let numel: usize = shape.iter().product();
         assert_eq!(data.len(), numel, "data length mismatch with shape");
@@ -33,6 +64,28 @@ impl Tensor {
         }
     }
 
+    /// Create a tensor from raw bytes. Used for dtypes without a Rust type
+    /// (e.g. FP8 E4M3) where we store the bit pattern as-is.
+    pub fn from_raw_bytes(data: &[u8], shape: &[usize], dtype: DType) -> Self {
+        let numel: usize = shape.iter().product();
+        assert_eq!(
+            data.len(),
+            numel * dtype.size_bytes(),
+            "raw bytes length {} != expected {} (numel={} * elem_size={})",
+            data.len(),
+            numel * dtype.size_bytes(),
+            numel,
+            dtype.size_bytes()
+        );
+        Self {
+            storage: Storage::cpu(data.to_vec()),
+            shape: Dims::from_slice(shape),
+            strides: shape::contiguous_strides(shape),
+            offset: 0,
+            dtype,
+        }
+    }
+
     pub fn zeros(shape: &[usize], dtype: DType, device: Device) -> Self {
         let numel = shape::num_elements(shape);
         let len_bytes = numel * dtype.size_bytes();
@@ -46,25 +99,56 @@ impl Tensor {
         }
     }
 
+    /// Allocate a tensor **without zeroing** the backing memory.
+    /// The buffer may contain stale data. Only use when the calling kernel
+    /// will fully overwrite every element before any read.
+    pub fn empty(shape: &[usize], dtype: DType, device: Device) -> Self {
+        let numel = shape::num_elements(shape);
+        let len_bytes = numel * dtype.size_bytes();
+        let storage = Storage::empty(len_bytes, device).expect("alloc failed");
+        Self {
+            storage,
+            shape: Dims::from_slice(shape),
+            strides: shape::contiguous_strides(shape),
+            offset: 0,
+            dtype,
+        }
+    }
+
     pub fn ones(shape: &[usize], dtype: DType) -> Self {
         let numel = shape::num_elements(shape);
         match dtype {
             DType::F32 => Self::from_slice(&vec![1.0f32; numel], shape),
             DType::F16 => Self::from_slice(&vec![half::f16::from_f32(1.0); numel], shape),
             DType::BF16 => Self::from_slice(&vec![half::bf16::from_f32(1.0); numel], shape),
+            DType::FP8E4M3 => panic!("ones() not supported for FP8E4M3"),
         }
     }
 
     // --- Properties ---
 
-    pub fn shape(&self) -> &[usize] { &self.shape }
-    pub fn strides(&self) -> &[usize] { &self.strides }
-    pub fn dtype(&self) -> DType { self.dtype }
-    pub fn ndim(&self) -> usize { self.shape.len() }
-    pub fn numel(&self) -> usize { shape::num_elements(&self.shape) }
-    pub fn offset(&self) -> usize { self.offset }
+    pub fn shape(&self) -> &[usize] {
+        &self.shape
+    }
+    pub fn strides(&self) -> &[usize] {
+        &self.strides
+    }
+    pub fn dtype(&self) -> DType {
+        self.dtype
+    }
+    pub fn ndim(&self) -> usize {
+        self.shape.len()
+    }
+    pub fn numel(&self) -> usize {
+        shape::num_elements(&self.shape)
+    }
+    pub fn offset(&self) -> usize {
+        self.offset
+    }
 
-    pub fn device(&self) -> Device { self.storage.device() }
+    pub fn device(&self) -> Device {
+        self.storage.device()
+    }
 
     pub fn is_contiguous(&self) -> bool {
         shape::is_contiguous(&self.shape, &self.strides)
@@ -85,6 +169,21 @@ impl Tensor {
         }
     }
 
+    /// Zero-copy slice along `dim`: keeps elements `[start, start+len)`.
+    pub fn narrow(&self, dim: usize, start: usize, len: usize) -> Self {
+        assert!(dim < self.ndim());
+        assert!(start + len <= self.shape[dim], "narrow out of bounds");
+        let mut new_shape = self.shape.clone();
+        new_shape[dim] = len;
+        Self {
+            storage: self.storage.clone(),
+            shape: new_shape,
+            strides: self.strides.clone(),
+            offset: self.offset + start * self.strides[dim],
+            dtype: self.dtype,
+        }
+    }
+
     pub fn transpose(&self, dim0: usize, dim1: usize) -> Self {
         assert!(dim0 < self.ndim() && dim1 < self.ndim());
         let mut new_shape = self.shape.clone();
@@ -118,10 +217,19 @@ impl Tensor {
     pub fn unsqueeze(&self, dim: usize) -> Self {
         assert!(dim <= self.ndim());
         let mut new_shape = self.shape.clone();
-        let mut new_strides = self.strides.clone();
         new_shape.insert(dim, 1);
-        let stride_val = if dim < self.strides.len() { self.strides[dim] } else { 1 };
-        new_strides.insert(dim, stride_val);
+        let new_strides = if self.is_contiguous() {
+            shape::contiguous_strides(&new_shape)
+        } else {
+            let mut s = self.strides.clone();
+            let stride_val = if dim < self.strides.len() {
+                self.strides[dim]
+            } else {
+                1
+            };
+            s.insert(dim, stride_val);
+            s
+        };
         Self {
             storage: self.storage.clone(),
             shape: new_shape,
@@ -137,8 +245,16 @@ impl Tensor {
         if self.is_contiguous() {
             return self.clone();
         }
-        // Copy to contiguous layout on CPU
-        assert_eq!(self.device(), Device::Cpu, "contiguous() on GPU not yet supported");
+        // For GPU tensors: use the registered GPU kernel if available,
+        // otherwise fall back to CPU round-trip.
+        if matches!(self.device(), Device::Cuda(_)) {
+            if let Some(gpu_fn) = GPU_CONTIGUOUS_FN.get() {
+                return gpu_fn(self);
+            }
+            let cpu = self.to_device(Device::Cpu);
+            let contig = cpu.contiguous();
+            return contig.to_device(self.device());
+        }
         let numel = self.numel();
         let elem_size = self.dtype.size_bytes();
         let src_bytes = self.storage.as_cpu_bytes();
@@ -147,7 +263,12 @@ impl Tensor {
         let ndim = self.ndim();
         let mut idx = vec![0usize; ndim];
         for flat in 0..numel {
-            let src_offset = self.offset + idx.iter().zip(self.strides.iter()).map(|(i, s)| i * s).sum::<usize>();
+            let src_offset = self.offset
+                + idx
+                    .iter()
+                    .zip(self.strides.iter())
+                    .map(|(i, s)| i * s)
+                    .sum::<usize>();
             let src_byte_offset = src_offset * elem_size;
             let dst_byte_offset = flat * elem_size;
             dst[dst_byte_offset..dst_byte_offset + elem_size]
@@ -173,17 +294,21 @@ impl Tensor {
     // --- Device transfer ---
 
     pub fn to_device(&self, device: Device) -> Self {
-        let t = if self.is_contiguous() { self.clone() } else { self.contiguous() };
-        if t.device() == device {
-            return t;
+        if self.device() == device {
+            return self.clone();
         }
-        let new_storage = t.storage.to_device(device).expect("device transfer failed");
+        // Transfer the raw storage (preserving strides/offset).
+        // Non-contiguous layout is preserved — the user can call contiguous() after.
+        let new_storage = self
+            .storage
+            .to_device(device)
+            .expect("device transfer failed");
         Self {
             storage: new_storage,
-            shape: t.shape,
-            strides: t.strides,
-            offset: 0,
-            dtype: t.dtype,
+            shape: self.shape.clone(),
+            strides: self.strides.clone(),
+            offset: self.offset,
+            dtype: self.dtype,
         }
     }
 
@@ -201,6 +326,17 @@ impl Tensor {
         unsafe { std::slice::from_raw_parts(bytes[start..].as_ptr() as *const T, len) }
     }
 
+    /// Raw byte access for dtypes without a Rust type (e.g. FP8).
+    pub fn as_raw_bytes(&self) -> &[u8] {
+        assert!(self.is_contiguous(), "as_raw_bytes requires contiguous");
+        assert_eq!(self.device(), Device::Cpu, "as_raw_bytes requires CPU");
+        let bytes = self.storage.as_cpu_bytes();
+        let elem_size = self.dtype.size_bytes();
+        let start = self.offset * elem_size;
+        let len = self.numel() * elem_size;
+        &bytes[start..start + len]
+    }
+
     /// Raw pointer to storage start (for GPU kernel launch).
     pub fn data_ptr(&self) -> *const u8 {
         match self.device() {
@@ -215,14 +351,75 @@ impl Tensor {
         }
     }
 
-    pub fn storage(&self) -> &Storage { &self.storage }
+    pub fn storage(&self) -> &Storage {
+        &self.storage
+    }
 }
 
 impl std::fmt::Debug for Tensor {
     fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
         write!(
-            f, "Tensor(shape={:?}, dtype={}, device={}, contiguous={})",
-            self.shape.as_slice(), self.dtype, self.device(), self.is_contiguous()
+            f,
+            "Tensor(shape={:?}, dtype={}, device={}, contiguous={})",
+            self.shape.as_slice(),
+            self.dtype,
+            self.device(),
+            self.is_contiguous()
         )
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn contiguous_2d() -> Tensor {
+        Tensor::from_slice(&[1.0f32; 12], &[3, 4])
+    }
+
+    #[test]
+    fn unsqueeze_dim0_contiguous() {
+        let t = contiguous_2d();
+        let u = t.unsqueeze(0);
+        assert_eq!(u.shape(), &[1, 3, 4]);
+        assert!(u.is_contiguous());
+        assert_eq!(u.strides(), &[12, 4, 1]);
+    }
+
+    #[test]
+    fn unsqueeze_dim1_contiguous() {
+        let t = contiguous_2d();
+        let u = t.unsqueeze(1);
+        assert_eq!(u.shape(), &[3, 1, 4]);
+        assert!(u.is_contiguous());
+        assert_eq!(u.strides(), &[4, 4, 1]);
+    }
+
+    #[test]
+    fn unsqueeze_dim2_contiguous() {
+        let t = contiguous_2d();
+        let u = t.unsqueeze(2);
+        assert_eq!(u.shape(), &[3, 4, 1]);
+        assert!(u.is_contiguous());
+        assert_eq!(u.strides(), &[4, 1, 1]);
+    }
+
+    #[test]
+    fn unsqueeze_noncontiguous() {
+        // Transpose makes [3,4] into [4,3] with strides [1,4] (non-contiguous)
+        let t = contiguous_2d().transpose(0, 1);
+        assert!(!t.is_contiguous());
+        let u = t.unsqueeze(0);
+        assert_eq!(u.shape(), &[1, 4, 3]);
+        // Non-contiguous path: stride_val copied from strides[0]=1
+        assert_eq!(u.strides(), &[1, 1, 4]);
+    }
+
+    #[test]
+    fn unsqueeze_squeeze_roundtrip() {
+        let t = contiguous_2d();
+        let u = t.unsqueeze(1).squeeze(1);
+        assert_eq!(u.shape(), t.shape());
+        assert!(u.is_contiguous());
+    }
+}

crates/xserv-tensor/tests/integration.rs

@@ -32,7 +32,11 @@ fn test_zeros_and_ones() {
 
 #[test]
 fn test_bf16_tensor() {
-    let data: Vec<bf16> = vec![bf16::from_f32(1.0), bf16::from_f32(2.5), bf16::from_f32(-3.0)];
+    let data: Vec<bf16> = vec![
+        bf16::from_f32(1.0),
+        bf16::from_f32(2.5),
+        bf16::from_f32(-3.0),
+    ];
     let t = Tensor::from_slice(&data, &[3]);
     assert_eq!(t.dtype(), DType::BF16);
     let out = t.as_slice::<bf16>();

crates/xserv-tokenizer/Cargo.toml

@@ -0,0 +1,9 @@
+[package]
+name = "xserv-tokenizer"
+version.workspace = true
+edition.workspace = true
+
+[dependencies]
+serde.workspace = true
+serde_json.workspace = true
+regex.workspace = true

crates/xserv-tokenizer/src/bpe.rs

@@ -0,0 +1,452 @@
+use regex::Regex;
+use serde::Deserialize;
+use std::collections::HashMap;
+use std::path::Path;
+
+pub struct Tokenizer {
+    encoder: HashMap<Vec<u8>, u32>,
+    decoder: Vec<Vec<u8>>,
+    merge_ranks: HashMap<(u32, u32), usize>,
+    special_tokens: HashMap<String, u32>,
+    #[allow(dead_code)]
+    special_token_ids: HashMap<u32, String>,
+    pre_tokenize_re: Regex,
+    eos_token_id: Option<u32>,
+    eos_token_ids: Vec<u32>,
+    byte_fallback: bool,
+}
+
+#[derive(Deserialize)]
+struct TokenizerJson {
+    model: ModelSection,
+    #[serde(default)]
+    added_tokens: Vec<AddedToken>,
+    #[serde(default)]
+    pre_tokenizer: Option<PreTokenizerSection>,
+}
+
+#[derive(Deserialize)]
+struct PreTokenizerSection {
+    #[serde(default, rename = "type")]
+    kind: Option<String>,
+    #[serde(default)]
+    pattern: Option<PatternSpec>,
+    #[serde(default)]
+    pretokenizers: Option<Vec<PreTokenizerSection>>,
+}
+
+#[derive(Deserialize)]
+struct PatternSpec {
+    #[serde(rename = "Regex")]
+    regex: Option<String>,
+}
+
+#[derive(Deserialize)]
+struct ModelSection {
+    vocab: HashMap<String, u32>,
+    merges: Vec<MergeEntry>,
+    #[serde(default)]
+    byte_fallback: bool,
+}
+
+#[derive(Deserialize)]
+#[serde(untagged)]
+enum MergeEntry {
+    Str(String),
+    Pair(Vec<String>),
+}
+
+#[derive(Deserialize)]
+struct AddedToken {
+    id: u32,
+    content: String,
+    #[allow(dead_code)]
+    special: bool,
+}
+
+impl Tokenizer {
+    pub fn from_file(path: &Path) -> Self {
+        let data = std::fs::read_to_string(path)
+            .unwrap_or_else(|e| panic!("failed to read {}: {e}", path.display()));
+        let tj: TokenizerJson = serde_json::from_str(&data)
+            .unwrap_or_else(|e| panic!("failed to parse tokenizer.json: {e}"));
+
+        // Build encoder: token bytes → ID
+        // All HF tokenizers use GPT-2 byte-to-unicode mapping for vocab keys.
+        let mut encoder = HashMap::new();
+        for (token_str, &id) in &tj.model.vocab {
+            let bytes = token_str_to_bytes(token_str);
+            encoder.insert(bytes, id);
+        }
+
+        // Build decoder: ID → token bytes
+        let max_id = tj.model.vocab.values().copied().max().unwrap_or(0);
+        let added_max = tj.added_tokens.iter().map(|t| t.id).max().unwrap_or(0);
+        let vocab_size = (max_id.max(added_max) + 1) as usize;
+        let mut decoder = vec![vec![]; vocab_size];
+        for (token_str, &id) in &tj.model.vocab {
+            decoder[id as usize] = token_str_to_bytes(token_str);
+        }
+
+        // Parse merges (supports both "a b" string format and ["a", "b"] array format)
+        let byte_fallback = tj.model.byte_fallback;
+        let mut merge_ranks = HashMap::new();
+        for (rank, entry) in tj.model.merges.iter().enumerate() {
+            let (a_str, b_str) = match entry {
+                MergeEntry::Str(s) => {
+                    let parts: Vec<&str> = s.splitn(2, ' ').collect();
+                    if parts.len() != 2 {
+                        continue;
+                    }
+                    (parts[0].to_string(), parts[1].to_string())
+                }
+                MergeEntry::Pair(v) => {
+                    if v.len() != 2 {
+                        continue;
+                    }
+                    (v[0].clone(), v[1].clone())
+                }
+            };
+            let a_bytes = token_str_to_bytes(&a_str);
+            let b_bytes = token_str_to_bytes(&b_str);
+            if let (Some(&a_id), Some(&b_id)) = (encoder.get(&a_bytes), encoder.get(&b_bytes)) {
+                merge_ranks.insert((a_id, b_id), rank);
+            }
+        }
+
+        // Added tokens are matched as indivisible tokens by HF tokenizers,
+        // even when their `special` flag is false (for example Qwen3's
+        // <think> and </think> tokens).
+        let mut special_tokens = HashMap::new();
+        let mut special_token_ids = HashMap::new();
+        for at in &tj.added_tokens {
+            special_tokens.insert(at.content.clone(), at.id);
+            special_token_ids.insert(at.id, at.content.clone());
+            decoder.resize(decoder.len().max(at.id as usize + 1), vec![]);
+            decoder[at.id as usize] = at.content.as_bytes().to_vec();
+        }
+        // End-of-generation tokens, in priority order. Families differ:
+        // Qwen uses <|im_end|>, Llama <|end_of_text|>, GPT-2 <|endoftext|>.
+        // gpt-oss (harmony) ends the assistant turn with <|return|> and also
+        // treats <|call|> (tool call) and <|endoftext|> as terminators
+        // (see generation_config.json eos_token_id = [200002, 199999, 200012]).
+        let eos_names = [
+            "<|im_end|>",
+            "<|end_of_text|>",
+            "<|return|>",
+            "<|call|>",
+            "<|endoftext|>",
+        ];
+        let mut eos_token_ids: Vec<u32> = Vec::new();
+        for name in eos_names {
+            if let Some(&id) = special_tokens.get(name) {
+                if !eos_token_ids.contains(&id) {
+                    eos_token_ids.push(id);
+                }
+            }
+        }
+        let eos_token_id = eos_token_ids.first().copied();
+
+        // Pre-tokenization regex: prefer the model's own regex from tokenizer.json,
+        // fall back to GPT-2/Qwen heuristic if not present or unsupported.
+        let model_regex = tj.pre_tokenizer.as_ref().and_then(|pt| {
+            // Direct Split with regex
+            if pt.kind.as_deref() == Some("Split") {
+                return pt.pattern.as_ref().and_then(|p| p.regex.clone());
+            }
+            // Sequence → find the Split entry
+            if let Some(subs) = &pt.pretokenizers {
+                for sub in subs {
+                    if sub.kind.as_deref() == Some("Split") {
+                        if let Some(r) = sub.pattern.as_ref().and_then(|p| p.regex.clone()) {
+                            return Some(r);
+                        }
+                    }
+                }
+            }
+            None
+        });
+
+        let pre_tokenize_re = if let Some(ref pat) = model_regex {
+            // Strip unsupported lookahead (?!\S) — Rust regex doesn't support it.
+            // The lookahead only affects trailing-whitespace edge cases.
+            let cleaned = pat.replace(r"(?!\S)", "");
+            match Regex::new(&cleaned) {
+                Ok(re) => re,
+                Err(e) => {
+                    eprintln!("warning: model pre_tokenizer regex failed ({e}), using fallback");
+                    if byte_fallback {
+                        Regex::new(r"[\p{L}\p{N}]+|[^\s\p{L}\p{N}]|\s+").unwrap()
+                    } else {
+                        Regex::new(
+                            r"'s|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+",
+                        )
+                        .unwrap()
+                    }
+                }
+            }
+        } else if byte_fallback {
+            Regex::new(r"[\p{L}\p{N}]+|[^\s\p{L}\p{N}]|\s+").unwrap()
+        } else {
+            Regex::new(r"'s|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+").unwrap()
+        };
+
+        Self {
+            encoder,
+            decoder,
+            merge_ranks,
+            special_tokens,
+            special_token_ids,
+            pre_tokenize_re,
+            eos_token_id,
+            eos_token_ids,
+            byte_fallback,
+        }
+    }
+
+    pub fn encode(&self, text: &str) -> Vec<u32> {
+        let mut tokens = Vec::new();
+
+        // Check for special tokens first (split around them)
+        let mut remaining = text;
+        while !remaining.is_empty() {
+            // Find earliest special token
+            let mut earliest: Option<(usize, &str, u32)> = None;
+            for (st, &id) in &self.special_tokens {
+                if let Some(pos) = remaining.find(st.as_str()) {
+                    if earliest.is_none() || pos < earliest.unwrap().0 {
+                        earliest = Some((pos, st, id));
+                    }
+                }
+            }
+
+            if let Some((pos, st, id)) = earliest {
+                if pos > 0 {
+                    self.encode_ordinary(&remaining[..pos], &mut tokens);
+                }
+                tokens.push(id);
+                remaining = &remaining[pos + st.len()..];
+            } else {
+                self.encode_ordinary(remaining, &mut tokens);
+                break;
+            }
+        }
+
+        tokens
+    }
+
+    fn encode_ordinary(&self, text: &str, out: &mut Vec<u32>) {
+        for mat in self.pre_tokenize_re.find_iter(text) {
+            let word = mat.as_str();
+            // Try to encode the whole word first
+            if let Some(&id) = self.encoder.get(word.as_bytes()) {
+                out.push(id);
+                continue;
+            }
+            // Fall back to per-byte encoding
+            let word_bytes: Vec<u8> = word.bytes().collect();
+            let mut token_ids: Vec<u32> = word_bytes.iter().filter_map(|&b| {
+                if let Some(&id) = self.encoder.get(&vec![b]) {
+                    Some(id)
+                } else if self.byte_fallback {
+                    let hex_token = format!("<0x{:02X}>", b);
+                    if let Some(&id) = self.special_tokens.get(&hex_token) {
+                        Some(id)
+                    } else if let Some(&id) = self.encoder.get(hex_token.as_bytes()) {
+                        Some(id)
+                    } else if let Some(&unk_id) = self.special_tokens.get("<unk>") {
+                        eprintln!("warning: byte 0x{b:02X} not in vocab, using <unk> token");
+                        Some(unk_id)
+                    } else {
+                        eprintln!("warning: byte 0x{b:02X} not in vocab and no fallback token, using token 0");
+                        Some(0)
+                    }
+                } else {
+                    eprintln!("warning: byte {b} (0x{b:02X}) not in vocab, skipping");
+                    None
+                }
+            }).collect();
+
+            // BPE merges
+            loop {
+                if token_ids.len() < 2 {
+                    break;
+                }
+                let mut best_rank = usize::MAX;
+                let mut best_idx = 0;
+                for i in 0..token_ids.len() - 1 {
+                    if let Some(&rank) = self.merge_ranks.get(&(token_ids[i], token_ids[i + 1])) {
+                        if rank < best_rank {
+                            best_rank = rank;
+                            best_idx = i;
+                        }
+                    }
+                }
+                if best_rank == usize::MAX {
+                    break;
+                }
+
+                let merged_bytes = [
+                    self.decoder[token_ids[best_idx] as usize].as_slice(),
+                    self.decoder[token_ids[best_idx + 1] as usize].as_slice(),
+                ]
+                .concat();
+                let merged_id = *self.encoder.get(&merged_bytes).unwrap_or_else(|| {
+                    panic!("merged token not in vocab");
+                });
+                token_ids[best_idx] = merged_id;
+                token_ids.remove(best_idx + 1);
+            }
+
+            out.extend_from_slice(&token_ids);
+        }
+    }
+
+    pub fn decode(&self, token_ids: &[u32]) -> String {
+        let mut bytes = Vec::new();
+        for &id in token_ids {
+            if let Some(b) = self.decoder.get(id as usize) {
+                bytes.extend_from_slice(b);
+            }
+        }
+        String::from_utf8_lossy(&bytes).into_owned()
+    }
+
+    pub fn decode_token_stream(&self, token_id: u32, pending: &mut Vec<u8>) -> String {
+        if let Some(bytes) = self.decoder.get(token_id as usize) {
+            pending.extend_from_slice(bytes);
+        }
+        take_valid_utf8(pending)
+    }
+
+    pub fn flush_decode_stream(&self, pending: &mut Vec<u8>) -> String {
+        let text = String::from_utf8_lossy(pending).into_owned();
+        pending.clear();
+        text
+    }
+
+    pub fn eos_token_id(&self) -> Option<u32> {
+        self.eos_token_id
+    }
+
+    /// True if `id` is any end-of-generation token (a model may have several;
+    /// gpt-oss/harmony ends on <|return|>, <|call|>, or <|endoftext|>).
+    pub fn is_eos(&self, id: u32) -> bool {
+        self.eos_token_ids.contains(&id)
+    }
+
+    pub fn vocab_size(&self) -> usize {
+        self.decoder.len()
+    }
+
+    pub fn special_token_id(&self, name: &str) -> Option<u32> {
+        self.special_tokens.get(name).copied()
+    }
+}
+
+/// Convert a token string from HF vocab (which uses Unicode replacements for bytes)
+/// back to raw bytes. GPT-2 uses a byte-to-unicode mapping where e.g. byte 0x20 (space)
+/// is represented as 'Ġ' (U+0120).
+fn token_str_to_bytes(s: &str) -> Vec<u8> {
+    s.chars().map(|c| unicode_to_byte(c)).collect()
+}
+
+fn take_valid_utf8(pending: &mut Vec<u8>) -> String {
+    match std::str::from_utf8(pending) {
+        Ok(text) => {
+            let text = text.to_string();
+            pending.clear();
+            text
+        }
+        Err(err) => {
+            let valid_up_to = err.valid_up_to();
+            if valid_up_to == 0 {
+                if let Some(error_len) = err.error_len() {
+                    let invalid_len = error_len.min(pending.len());
+                    let text = String::from_utf8_lossy(&pending[..invalid_len]).into_owned();
+                    pending.drain(..invalid_len);
+                    return text;
+                }
+                return String::new();
+            }
+            let text = String::from_utf8_lossy(&pending[..valid_up_to]).into_owned();
+            pending.drain(..valid_up_to);
+            text
+        }
+    }
+}
+
+/// Convert a Unicode char back to the byte it represents in GPT-2 encoding.
+fn unicode_to_byte(c: char) -> u8 {
+    // Build the inverse map on first use
+    use std::sync::OnceLock;
+    static INV_MAP: OnceLock<HashMap<u32, u8>> = OnceLock::new();
+
+    let map = INV_MAP.get_or_init(|| {
+        let mut m = HashMap::new();
+        // Build GPT-2's bytes_to_unicode forward map, then invert
+        let mut n = 0u32;
+        for b in 0..=255u16 {
+            let byte = b as u8;
+            let unicode = match byte {
+                0x21..=0x7E | 0xA1..=0xAC | 0xAE..=0xFF => byte as u32,
+                _ => {
+                    let u = 256 + n;
+                    n += 1;
+                    u
+                }
+            };
+            m.insert(unicode, byte);
+        }
+        m
+    });
+
+    *map.get(&(c as u32))
+        .unwrap_or_else(|| panic!("unmapped unicode char U+{:04X} in tokenizer", c as u32))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::{Tokenizer, take_valid_utf8};
+
+    #[test]
+    fn qwen_added_tokens_are_indivisible_and_im_end_is_eos() {
+        let path =
+            std::env::temp_dir().join(format!("xserv-tokenizer-test-{}.json", std::process::id()));
+        std::fs::write(
+            &path,
+            r#"{
+                "model": {
+                    "vocab": {},
+                    "merges": [],
+                    "byte_fallback": false
+                },
+                "added_tokens": [
+                    {"id":151643,"content":"<|endoftext|>","special":true},
+                    {"id":151644,"content":"<|im_start|>","special":true},
+                    {"id":151645,"content":"<|im_end|>","special":true},
+                    {"id":151667,"content":"<think>","special":false},
+                    {"id":151668,"content":"</think>","special":false}
+                ]
+            }"#,
+        )
+        .unwrap();
+
+        let tokenizer = Tokenizer::from_file(&path);
+        let _ = std::fs::remove_file(&path);
+
+        assert_eq!(tokenizer.eos_token_id(), Some(151645));
+        assert_eq!(tokenizer.encode("<think>"), vec![151667]);
+        assert_eq!(tokenizer.encode("</think>"), vec![151668]);
+        assert_eq!(tokenizer.decode(&[151645]), "<|im_end|>");
+    }
+
+    #[test]
+    fn stream_decode_buffers_incomplete_utf8() {
+        let mut pending = vec![0xF0, 0x9F];
+        assert_eq!(take_valid_utf8(&mut pending), "");
+        pending.extend_from_slice(&[0x98, 0x8A, b'!']);
+        assert_eq!(take_valid_utf8(&mut pending), "😊!");
+        assert!(pending.is_empty());
+    }
+}

crates/xserv-tokenizer/src/lib.rs

@@ -0,0 +1,3 @@
+pub mod bpe;
+
+pub use bpe::Tokenizer;

csrc/activation/activations.cu

@@ -1,5 +1,6 @@
 #include <cuda_bf16.h>
 #include <math.h>
+#include "../common.cuh"
 
 // GELU (tanh approximation):
 // gelu(x) = 0.5 * x * (1 + tanh(sqrt(2/pi) * (x + 0.044715 * x^3)))
@@ -35,12 +36,85 @@ __global__ void silu_bf16(const __nv_bfloat16* x, __nv_bfloat16* out, int n) {
     if (idx < n) out[idx] = __float2bfloat16(silu_f(__bfloat162float(x[idx])));
 }
 
+__global__ void scale_f32_kernel(const float* x, float* out, float scale, int n) {
+    int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    if (idx < n) out[idx] = x[idx] * scale;
+}
+
+__global__ void scale_bf16_kernel(const __nv_bfloat16* x, __nv_bfloat16* out, float scale, int n) {
+    int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    if (idx < n) out[idx] = __float2bfloat16(__bfloat162float(x[idx]) * scale);
+}
+
+// Fused SiLU×Mul: out = silu(gate) * up
+__global__ void silu_mul_bf16_kernel(const __nv_bfloat16* gate, const __nv_bfloat16* up,
+                                     __nv_bfloat16* out, int n) {
+    int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    if (idx < n) {
+        float g = __bfloat162float(gate[idx]);
+        float u = __bfloat162float(up[idx]);
+        float silu_g = g / (1.0f + expf(-g));
+        out[idx] = __float2bfloat16(silu_g * u);
+    }
+}
+
+// gpt-oss GLU: gate_up is [N, 2*D] with interleaved columns (gate=even, up=odd).
+// gate = gate_up[::2].clamp(max=limit)
+// up   = gate_up[1::2].clamp(-limit, limit)
+// glu  = gate * sigmoid(gate * alpha)
+// out  = (up + 1) * glu
+// Output: [N, D]
+__global__ void gpt_oss_glu_bf16_kernel(const __nv_bfloat16* gate_up, __nv_bfloat16* out,
+                                         int n_elements, float alpha, float limit) {
+    int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    if (idx < n_elements) {
+        float g = __bfloat162float(gate_up[idx * 2]);
+        float u = __bfloat162float(gate_up[idx * 2 + 1]);
+        g = fminf(g, limit);
+        u = fmaxf(fminf(u, limit), -limit);
+        float glu = g / (1.0f + expf(-g * alpha));
+        out[idx] = __float2bfloat16((u + 1.0f) * glu);
+    }
+}
+
+// Element-wise add: out = a + b
+__global__ void add_f32_kernel(const float* a, const float* b, float* out, int n) {
+    int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    if (idx < n) out[idx] = a[idx] + b[idx];
+}
+__global__ void add_bf16_kernel(const __nv_bfloat16* a, const __nv_bfloat16* b, __nv_bfloat16* out, int n) {
+    int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    if (idx < n) out[idx] = __float2bfloat16(__bfloat162float(a[idx]) + __bfloat162float(b[idx]));
+}
+
+// Row-broadcast bias add: out[r, c] = x[r, c] + bias[c]
+__global__ void bias_add_2d_bf16_kernel(
+    const __nv_bfloat16* __restrict__ x, const __nv_bfloat16* __restrict__ bias,
+    __nv_bfloat16* __restrict__ out, int rows, int cols
+) {
+    int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    if (idx >= rows * cols) return;
+    float v = __bfloat162float(x[idx]) + __bfloat162float(bias[idx % cols]);
+    out[idx] = __float2bfloat16(v);
+}
+
+// Element-wise mul: out = a * b
+__global__ void mul_f32_kernel(const float* a, const float* b, float* out, int n) {
+    int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    if (idx < n) out[idx] = a[idx] * b[idx];
+}
+__global__ void mul_bf16_kernel(const __nv_bfloat16* a, const __nv_bfloat16* b, __nv_bfloat16* out, int n) {
+    int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    if (idx < n) out[idx] = __float2bfloat16(__bfloat162float(a[idx]) * __bfloat162float(b[idx]));
+}
+
 extern "C" {
 
 void launch_gelu_f32(const void* x, void* out, int n, void* stream) {
     int block = 256;
     int grid = (n + block - 1) / block;
     gelu_f32<<<grid, block, 0, (cudaStream_t)stream>>>((const float*)x, (float*)out, n);
+    CUDA_CHECK_LAST_ERROR();
 }
 
 void launch_gelu_bf16(const void* x, void* out, int n, void* stream) {
@@ -48,12 +122,14 @@ void launch_gelu_bf16(const void* x, void* out, int n, void* stream) {
     int grid = (n + block - 1) / block;
     gelu_bf16<<<grid, block, 0, (cudaStream_t)stream>>>(
         (const __nv_bfloat16*)x, (__nv_bfloat16*)out, n);
+    CUDA_CHECK_LAST_ERROR();
 }
 
 void launch_silu_f32(const void* x, void* out, int n, void* stream) {
     int block = 256;
     int grid = (n + block - 1) / block;
     silu_f32<<<grid, block, 0, (cudaStream_t)stream>>>((const float*)x, (float*)out, n);
+    CUDA_CHECK_LAST_ERROR();
 }
 
 void launch_silu_bf16(const void* x, void* out, int n, void* stream) {
@@ -61,6 +137,77 @@ void launch_silu_bf16(const void* x, void* out, int n, void* stream) {
     int grid = (n + block - 1) / block;
     silu_bf16<<<grid, block, 0, (cudaStream_t)stream>>>(
         (const __nv_bfloat16*)x, (__nv_bfloat16*)out, n);
+    CUDA_CHECK_LAST_ERROR();
+}
+
+void launch_scale_f32(const void* x, void* out, float scale, int n, void* stream) {
+    int block = 256;
+    int grid = (n + block - 1) / block;
+    scale_f32_kernel<<<grid, block, 0, (cudaStream_t)stream>>>(
+        (const float*)x, (float*)out, scale, n);
+    CUDA_CHECK_LAST_ERROR();
+}
+
+void launch_scale_bf16(const void* x, void* out, float scale, int n, void* stream) {
+    int block = 256;
+    int grid = (n + block - 1) / block;
+    scale_bf16_kernel<<<grid, block, 0, (cudaStream_t)stream>>>(
+        (const __nv_bfloat16*)x, (__nv_bfloat16*)out, scale, n);
+    CUDA_CHECK_LAST_ERROR();
+}
+
+void launch_add_f32(const void* a, const void* b, void* out, int n, void* stream) {
+    int block = 256;
+    int grid = (n + block - 1) / block;
+    add_f32_kernel<<<grid, block, 0, (cudaStream_t)stream>>>(
+        (const float*)a, (const float*)b, (float*)out, n);
+    CUDA_CHECK_LAST_ERROR();
+}
+void launch_add_bf16(const void* a, const void* b, void* out, int n, void* stream) {
+    int block = 256;
+    int grid = (n + block - 1) / block;
+    add_bf16_kernel<<<grid, block, 0, (cudaStream_t)stream>>>(
+        (const __nv_bfloat16*)a, (const __nv_bfloat16*)b, (__nv_bfloat16*)out, n);
+    CUDA_CHECK_LAST_ERROR();
+}
+void launch_bias_add_2d_bf16(const void* x, const void* bias, void* out, int rows, int cols, void* stream) {
+    int n = rows * cols;
+    int block = 256;
+    int grid = (n + block - 1) / block;
+    bias_add_2d_bf16_kernel<<<grid, block, 0, (cudaStream_t)stream>>>(
+        (const __nv_bfloat16*)x, (const __nv_bfloat16*)bias, (__nv_bfloat16*)out, rows, cols);
+    CUDA_CHECK_LAST_ERROR();
+}
+void launch_mul_f32(const void* a, const void* b, void* out, int n, void* stream) {
+    int block = 256;
+    int grid = (n + block - 1) / block;
+    mul_f32_kernel<<<grid, block, 0, (cudaStream_t)stream>>>(
+        (const float*)a, (const float*)b, (float*)out, n);
+    CUDA_CHECK_LAST_ERROR();
+}
+void launch_mul_bf16(const void* a, const void* b, void* out, int n, void* stream) {
+    int block = 256;
+    int grid = (n + block - 1) / block;
+    mul_bf16_kernel<<<grid, block, 0, (cudaStream_t)stream>>>(
+        (const __nv_bfloat16*)a, (const __nv_bfloat16*)b, (__nv_bfloat16*)out, n);
+    CUDA_CHECK_LAST_ERROR();
+}
+
+void launch_silu_mul_bf16(const void* gate, const void* up, void* out, int n, void* stream) {
+    int block = 256;
+    int grid = (n + block - 1) / block;
+    silu_mul_bf16_kernel<<<grid, block, 0, (cudaStream_t)stream>>>(
+        (const __nv_bfloat16*)gate, (const __nv_bfloat16*)up, (__nv_bfloat16*)out, n);
+    CUDA_CHECK_LAST_ERROR();
+}
+
+void launch_gpt_oss_glu_bf16(const void* gate_up, void* out, int n_elements,
+                              float alpha, float limit, void* stream) {
+    int block = 256;
+    int grid = (n_elements + block - 1) / block;
+    gpt_oss_glu_bf16_kernel<<<grid, block, 0, (cudaStream_t)stream>>>(
+        (const __nv_bfloat16*)gate_up, (__nv_bfloat16*)out, n_elements, alpha, limit);
+    CUDA_CHECK_LAST_ERROR();
 }
 
 }

csrc/attention/causal_mask.cu

@@ -0,0 +1,59 @@
+#include <cuda_bf16.h>
+#include "../common.cuh"
+
+// Apply causal mask: set scores[row][col] = -inf where col > row + offset.
+// offset is used for KV cache: when query starts at position `offset`,
+// we allow attending to positions [0, offset + row].
+// scores: [batch, rows, cols]  (flattened batch×heads)
+
+__global__ void causal_mask_f32(
+    float* __restrict__ scores,
+    int rows, int cols, int offset
+) {
+    int batch_idx = blockIdx.z;
+    int row = blockIdx.y;
+    int col = blockIdx.x * blockDim.x + threadIdx.x;
+
+    if (col < cols && col > row + offset) {
+        // 64-bit index: batch * rows * cols overflows int32 at moderate batch
+        // and long context (e.g. batch=128 * heads=28 * seq=32768).
+        long long idx = ((long long)batch_idx * rows + row) * cols + col;
+        scores[idx] = -INFINITY;
+    }
+}
+
+__global__ void causal_mask_bf16(
+    __nv_bfloat16* __restrict__ scores,
+    int rows, int cols, int offset
+) {
+    int batch_idx = blockIdx.z;
+    int row = blockIdx.y;
+    int col = blockIdx.x * blockDim.x + threadIdx.x;
+
+    if (col < cols && col > row + offset) {
+        long long idx = ((long long)batch_idx * rows + row) * cols + col;
+        scores[idx] = __float2bfloat16(-INFINITY);
+    }
+}
+
+extern "C" {
+
+void launch_causal_mask_f32(void* scores, int batch, int rows, int cols,
+                            int offset, void* stream) {
+    int block = 256;
+    dim3 grid((cols + block - 1) / block, rows, batch);
+    causal_mask_f32<<<grid, block, 0, (cudaStream_t)stream>>>(
+        (float*)scores, rows, cols, offset);
+    CUDA_CHECK_LAST_ERROR();
+}
+
+void launch_causal_mask_bf16(void* scores, int batch, int rows, int cols,
+                             int offset, void* stream) {
+    int block = 256;
+    dim3 grid((cols + block - 1) / block, rows, batch);
+    causal_mask_bf16<<<grid, block, 0, (cudaStream_t)stream>>>(
+        (__nv_bfloat16*)scores, rows, cols, offset);
+    CUDA_CHECK_LAST_ERROR();
+}
+
+}

csrc/attention/flash_attention.cu

@@ -0,0 +1,616 @@
+#include <cuda_bf16.h>
+#include <float.h>
+#include "../common.cuh"
+
+// Flash Attention 2 forward kernel for BF16 with FP32 accumulation.
+//
+// Algorithm: outer loop over Q tiles (BR rows), inner loop over K/V tiles (BC rows).
+// Uses online softmax — no O(S^2) memory.
+//
+// Layout: Q [batch, num_q_heads, q_len, head_dim]
+//         K [batch, num_kv_heads, kv_len, head_dim]
+//         V [batch, num_kv_heads, kv_len, head_dim]
+//         O [batch, num_q_heads, q_len, head_dim]
+//
+// Shared memory (BF16):
+//   smem_q[BR][head_dim]   — 64 * 128 * 2 = 16 KB (loaded once per Q tile)
+//   smem_kv[BC][head_dim]  — 64 * 128 * 2 = 16 KB (alternates K and V)
+//   Total: 32 KB (fits in default 48 KB shared memory)
+
+#define BR 64
+#define BC 64
+#define THREADS_PER_BLOCK 128
+
+__global__ void flash_attention_bf16_kernel(
+    const __nv_bfloat16* __restrict__ Q,
+    const __nv_bfloat16* __restrict__ K,
+    const __nv_bfloat16* __restrict__ V,
+    __nv_bfloat16* __restrict__ O,
+    int num_q_heads, int num_kv_heads,
+    int q_len, int kv_len, int head_dim,
+    float scale, int causal
+) {
+    // Grid: (ceil(q_len / BR), batch * num_q_heads)
+    int q_tile_idx = blockIdx.x;
+    int bh = blockIdx.y;
+    int batch_idx = bh / num_q_heads;
+    int q_head = bh % num_q_heads;
+
+    // GQA: map Q head to KV head
+    int heads_per_group = num_q_heads / num_kv_heads;
+    int kv_head = q_head / heads_per_group;
+
+    int q_tile_start = q_tile_idx * BR;
+    if (q_tile_start >= q_len) return;
+    int q_tile_rows = min(BR, q_len - q_tile_start);
+
+    // Pointers to this batch/head's data
+    const __nv_bfloat16* Q_head = Q + ((long long)batch_idx * num_q_heads + q_head) * q_len * head_dim;
+    const __nv_bfloat16* K_head = K + ((long long)batch_idx * num_kv_heads + kv_head) * kv_len * head_dim;
+    const __nv_bfloat16* V_head = V + ((long long)batch_idx * num_kv_heads + kv_head) * kv_len * head_dim;
+    __nv_bfloat16* O_head = O + ((long long)batch_idx * num_q_heads + q_head) * q_len * head_dim;
+
+    int tid = threadIdx.x;
+
+    // Dynamic shared memory
+    extern __shared__ __nv_bfloat16 smem[];
+    __nv_bfloat16* smem_q  = smem;                 // BR * head_dim elements
+    __nv_bfloat16* smem_kv = smem + BR * head_dim;  // BC * head_dim elements
+
+    // ---- Load Q tile into shared memory (cooperative) ----
+    int q_elems = q_tile_rows * head_dim;
+    for (int i = tid; i < q_elems; i += THREADS_PER_BLOCK) {
+        int row = i / head_dim;
+        int col = i % head_dim;
+        smem_q[row * head_dim + col] = Q_head[(q_tile_start + row) * head_dim + col];
+    }
+    // Zero-pad if q_tile_rows < BR
+    for (int i = q_elems + tid; i < BR * head_dim; i += THREADS_PER_BLOCK) {
+        smem_q[i] = __float2bfloat16(0.0f);
+    }
+    __syncthreads();
+
+    // Thread t (0 <= t < q_tile_rows) owns Q row t
+    bool owns_row = (tid < q_tile_rows);
+
+    // Per-thread FP32 accumulators (head_dim up to 128)
+    float O_acc[128];
+    float m_val = -INFINITY;
+    float l_val = 0.0f;
+    if (owns_row) {
+        for (int d = 0; d < head_dim; d++) {
+            O_acc[d] = 0.0f;
+        }
+    }
+
+    // kv_offset handles cached KV longer than Q (decode step)
+    int kv_offset = kv_len - q_len;
+    int num_kv_tiles = (kv_len + BC - 1) / BC;
+
+    // ---- Inner loop over K/V tiles ----
+    for (int j = 0; j < num_kv_tiles; j++) {
+        int kv_tile_start = j * BC;
+        int kv_tile_cols = min(BC, kv_len - kv_tile_start);
+
+        // Causal: skip entire tile if all K positions are in the future
+        if (causal) {
+            int max_allowed_kv = (q_tile_start + q_tile_rows - 1) + kv_offset;
+            if (kv_tile_start > max_allowed_kv) {
+                continue;
+            }
+        }
+
+        // ---- Load K tile into smem_kv ----
+        int kv_elems = kv_tile_cols * head_dim;
+        for (int i = tid; i < kv_elems; i += THREADS_PER_BLOCK) {
+            int row = i / head_dim;
+            int col = i % head_dim;
+            smem_kv[row * head_dim + col] = K_head[(kv_tile_start + row) * head_dim + col];
+        }
+        for (int i = kv_elems + tid; i < BC * head_dim; i += THREADS_PER_BLOCK) {
+            smem_kv[i] = __float2bfloat16(0.0f);
+        }
+        __syncthreads();
+
+        // ---- Compute S = Q @ K^T * scale, causal mask, online softmax ----
+        float P[BC];
+
+        if (owns_row) {
+            float row_max = -INFINITY;
+            for (int c = 0; c < kv_tile_cols; c++) {
+                float dot = 0.0f;
+                for (int d = 0; d < head_dim; d++) {
+                    dot += __bfloat162float(smem_q[tid * head_dim + d])
+                         * __bfloat162float(smem_kv[c * head_dim + d]);
+                }
+                float s = dot * scale;
+
+                if (causal) {
+                    int q_pos = q_tile_start + tid;
+                    int kv_pos = kv_tile_start + c;
+                    if (kv_pos > q_pos + kv_offset) {
+                        s = -INFINITY;
+                    }
+                }
+
+                P[c] = s;  // store score temporarily in P
+                row_max = fmaxf(row_max, s);
+            }
+
+            // Online softmax: m_new, P = exp(S - m_new), l_new
+            float m_new = fmaxf(m_val, row_max);
+
+            float psum = 0.0f;
+            for (int c = 0; c < kv_tile_cols; c++) {
+                P[c] = expf(P[c] - m_new);
+                psum += P[c];
+            }
+
+            // Rescale previous accumulator
+            float correction = expf(m_val - m_new);
+            l_val = correction * l_val + psum;
+
+            for (int d = 0; d < head_dim; d++) {
+                O_acc[d] *= correction;
+            }
+
+            m_val = m_new;
+        }
+
+        // Sync before overwriting smem_kv with V tile
+        __syncthreads();
+
+        // ---- Load V tile (reuse smem_kv) ----
+        int v_elems = kv_tile_cols * head_dim;
+        for (int i = tid; i < v_elems; i += THREADS_PER_BLOCK) {
+            int row = i / head_dim;
+            int col = i % head_dim;
+            smem_kv[row * head_dim + col] = V_head[(kv_tile_start + row) * head_dim + col];
+        }
+        for (int i = v_elems + tid; i < BC * head_dim; i += THREADS_PER_BLOCK) {
+            smem_kv[i] = __float2bfloat16(0.0f);
+        }
+        __syncthreads();
+
+        // ---- Accumulate O += P @ V_tile ----
+        if (owns_row) {
+            for (int c = 0; c < kv_tile_cols; c++) {
+                float p = P[c];
+                if (p != 0.0f) {
+                    for (int d = 0; d < head_dim; d++) {
+                        O_acc[d] += p * __bfloat162float(smem_kv[c * head_dim + d]);
+                    }
+                }
+            }
+        }
+
+        __syncthreads();
+    }
+
+    // ---- Final normalize and write output (convert FP32 → BF16) ----
+    if (owns_row) {
+        float inv_l = (l_val > 0.0f) ? (1.0f / l_val) : 0.0f;
+        int global_row = q_tile_start + tid;
+        for (int d = 0; d < head_dim; d++) {
+            O_head[global_row * head_dim + d] = __float2bfloat16(O_acc[d] * inv_l);
+        }
+    }
+}
+
+// Flash Attention 2 forward with gpt-oss attention sinks + optional sliding window.
+// Identical to flash_attention_bf16_kernel, plus:
+//   - sinks: [num_q_heads] BF16 — a per-head extra softmax logit (no value),
+//     folded into the denominator after the K/V tiles (exactly as the decode
+//     sink kernel does).
+//   - window_size > 0: sliding-window mask. Query at global position p attends
+//     to keys k with p - window_size < k <= p (matches HF gpt-oss).
+__global__ void flash_attention_sinks_bf16_kernel(
+    const __nv_bfloat16* __restrict__ Q,
+    const __nv_bfloat16* __restrict__ K,
+    const __nv_bfloat16* __restrict__ V,
+    __nv_bfloat16* __restrict__ O,
+    const __nv_bfloat16* __restrict__ sinks, // [num_q_heads] or NULL
+    int num_q_heads, int num_kv_heads,
+    int q_len, int kv_len, int head_dim,
+    float scale, int causal, int window_size
+) {
+    int q_tile_idx = blockIdx.x;
+    int bh = blockIdx.y;
+    int batch_idx = bh / num_q_heads;
+    int q_head = bh % num_q_heads;
+
+    int heads_per_group = num_q_heads / num_kv_heads;
+    int kv_head = q_head / heads_per_group;
+
+    int q_tile_start = q_tile_idx * BR;
+    if (q_tile_start >= q_len) return;
+    int q_tile_rows = min(BR, q_len - q_tile_start);
+
+    const __nv_bfloat16* Q_head = Q + ((long long)batch_idx * num_q_heads + q_head) * q_len * head_dim;
+    const __nv_bfloat16* K_head = K + ((long long)batch_idx * num_kv_heads + kv_head) * kv_len * head_dim;
+    const __nv_bfloat16* V_head = V + ((long long)batch_idx * num_kv_heads + kv_head) * kv_len * head_dim;
+    __nv_bfloat16* O_head = O + ((long long)batch_idx * num_q_heads + q_head) * q_len * head_dim;
+
+    int tid = threadIdx.x;
+
+    extern __shared__ __nv_bfloat16 smem[];
+    __nv_bfloat16* smem_q  = smem;
+    __nv_bfloat16* smem_kv = smem + BR * head_dim;
+
+    int q_elems = q_tile_rows * head_dim;
+    for (int i = tid; i < q_elems; i += THREADS_PER_BLOCK) {
+        int row = i / head_dim;
+        int col = i % head_dim;
+        smem_q[row * head_dim + col] = Q_head[(q_tile_start + row) * head_dim + col];
+    }
+    for (int i = q_elems + tid; i < BR * head_dim; i += THREADS_PER_BLOCK) {
+        smem_q[i] = __float2bfloat16(0.0f);
+    }
+    __syncthreads();
+
+    bool owns_row = (tid < q_tile_rows);
+
+    float O_acc[128];
+    float m_val = -INFINITY;
+    float l_val = 0.0f;
+    if (owns_row) {
+        for (int d = 0; d < head_dim; d++) O_acc[d] = 0.0f;
+    }
+
+    int kv_offset = kv_len - q_len;
+    int num_kv_tiles = (kv_len + BC - 1) / BC;
+
+    for (int j = 0; j < num_kv_tiles; j++) {
+        int kv_tile_start = j * BC;
+        int kv_tile_cols = min(BC, kv_len - kv_tile_start);
+
+        if (causal) {
+            int max_allowed_kv = (q_tile_start + q_tile_rows - 1) + kv_offset;
+            if (kv_tile_start > max_allowed_kv) continue;
+        }
+
+        int kv_elems = kv_tile_cols * head_dim;
+        for (int i = tid; i < kv_elems; i += THREADS_PER_BLOCK) {
+            int row = i / head_dim;
+            int col = i % head_dim;
+            smem_kv[row * head_dim + col] = K_head[(kv_tile_start + row) * head_dim + col];
+        }
+        for (int i = kv_elems + tid; i < BC * head_dim; i += THREADS_PER_BLOCK) {
+            smem_kv[i] = __float2bfloat16(0.0f);
+        }
+        __syncthreads();
+
+        float P[BC];
+
+        if (owns_row) {
+            float row_max = -INFINITY;
+            int q_pos = q_tile_start + tid + kv_offset; // global query position
+            for (int c = 0; c < kv_tile_cols; c++) {
+                float dot = 0.0f;
+                for (int d = 0; d < head_dim; d++) {
+                    dot += __bfloat162float(smem_q[tid * head_dim + d])
+                         * __bfloat162float(smem_kv[c * head_dim + d]);
+                }
+                float s = dot * scale;
+
+                int kv_pos = kv_tile_start + c;
+                if (causal && kv_pos > q_pos) {
+                    s = -INFINITY;
+                }
+                // Sliding window: drop keys older than the window.
+                if (window_size > 0 && kv_pos <= q_pos - window_size) {
+                    s = -INFINITY;
+                }
+
+                P[c] = s;
+                row_max = fmaxf(row_max, s);
+            }
+
+            // A fully-masked KV tile (every key causal- or window-masked) has
+            // row_max == -INFINITY. Folding it in computes expf(-inf - (-inf))
+            // = NaN, and a later valid tile's 0*NaN correction then poisons the
+            // whole row. This happens for sliding-window layers whenever a
+            // query's window starts past an early tile (the causal `continue`
+            // above only skips fully-future tiles, not out-of-window ones).
+            // A masked tile contributes nothing to the softmax — skip it.
+            if (row_max != -INFINITY) {
+                float m_new = fmaxf(m_val, row_max);
+                float psum = 0.0f;
+                for (int c = 0; c < kv_tile_cols; c++) {
+                    P[c] = expf(P[c] - m_new);
+                    psum += P[c];
+                }
+                float correction = expf(m_val - m_new);
+                l_val = correction * l_val + psum;
+                for (int d = 0; d < head_dim; d++) O_acc[d] *= correction;
+                m_val = m_new;
+            } else {
+                for (int c = 0; c < kv_tile_cols; c++) P[c] = 0.0f;
+            }
+        }
+
+        __syncthreads();
+
+        int v_elems = kv_tile_cols * head_dim;
+        for (int i = tid; i < v_elems; i += THREADS_PER_BLOCK) {
+            int row = i / head_dim;
+            int col = i % head_dim;
+            smem_kv[row * head_dim + col] = V_head[(kv_tile_start + row) * head_dim + col];
+        }
+        for (int i = v_elems + tid; i < BC * head_dim; i += THREADS_PER_BLOCK) {
+            smem_kv[i] = __float2bfloat16(0.0f);
+        }
+        __syncthreads();
+
+        if (owns_row) {
+            for (int c = 0; c < kv_tile_cols; c++) {
+                float p = P[c];
+                if (p != 0.0f) {
+                    for (int d = 0; d < head_dim; d++) {
+                        O_acc[d] += p * __bfloat162float(smem_kv[c * head_dim + d]);
+                    }
+                }
+            }
+        }
+        __syncthreads();
+    }
+
+    // Fold in the per-head attention sink (extra logit, no value contribution).
+    if (owns_row && sinks != nullptr) {
+        float sink_logit = __bfloat162float(sinks[q_head]);
+        float m_new = fmaxf(m_val, sink_logit);
+        float correction = expf(m_val - m_new);
+        l_val = correction * l_val + expf(sink_logit - m_new);
+        for (int d = 0; d < head_dim; d++) O_acc[d] *= correction;
+        m_val = m_new;
+    }
+
+    if (owns_row) {
+        float inv_l = (l_val > 0.0f) ? (1.0f / l_val) : 0.0f;
+        int global_row = q_tile_start + tid;
+        for (int d = 0; d < head_dim; d++) {
+            O_head[global_row * head_dim + d] = __float2bfloat16(O_acc[d] * inv_l);
+        }
+    }
+}
+
+// ============================================================
+// Decode Attention kernel: optimized for Q_len=1 (single-token decode).
+// Parallelizes across KV sequence dimension instead of Q rows.
+//
+// Grid: (batch * num_q_heads, 1) — one block per Q head
+// Block: 256 threads — each thread handles ceil(kv_len / 256) KV positions
+// Uses online softmax reduction across threads.
+// ============================================================
+
+#define DECODE_THREADS 256
+#define HEAD_DIM_MAX 128
+
+__global__ void decode_attention_bf16_kernel(
+    const __nv_bfloat16* __restrict__ Q,
+    const __nv_bfloat16* __restrict__ K,
+    const __nv_bfloat16* __restrict__ V,
+    __nv_bfloat16* __restrict__ O,
+    int num_q_heads, int num_kv_heads,
+    int kv_len, int head_dim,
+    float scale
+) {
+    int bh = blockIdx.x;
+    int batch_idx = bh / num_q_heads;
+    int q_head = bh % num_q_heads;
+
+    // GQA mapping
+    int heads_per_group = num_q_heads / num_kv_heads;
+    int kv_head = q_head / heads_per_group;
+
+    int tid = threadIdx.x;
+
+    // Pointers to this batch/head's data
+    // Q: [batch, num_q_heads, 1, head_dim]
+    const __nv_bfloat16* Q_ptr = Q + ((long long)batch_idx * num_q_heads + q_head) * head_dim;
+    // K/V: [batch, num_kv_heads, kv_len, head_dim]
+    const __nv_bfloat16* K_base = K + ((long long)batch_idx * num_kv_heads + kv_head) * kv_len * head_dim;
+    const __nv_bfloat16* V_base = V + ((long long)batch_idx * num_kv_heads + kv_head) * kv_len * head_dim;
+    __nv_bfloat16* O_ptr = O + ((long long)batch_idx * num_q_heads + q_head) * head_dim;
+
+    // Load Q vector into registers (head_dim <= 128)
+    float q_reg[HEAD_DIM_MAX];
+    for (int d = 0; d < head_dim; d++) {
+        q_reg[d] = __bfloat162float(Q_ptr[d]);
+    }
+
+    // Each thread processes a chunk of KV positions
+    // Thread tid handles positions: tid, tid+DECODE_THREADS, tid+2*DECODE_THREADS, ...
+    float local_max = -INFINITY;
+    float local_sum = 0.0f;
+    float local_O[HEAD_DIM_MAX];
+    for (int d = 0; d < head_dim; d++) {
+        local_O[d] = 0.0f;
+    }
+
+    for (int pos = tid; pos < kv_len; pos += DECODE_THREADS) {
+        // Compute dot(Q, K[pos]) * scale
+        const __nv_bfloat16* K_pos = K_base + pos * head_dim;
+        float dot = 0.0f;
+        for (int d = 0; d < head_dim; d++) {
+            dot += q_reg[d] * __bfloat162float(K_pos[d]);
+        }
+        float s = dot * scale;
+
+        // Online softmax update
+        float new_max = fmaxf(local_max, s);
+        float correction = expf(local_max - new_max);
+        float p = expf(s - new_max);
+
+        // Rescale running sum and O
+        local_sum = local_sum * correction + p;
+        for (int d = 0; d < head_dim; d++) {
+            local_O[d] = local_O[d] * correction;
+        }
+
+        // Accumulate V[pos] weighted by p
+        const __nv_bfloat16* V_pos = V_base + pos * head_dim;
+        for (int d = 0; d < head_dim; d++) {
+            local_O[d] += p * __bfloat162float(V_pos[d]);
+        }
+
+        local_max = new_max;
+    }
+
+    // --- Block-level online softmax reduction ---
+    // We need to combine (local_max, local_sum, local_O) across all threads.
+    // Strategy: reduce max, then each thread rescales, then reduce sum and O.
+
+    // Shared memory for reduction
+    __shared__ float smem_max[32];   // one per warp
+    __shared__ float smem_sum[32];
+    __shared__ float smem_O_warp[32][HEAD_DIM_MAX];
+
+    // Step 1: Block-wide max reduction
+    int lane = tid & 31;
+    int warp_id = tid >> 5;
+    int num_warps = DECODE_THREADS >> 5; // 8 warps
+
+    float warp_max = local_max;
+    #pragma unroll
+    for (int offset = 16; offset > 0; offset >>= 1)
+        warp_max = fmaxf(warp_max, __shfl_down_sync(0xffffffff, warp_max, offset));
+    if (lane == 0) smem_max[warp_id] = warp_max;
+    __syncthreads();
+
+    float global_max;
+    if (tid == 0) {
+        global_max = smem_max[0];
+        for (int i = 1; i < num_warps; i++)
+            global_max = fmaxf(global_max, smem_max[i]);
+        smem_max[0] = global_max;
+    }
+    __syncthreads();
+    global_max = smem_max[0];
+
+    // Step 2: Each thread rescales its local_sum and local_O with global_max
+    float rescale = (local_max == -INFINITY) ? 0.0f : expf(local_max - global_max);
+    local_sum *= rescale;
+    for (int d = 0; d < head_dim; d++) {
+        local_O[d] *= rescale;
+    }
+
+    // Step 3: Reduce sum across block
+    float warp_sum = local_sum;
+    #pragma unroll
+    for (int offset = 16; offset > 0; offset >>= 1)
+        warp_sum += __shfl_down_sync(0xffffffff, warp_sum, offset);
+    if (lane == 0) smem_sum[warp_id] = warp_sum;
+    __syncthreads();
+
+    float global_sum;
+    if (tid == 0) {
+        global_sum = 0.0f;
+        for (int i = 0; i < num_warps; i++)
+            global_sum += smem_sum[i];
+        smem_sum[0] = global_sum;
+    }
+    __syncthreads();
+    global_sum = smem_sum[0];
+
+    // Step 4: Reduce O across block, dim by dim. Store one partial per warp
+    // and sum in warp-id order; atomicAdd made greedy decode nondeterministic
+    // when logits were close (same fix pattern as paged_attention.cu / gemv.cu).
+    float inv_sum = (global_sum > 0.0f) ? (1.0f / global_sum) : 0.0f;
+
+    for (int i = tid; i < 32 * HEAD_DIM_MAX; i += DECODE_THREADS) {
+        reinterpret_cast<float*>(smem_O_warp)[i] = 0.0f;
+    }
+    __syncthreads();
+
+    for (int d = 0; d < head_dim; d++) {
+        float val = local_O[d];
+        #pragma unroll
+        for (int offset = 16; offset > 0; offset >>= 1)
+            val += __shfl_down_sync(0xffffffff, val, offset);
+        if (lane == 0) smem_O_warp[warp_id][d] = val;
+    }
+    __syncthreads();
+
+    // Thread 0..head_dim-1 write final output
+    for (int d = tid; d < head_dim; d += DECODE_THREADS) {
+        float out = 0.0f;
+        for (int i = 0; i < num_warps; i++) out += smem_O_warp[i][d];
+        O_ptr[d] = __float2bfloat16(out * inv_sum);
+    }
+}
+
+extern "C" {
+
+void launch_flash_attention_bf16(
+    const void* Q, const void* K, const void* V, void* O,
+    int batch, int num_q_heads, int num_kv_heads,
+    int q_len, int kv_len, int head_dim,
+    float scale, int causal, void* stream
+) {
+    int q_tiles = (q_len + BR - 1) / BR;
+    dim3 grid(q_tiles, batch * num_q_heads);
+    int block = THREADS_PER_BLOCK;
+
+    // Shared memory: smem_q[BR * head_dim] + smem_kv[BC * head_dim], all BF16
+    int smem_bytes = (BR + BC) * head_dim * (int)sizeof(__nv_bfloat16);
+
+    flash_attention_bf16_kernel<<<grid, block, smem_bytes, (cudaStream_t)stream>>>(
+        (const __nv_bfloat16*)Q,
+        (const __nv_bfloat16*)K,
+        (const __nv_bfloat16*)V,
+        (__nv_bfloat16*)O,
+        num_q_heads, num_kv_heads,
+        q_len, kv_len, head_dim,
+        scale, causal
+    );
+    CUDA_CHECK_LAST_ERROR();
+}
+
+void launch_flash_attention_sinks_bf16(
+    const void* Q, const void* K, const void* V, void* O,
+    const void* sinks,
+    int batch, int num_q_heads, int num_kv_heads,
+    int q_len, int kv_len, int head_dim,
+    float scale, int causal, int window_size, void* stream
+) {
+    int q_tiles = (q_len + BR - 1) / BR;
+    dim3 grid(q_tiles, batch * num_q_heads);
+    int block = THREADS_PER_BLOCK;
+    int smem_bytes = (BR + BC) * head_dim * (int)sizeof(__nv_bfloat16);
+
+    flash_attention_sinks_bf16_kernel<<<grid, block, smem_bytes, (cudaStream_t)stream>>>(
+        (const __nv_bfloat16*)Q,
+        (const __nv_bfloat16*)K,
+        (const __nv_bfloat16*)V,
+        (__nv_bfloat16*)O,
+        (const __nv_bfloat16*)sinks,
+        num_q_heads, num_kv_heads,
+        q_len, kv_len, head_dim,
+        scale, causal, window_size
+    );
+    CUDA_CHECK_LAST_ERROR();
+}
+
+void launch_decode_attention_bf16(
+    const void* Q, const void* K, const void* V, void* O,
+    int batch, int num_q_heads, int num_kv_heads,
+    int kv_len, int head_dim,
+    float scale, int causal, void* stream
+) {
+    int grid = batch * num_q_heads;
+    int block = DECODE_THREADS;
+
+    decode_attention_bf16_kernel<<<grid, block, 0, (cudaStream_t)stream>>>(
+        (const __nv_bfloat16*)Q,
+        (const __nv_bfloat16*)K,
+        (const __nv_bfloat16*)V,
+        (__nv_bfloat16*)O,
+        num_q_heads, num_kv_heads,
+        kv_len, head_dim,
+        scale
+    );
+    CUDA_CHECK_LAST_ERROR();
+}
+
+}

csrc/attention/paged_attention.cu

@@ -0,0 +1,614 @@
+#include <cuda_bf16.h>
+#include <float.h>
+#include "../common.cuh"
+
+// Paged decode attention kernel for BF16 with FP32 accumulation.
+//
+// Reads K/V from a paged pool indexed by a per-sequence block table.
+// One CUDA block per (sequence, q_head). Each block streams over the
+// sequence's KV positions and accumulates attention output via online
+// softmax.
+//
+// Layouts:
+//   Q          [batch, num_q_heads, 1, head_dim]                    BF16
+//   K_cache    [num_blocks, num_kv_heads, BLOCK_SIZE, head_dim]     BF16
+//   V_cache    same
+//   block_tables  [max_seqs, max_blocks_per_seq]                    int32
+//                 — the i-th sequence in this launch reads row
+//                   block_tables[seq_slot[i] * stride + ...].
+//                 For simplicity the launch passes a packed row table
+//                 [batch, max_blocks_per_seq] (already gathered for the
+//                 active batch) so we just index by blockIdx.x_seq.
+//   context_lens  [batch] int32  — number of valid tokens per sequence.
+//
+// One CUDA block: 256 threads, head_dim <= 128.
+
+#define PAGED_BLOCK_SIZE 16
+#define PAGED_THREADS 256
+#define PAGED_HEAD_DIM_MAX 128
+
+__global__ void paged_decode_attention_bf16_kernel(
+    const __nv_bfloat16* __restrict__ Q,
+    const __nv_bfloat16* __restrict__ K_cache,
+    const __nv_bfloat16* __restrict__ V_cache,
+    __nv_bfloat16* __restrict__ O,
+    const int* __restrict__ block_tables,    // [batch, max_blocks_per_seq]
+    const int* __restrict__ context_lens,    // [batch]
+    int num_q_heads, int num_kv_heads,
+    int head_dim, int max_blocks_per_seq,
+    float scale
+) {
+    int seq_idx = blockIdx.y;          // batch dim
+    int q_head  = blockIdx.x;          // 0 .. num_q_heads-1
+    int tid     = threadIdx.x;
+
+    int kv_len = context_lens[seq_idx];
+    if (kv_len <= 0) {
+        // Nothing to attend over; zero output for safety.
+        if (tid < head_dim) {
+            O[((long long)seq_idx * num_q_heads + q_head) * head_dim + tid] =
+                __float2bfloat16(0.0f);
+        }
+        return;
+    }
+
+    // GQA mapping
+    int heads_per_group = num_q_heads / num_kv_heads;
+    int kv_head = q_head / heads_per_group;
+
+    // Pointers
+    const __nv_bfloat16* Q_ptr = Q +
+        ((long long)seq_idx * num_q_heads + q_head) * head_dim;
+    __nv_bfloat16* O_ptr = O +
+        ((long long)seq_idx * num_q_heads + q_head) * head_dim;
+    const int* bt = block_tables + (long long)seq_idx * max_blocks_per_seq;
+
+    // Load Q vector into registers.
+    float q_reg[PAGED_HEAD_DIM_MAX];
+    for (int d = 0; d < head_dim; d++) {
+        q_reg[d] = __bfloat162float(Q_ptr[d]);
+    }
+
+    // Per-thread online softmax state.
+    float local_max = -INFINITY;
+    float local_sum = 0.0f;
+    float local_O[PAGED_HEAD_DIM_MAX];
+    for (int d = 0; d < head_dim; d++) local_O[d] = 0.0f;
+
+    int kv_stride_block = num_kv_heads * PAGED_BLOCK_SIZE * head_dim;
+    int kv_stride_head  = PAGED_BLOCK_SIZE * head_dim;
+
+    // Each thread handles positions tid, tid+PAGED_THREADS, ...
+    for (int pos = tid; pos < kv_len; pos += PAGED_THREADS) {
+        int logical_blk = pos / PAGED_BLOCK_SIZE;
+        int slot_in_blk = pos % PAGED_BLOCK_SIZE;
+        int phys_blk = bt[logical_blk];
+
+        const __nv_bfloat16* K_pos = K_cache
+            + (long long)phys_blk * kv_stride_block
+            + kv_head * kv_stride_head
+            + slot_in_blk * head_dim;
+        const __nv_bfloat16* V_pos = V_cache
+            + (long long)phys_blk * kv_stride_block
+            + kv_head * kv_stride_head
+            + slot_in_blk * head_dim;
+
+        // dot(Q, K[pos]) * scale
+        float dot = 0.0f;
+        for (int d = 0; d < head_dim; d++) {
+            dot += q_reg[d] * __bfloat162float(K_pos[d]);
+        }
+        float s = dot * scale;
+
+        float new_max = fmaxf(local_max, s);
+        float correction = expf(local_max - new_max);
+        float p = expf(s - new_max);
+
+        local_sum = local_sum * correction + p;
+        for (int d = 0; d < head_dim; d++) local_O[d] *= correction;
+
+        // Accumulate weighted V.
+        for (int d = 0; d < head_dim; d++) {
+            local_O[d] += p * __bfloat162float(V_pos[d]);
+        }
+
+        local_max = new_max;
+    }
+
+    // ---- Block-level online softmax reduction ----
+    __shared__ float smem_max[32];
+    __shared__ float smem_sum[32];
+    __shared__ float smem_O_warp[32][PAGED_HEAD_DIM_MAX];
+
+    int lane = tid & 31;
+    int warp_id = tid >> 5;
+    int num_warps = PAGED_THREADS >> 5;
+
+    // Step 1: block-wide max
+    float warp_max = local_max;
+    #pragma unroll
+    for (int offset = 16; offset > 0; offset >>= 1)
+        warp_max = fmaxf(warp_max, __shfl_down_sync(0xffffffff, warp_max, offset));
+    if (lane == 0) smem_max[warp_id] = warp_max;
+    __syncthreads();
+
+    float global_max;
+    if (tid == 0) {
+        global_max = smem_max[0];
+        for (int i = 1; i < num_warps; i++)
+            global_max = fmaxf(global_max, smem_max[i]);
+        smem_max[0] = global_max;
+    }
+    __syncthreads();
+    global_max = smem_max[0];
+
+    // Step 2: rescale local state to global_max
+    float rescale = (local_max == -INFINITY) ? 0.0f : expf(local_max - global_max);
+    local_sum *= rescale;
+    for (int d = 0; d < head_dim; d++) local_O[d] *= rescale;
+
+    // Step 3: reduce sum
+    float warp_sum = local_sum;
+    #pragma unroll
+    for (int offset = 16; offset > 0; offset >>= 1)
+        warp_sum += __shfl_down_sync(0xffffffff, warp_sum, offset);
+    if (lane == 0) smem_sum[warp_id] = warp_sum;
+    __syncthreads();
+
+    float global_sum;
+    if (tid == 0) {
+        global_sum = 0.0f;
+        for (int i = 0; i < num_warps; i++) global_sum += smem_sum[i];
+        smem_sum[0] = global_sum;
+    }
+    __syncthreads();
+    global_sum = smem_sum[0];
+
+    // Step 4: reduce O across block, dim by dim. Store one partial per warp
+    // and sum in warp-id order; atomicAdd made greedy decode nondeterministic
+    // when logits were close.
+    for (int i = tid; i < 32 * PAGED_HEAD_DIM_MAX; i += PAGED_THREADS) {
+        reinterpret_cast<float*>(smem_O_warp)[i] = 0.0f;
+    }
+    __syncthreads();
+
+    for (int d = 0; d < head_dim; d++) {
+        float val = local_O[d];
+        #pragma unroll
+        for (int offset = 16; offset > 0; offset >>= 1)
+            val += __shfl_down_sync(0xffffffff, val, offset);
+        if (lane == 0) smem_O_warp[warp_id][d] = val;
+    }
+    __syncthreads();
+
+    float inv_sum = (global_sum > 0.0f) ? (1.0f / global_sum) : 0.0f;
+    for (int d = tid; d < head_dim; d += PAGED_THREADS) {
+        float out = 0.0f;
+        for (int i = 0; i < num_warps; i++) out += smem_O_warp[i][d];
+        O_ptr[d] = __float2bfloat16(out * inv_sum);
+    }
+}
+
+// Tree-aware paged decode attention: per-query mask lets sibling candidates
+// in the same batch attend to different subsets of newly-written K/V.
+// `tree_start`: position where newly-written K/V begins (typically pos_offset).
+// `tree_len`: number of newly-written K/V rows (= batch, one per query).
+// `tree_mask[i][j] = 1` iff query i attends to K/V at position `tree_start+j`.
+// Positions < tree_start are always attended (regular history).
+__global__ void paged_decode_attention_tree_bf16_kernel(
+    const __nv_bfloat16* __restrict__ Q,
+    const __nv_bfloat16* __restrict__ K_cache,
+    const __nv_bfloat16* __restrict__ V_cache,
+    __nv_bfloat16* __restrict__ O,
+    const int* __restrict__ block_tables,
+    const int* __restrict__ context_lens,
+    const int* __restrict__ tree_mask,       // [batch, tree_len] int32
+    int num_q_heads, int num_kv_heads,
+    int head_dim, int max_blocks_per_seq,
+    int tree_start, int tree_len,
+    float scale
+) {
+    int seq_idx = blockIdx.y;
+    int q_head  = blockIdx.x;
+    int tid     = threadIdx.x;
+
+    int kv_len = context_lens[seq_idx];
+    if (kv_len <= 0) {
+        if (tid < head_dim) {
+            O[((long long)seq_idx * num_q_heads + q_head) * head_dim + tid] =
+                __float2bfloat16(0.0f);
+        }
+        return;
+    }
+
+    int heads_per_group = num_q_heads / num_kv_heads;
+    int kv_head = q_head / heads_per_group;
+
+    const __nv_bfloat16* Q_ptr = Q +
+        ((long long)seq_idx * num_q_heads + q_head) * head_dim;
+    __nv_bfloat16* O_ptr = O +
+        ((long long)seq_idx * num_q_heads + q_head) * head_dim;
+    const int* bt = block_tables + (long long)seq_idx * max_blocks_per_seq;
+    const int* mask_row = tree_mask + (long long)seq_idx * tree_len;
+
+    float q_reg[PAGED_HEAD_DIM_MAX];
+    for (int d = 0; d < head_dim; d++) {
+        q_reg[d] = __bfloat162float(Q_ptr[d]);
+    }
+
+    float local_max = -INFINITY;
+    float local_sum = 0.0f;
+    float local_O[PAGED_HEAD_DIM_MAX];
+    for (int d = 0; d < head_dim; d++) local_O[d] = 0.0f;
+
+    int kv_stride_block = num_kv_heads * PAGED_BLOCK_SIZE * head_dim;
+    int kv_stride_head  = PAGED_BLOCK_SIZE * head_dim;
+
+    for (int pos = tid; pos < kv_len; pos += PAGED_THREADS) {
+        // Tree mask: skip positions in [tree_start, tree_start+tree_len) that
+        // the mask marks as 0. Everything else (history) is always attended.
+        if (pos >= tree_start && pos < tree_start + tree_len) {
+            if (mask_row[pos - tree_start] == 0) continue;
+        }
+
+        int logical_blk = pos / PAGED_BLOCK_SIZE;
+        int slot_in_blk = pos % PAGED_BLOCK_SIZE;
+        int phys_blk = bt[logical_blk];
+
+        const __nv_bfloat16* K_pos = K_cache
+            + (long long)phys_blk * kv_stride_block
+            + kv_head * kv_stride_head
+            + slot_in_blk * head_dim;
+        const __nv_bfloat16* V_pos = V_cache
+            + (long long)phys_blk * kv_stride_block
+            + kv_head * kv_stride_head
+            + slot_in_blk * head_dim;
+
+        float dot = 0.0f;
+        for (int d = 0; d < head_dim; d++) {
+            dot += q_reg[d] * __bfloat162float(K_pos[d]);
+        }
+        float s = dot * scale;
+
+        float new_max = fmaxf(local_max, s);
+        float correction = expf(local_max - new_max);
+        float p = expf(s - new_max);
+
+        local_sum = local_sum * correction + p;
+        for (int d = 0; d < head_dim; d++) local_O[d] *= correction;
+
+        for (int d = 0; d < head_dim; d++) {
+            local_O[d] += p * __bfloat162float(V_pos[d]);
+        }
+
+        local_max = new_max;
+    }
+
+    // Block-level reduction (identical to base kernel).
+    __shared__ float smem_max[32];
+    __shared__ float smem_sum[32];
+    __shared__ float smem_O_warp[32][PAGED_HEAD_DIM_MAX];
+
+    int lane = tid & 31;
+    int warp_id = tid >> 5;
+    int num_warps = PAGED_THREADS >> 5;
+
+    float warp_max = local_max;
+    #pragma unroll
+    for (int offset = 16; offset > 0; offset >>= 1)
+        warp_max = fmaxf(warp_max, __shfl_down_sync(0xffffffff, warp_max, offset));
+    if (lane == 0) smem_max[warp_id] = warp_max;
+    __syncthreads();
+
+    float global_max;
+    if (tid == 0) {
+        global_max = smem_max[0];
+        for (int i = 1; i < num_warps; i++)
+            global_max = fmaxf(global_max, smem_max[i]);
+        smem_max[0] = global_max;
+    }
+    __syncthreads();
+    global_max = smem_max[0];
+
+    float rescale = (local_max == -INFINITY) ? 0.0f : expf(local_max - global_max);
+    local_sum *= rescale;
+    for (int d = 0; d < head_dim; d++) local_O[d] *= rescale;
+
+    float warp_sum = local_sum;
+    #pragma unroll
+    for (int offset = 16; offset > 0; offset >>= 1)
+        warp_sum += __shfl_down_sync(0xffffffff, warp_sum, offset);
+    if (lane == 0) smem_sum[warp_id] = warp_sum;
+    __syncthreads();
+
+    float global_sum;
+    if (tid == 0) {
+        global_sum = 0.0f;
+        for (int i = 0; i < num_warps; i++) global_sum += smem_sum[i];
+        smem_sum[0] = global_sum;
+    }
+    __syncthreads();
+    global_sum = smem_sum[0];
+
+    for (int i = tid; i < 32 * PAGED_HEAD_DIM_MAX; i += PAGED_THREADS) {
+        reinterpret_cast<float*>(smem_O_warp)[i] = 0.0f;
+    }
+    __syncthreads();
+
+    for (int d = 0; d < head_dim; d++) {
+        float val = local_O[d];
+        #pragma unroll
+        for (int offset = 16; offset > 0; offset >>= 1)
+            val += __shfl_down_sync(0xffffffff, val, offset);
+        if (lane == 0) smem_O_warp[warp_id][d] = val;
+    }
+    __syncthreads();
+
+    float inv_sum = (global_sum > 0.0f) ? (1.0f / global_sum) : 0.0f;
+    for (int d = tid; d < head_dim; d += PAGED_THREADS) {
+        float out = 0.0f;
+        for (int i = 0; i < num_warps; i++) out += smem_O_warp[i][d];
+        O_ptr[d] = __float2bfloat16(out * inv_sum);
+    }
+}
+
+// Extended paged decode attention with attention sinks and sliding window.
+// sinks: [num_q_heads] BF16 — per-head extra logit appended before softmax.
+// window_size: >0 = sliding window (only attend to last `window_size` positions), 0 = full.
+__global__ void paged_decode_attention_sinks_bf16_kernel(
+    const __nv_bfloat16* __restrict__ Q,
+    const __nv_bfloat16* __restrict__ K_cache,
+    const __nv_bfloat16* __restrict__ V_cache,
+    __nv_bfloat16* __restrict__ O,
+    const int* __restrict__ block_tables,
+    const int* __restrict__ context_lens,
+    const __nv_bfloat16* __restrict__ sinks, // [num_q_heads] or NULL
+    int num_q_heads, int num_kv_heads,
+    int head_dim, int max_blocks_per_seq,
+    float scale, int window_size
+) {
+    int seq_idx = blockIdx.y;
+    int q_head  = blockIdx.x;
+    int tid     = threadIdx.x;
+
+    int kv_len = context_lens[seq_idx];
+    if (kv_len <= 0) {
+        if (tid < head_dim) {
+            O[((long long)seq_idx * num_q_heads + q_head) * head_dim + tid] =
+                __float2bfloat16(0.0f);
+        }
+        return;
+    }
+
+    int heads_per_group = num_q_heads / num_kv_heads;
+    int kv_head = q_head / heads_per_group;
+
+    const __nv_bfloat16* Q_ptr = Q +
+        ((long long)seq_idx * num_q_heads + q_head) * head_dim;
+    __nv_bfloat16* O_ptr = O +
+        ((long long)seq_idx * num_q_heads + q_head) * head_dim;
+    const int* bt = block_tables + (long long)seq_idx * max_blocks_per_seq;
+
+    // Sliding window: only attend to positions [kv_len - window_size, kv_len)
+    int start_pos = 0;
+    if (window_size > 0 && kv_len > window_size) {
+        start_pos = kv_len - window_size;
+    }
+
+    float q_reg[PAGED_HEAD_DIM_MAX];
+    for (int d = 0; d < head_dim; d++) {
+        q_reg[d] = __bfloat162float(Q_ptr[d]);
+    }
+
+    float local_max = -INFINITY;
+    float local_sum = 0.0f;
+    float local_O[PAGED_HEAD_DIM_MAX];
+    for (int d = 0; d < head_dim; d++) local_O[d] = 0.0f;
+
+    int kv_stride_block = num_kv_heads * PAGED_BLOCK_SIZE * head_dim;
+    int kv_stride_head  = PAGED_BLOCK_SIZE * head_dim;
+
+    int attend_len = kv_len - start_pos;
+    for (int rel = tid; rel < attend_len; rel += PAGED_THREADS) {
+        int pos = start_pos + rel;
+        int logical_blk = pos / PAGED_BLOCK_SIZE;
+        int slot_in_blk = pos % PAGED_BLOCK_SIZE;
+        int phys_blk = bt[logical_blk];
+
+        const __nv_bfloat16* K_pos = K_cache
+            + (long long)phys_blk * kv_stride_block
+            + kv_head * kv_stride_head
+            + slot_in_blk * head_dim;
+        const __nv_bfloat16* V_pos = V_cache
+            + (long long)phys_blk * kv_stride_block
+            + kv_head * kv_stride_head
+            + slot_in_blk * head_dim;
+
+        float dot = 0.0f;
+        for (int d = 0; d < head_dim; d++) {
+            dot += q_reg[d] * __bfloat162float(K_pos[d]);
+        }
+        float s = dot * scale;
+
+        float new_max = fmaxf(local_max, s);
+        float correction = expf(local_max - new_max);
+        float p = expf(s - new_max);
+
+        local_sum = local_sum * correction + p;
+        for (int d = 0; d < head_dim; d++) local_O[d] *= correction;
+        for (int d = 0; d < head_dim; d++) {
+            local_O[d] += p * __bfloat162float(V_pos[d]);
+        }
+        local_max = new_max;
+    }
+
+    // Include the sink logit (only thread 0 handles it to avoid double-counting)
+    float sink_logit = -INFINITY;
+    if (sinks != nullptr && tid == 0) {
+        sink_logit = __bfloat162float(sinks[q_head]);
+        float new_max = fmaxf(local_max, sink_logit);
+        float correction = expf(local_max - new_max);
+        float p = expf(sink_logit - new_max);
+        local_sum = local_sum * correction + p;
+        for (int d = 0; d < head_dim; d++) local_O[d] *= correction;
+        // Sink absorbs probability but produces no value output (p * 0)
+        local_max = new_max;
+    }
+
+    // ---- Block-level online softmax reduction (same as base kernel) ----
+    __shared__ float smem_max[32];
+    __shared__ float smem_sum[32];
+    __shared__ float smem_O_warp[32][PAGED_HEAD_DIM_MAX];
+
+    int lane = tid & 31;
+    int warp_id = tid >> 5;
+    int num_warps = PAGED_THREADS >> 5;
+
+    float warp_max = local_max;
+    #pragma unroll
+    for (int offset = 16; offset > 0; offset >>= 1)
+        warp_max = fmaxf(warp_max, __shfl_down_sync(0xffffffff, warp_max, offset));
+    if (lane == 0) smem_max[warp_id] = warp_max;
+    __syncthreads();
+
+    float global_max;
+    if (tid == 0) {
+        global_max = smem_max[0];
+        for (int i = 1; i < num_warps; i++)
+            global_max = fmaxf(global_max, smem_max[i]);
+        smem_max[0] = global_max;
+    }
+    __syncthreads();
+    global_max = smem_max[0];
+
+    float rescale = (local_max == -INFINITY) ? 0.0f : expf(local_max - global_max);
+    local_sum *= rescale;
+    for (int d = 0; d < head_dim; d++) local_O[d] *= rescale;
+
+    float warp_sum = local_sum;
+    #pragma unroll
+    for (int offset = 16; offset > 0; offset >>= 1)
+        warp_sum += __shfl_down_sync(0xffffffff, warp_sum, offset);
+    if (lane == 0) smem_sum[warp_id] = warp_sum;
+    __syncthreads();
+
+    float global_sum;
+    if (tid == 0) {
+        global_sum = 0.0f;
+        for (int i = 0; i < num_warps; i++) global_sum += smem_sum[i];
+        smem_sum[0] = global_sum;
+    }
+    __syncthreads();
+    global_sum = smem_sum[0];
+
+    for (int i = tid; i < 32 * PAGED_HEAD_DIM_MAX; i += PAGED_THREADS) {
+        reinterpret_cast<float*>(smem_O_warp)[i] = 0.0f;
+    }
+    __syncthreads();
+
+    for (int d = 0; d < head_dim; d++) {
+        float val = local_O[d];
+        #pragma unroll
+        for (int offset = 16; offset > 0; offset >>= 1)
+            val += __shfl_down_sync(0xffffffff, val, offset);
+        if (lane == 0) smem_O_warp[warp_id][d] = val;
+    }
+    __syncthreads();
+
+    float inv_sum = (global_sum > 0.0f) ? (1.0f / global_sum) : 0.0f;
+    for (int d = tid; d < head_dim; d += PAGED_THREADS) {
+        float out = 0.0f;
+        for (int i = 0; i < num_warps; i++) out += smem_O_warp[i][d];
+        O_ptr[d] = __float2bfloat16(out * inv_sum);
+    }
+}
+
+extern "C" {
+
+void launch_paged_decode_attention_bf16(
+    const void* Q,
+    const void* K_cache,
+    const void* V_cache,
+    void* O,
+    const int* block_tables,
+    const int* context_lens,
+    int batch, int num_q_heads, int num_kv_heads,
+    int head_dim, int max_blocks_per_seq,
+    float scale, void* stream
+) {
+    dim3 grid(num_q_heads, batch);
+    int block = PAGED_THREADS;
+
+    paged_decode_attention_bf16_kernel<<<grid, block, 0, (cudaStream_t)stream>>>(
+        (const __nv_bfloat16*)Q,
+        (const __nv_bfloat16*)K_cache,
+        (const __nv_bfloat16*)V_cache,
+        (__nv_bfloat16*)O,
+        block_tables, context_lens,
+        num_q_heads, num_kv_heads,
+        head_dim, max_blocks_per_seq,
+        scale
+    );
+    CUDA_CHECK_LAST_ERROR();
+}
+
+void launch_paged_decode_attention_tree_bf16(
+    const void* Q,
+    const void* K_cache,
+    const void* V_cache,
+    void* O,
+    const int* block_tables,
+    const int* context_lens,
+    const int* tree_mask,
+    int batch, int num_q_heads, int num_kv_heads,
+    int head_dim, int max_blocks_per_seq,
+    int tree_start, int tree_len,
+    float scale, void* stream
+) {
+    dim3 grid(num_q_heads, batch);
+    int block = PAGED_THREADS;
+
+    paged_decode_attention_tree_bf16_kernel<<<grid, block, 0, (cudaStream_t)stream>>>(
+        (const __nv_bfloat16*)Q,
+        (const __nv_bfloat16*)K_cache,
+        (const __nv_bfloat16*)V_cache,
+        (__nv_bfloat16*)O,
+        block_tables, context_lens, tree_mask,
+        num_q_heads, num_kv_heads,
+        head_dim, max_blocks_per_seq,
+        tree_start, tree_len,
+        scale
+    );
+    CUDA_CHECK_LAST_ERROR();
+}
+
+void launch_paged_decode_attention_sinks_bf16(
+    const void* Q,
+    const void* K_cache,
+    const void* V_cache,
+    void* O,
+    const int* block_tables,
+    const int* context_lens,
+    const void* sinks,
+    int batch, int num_q_heads, int num_kv_heads,
+    int head_dim, int max_blocks_per_seq,
+    float scale, int window_size, void* stream
+) {
+    dim3 grid(num_q_heads, batch);
+    int block = PAGED_THREADS;
+
+    paged_decode_attention_sinks_bf16_kernel<<<grid, block, 0, (cudaStream_t)stream>>>(
+        (const __nv_bfloat16*)Q,
+        (const __nv_bfloat16*)K_cache,
+        (const __nv_bfloat16*)V_cache,
+        (__nv_bfloat16*)O,
+        block_tables, context_lens,
+        (const __nv_bfloat16*)sinks,
+        num_q_heads, num_kv_heads,
+        head_dim, max_blocks_per_seq,
+        scale, window_size
+    );
+    CUDA_CHECK_LAST_ERROR();
+}
+
+}

csrc/attention/reshape_and_cache.cu

@@ -0,0 +1,215 @@
+#include <cuda_bf16.h>
+#include "../common.cuh"
+
+// Scatter [num_tokens] new K/V into a paged KV pool for ONE sequence.
+//
+// Source layouts (BF16, contiguous):
+//   k_src, v_src : [num_kv_heads, num_tokens, head_dim]   (head-major)
+//
+// Pool layouts (BF16, contiguous):
+//   k_pool, v_pool : [num_blocks_total, num_kv_heads, BLOCK_SIZE, head_dim]
+//
+// For token t (0 <= t < num_tokens):
+//   p           = start_pos + t
+//   logical_blk = p / BLOCK_SIZE
+//   slot_in_blk = p % BLOCK_SIZE
+//   phys        = block_ids[logical_blk]
+//   pool[phys, h, slot_in_blk, :] := src[h, t, :]
+//
+// Replaces a Rust-side per-token, per-head cudaMemcpy loop. With Qwen3-8B
+// (8 KV heads, 36 layers) and a 1024-token prefill, that loop fired
+// ~290k device-side memcpys; one kernel launch per layer is dramatically
+// less overhead.
+//
+// Grid : (num_tokens, num_kv_heads)
+// Block: head_dim threads (≤128 in practice; head_dim is padded to a
+//        multiple of 32 by the model and all our shipping configs are
+//        128, so a single warp's worth handles two slots in flight).
+
+__global__ void reshape_and_cache_bf16_kernel(
+    const __nv_bfloat16* __restrict__ k_src,
+    const __nv_bfloat16* __restrict__ v_src,
+    __nv_bfloat16* __restrict__ k_pool,
+    __nv_bfloat16* __restrict__ v_pool,
+    const int* __restrict__ block_ids,
+    int num_tokens, int num_heads,
+    int head_dim, int start_pos, int block_size
+) {
+    int t = blockIdx.x;
+    int h = blockIdx.y;
+    if (t >= num_tokens || h >= num_heads) return;
+
+    int p           = start_pos + t;
+    int logical_blk = p / block_size;
+    int slot_in_blk = p - logical_blk * block_size;
+    int phys        = block_ids[logical_blk];
+
+    long long src_off = ((long long)h * num_tokens + t) * head_dim;
+    long long dst_off = (((long long)phys * num_heads + h) * block_size + slot_in_blk) * head_dim;
+
+    int tid = threadIdx.x;
+    int blockSize = blockDim.x;
+
+    // Per-thread strided copy. head_dim is typically 128 and blockSize is
+    // 128, so each thread copies exactly one element — but the loop keeps
+    // the kernel correct for non-128 head_dim configs (Phi-style 64, etc.).
+    for (int d = tid; d < head_dim; d += blockSize) {
+        k_pool[dst_off + d] = k_src[src_off + d];
+        v_pool[dst_off + d] = v_src[src_off + d];
+    }
+}
+
+// Batched variant: writes one new K/V token per sequence into a paged
+// pool, indexed by a per-batch block table that also drives the paged
+// attention kernel. Used in the decode path where every seq advances
+// by exactly one position per step.
+//
+// Source layouts (BF16, contiguous):
+//   k_src, v_src : [batch, num_kv_heads, head_dim]
+//
+// Pool layouts (BF16, contiguous):
+//   k_pool, v_pool : [num_blocks_total, num_kv_heads, BLOCK_SIZE, head_dim]
+//
+// block_tables : int32 [batch, max_blocks_per_seq]
+// kv_lens      : int32 [batch]   (current seq_len BEFORE this step + 1
+//                                  — i.e. the same buffer paged attention
+//                                  reads. The new token's logical index
+//                                  is `kv_lens[b] - 1`.)
+//
+// Grid : (batch, num_kv_heads)
+// Block: head_dim threads.
+
+__global__ void reshape_and_cache_batched_bf16_kernel(
+    const __nv_bfloat16* __restrict__ k_src,
+    const __nv_bfloat16* __restrict__ v_src,
+    __nv_bfloat16* __restrict__ k_pool,
+    __nv_bfloat16* __restrict__ v_pool,
+    const int* __restrict__ block_tables,
+    const int* __restrict__ kv_lens,
+    int num_heads, int head_dim,
+    int block_size, int max_blocks_per_seq
+) {
+    int b = blockIdx.x;
+    int h = blockIdx.y;
+
+    int new_pos     = kv_lens[b] - 1;
+    int logical_blk = new_pos / block_size;
+    int slot_in_blk = new_pos - logical_blk * block_size;
+    int phys        = block_tables[b * max_blocks_per_seq + logical_blk];
+
+    long long src_off = ((long long)b * num_heads + h) * head_dim;
+    long long dst_off = (((long long)phys * num_heads + h) * block_size + slot_in_blk) * head_dim;
+
+    int tid = threadIdx.x;
+    int blockSize = blockDim.x;
+    for (int d = tid; d < head_dim; d += blockSize) {
+        k_pool[dst_off + d] = k_src[src_off + d];
+        v_pool[dst_off + d] = v_src[src_off + d];
+    }
+}
+
+extern "C" {
+
+void launch_reshape_and_cache_bf16(
+    const void* k_src, const void* v_src,
+    void* k_pool, void* v_pool,
+    const void* block_ids,
+    int num_tokens, int num_heads,
+    int head_dim, int start_pos, int block_size,
+    void* stream
+) {
+    if (num_tokens <= 0) return;
+    int threads = head_dim < 32 ? 32 : head_dim;
+    if (threads > 1024) threads = 1024;
+    dim3 grid(num_tokens, num_heads);
+    reshape_and_cache_bf16_kernel<<<grid, threads, 0, (cudaStream_t)stream>>>(
+        (const __nv_bfloat16*)k_src,
+        (const __nv_bfloat16*)v_src,
+        (__nv_bfloat16*)k_pool,
+        (__nv_bfloat16*)v_pool,
+        (const int*)block_ids,
+        num_tokens, num_heads,
+        head_dim, start_pos, block_size
+    );
+    CUDA_CHECK_LAST_ERROR();
+}
+
+void launch_reshape_and_cache_batched_bf16(
+    const void* k_src, const void* v_src,
+    void* k_pool, void* v_pool,
+    const void* block_tables, const void* kv_lens,
+    int batch, int num_heads,
+    int head_dim, int block_size, int max_blocks_per_seq,
+    void* stream
+) {
+    if (batch <= 0 || num_heads <= 0) return;
+    int threads = head_dim < 32 ? 32 : head_dim;
+    if (threads > 1024) threads = 1024;
+    dim3 grid(batch, num_heads);
+    reshape_and_cache_batched_bf16_kernel<<<grid, threads, 0, (cudaStream_t)stream>>>(
+        (const __nv_bfloat16*)k_src,
+        (const __nv_bfloat16*)v_src,
+        (__nv_bfloat16*)k_pool,
+        (__nv_bfloat16*)v_pool,
+        (const int*)block_tables,
+        (const int*)kv_lens,
+        num_heads, head_dim, block_size, max_blocks_per_seq
+    );
+    CUDA_CHECK_LAST_ERROR();
+}
+
+// Copy one token's K/V from src_pos to dst_pos within one pool.
+// Grid: (num_kv_heads,). Block: head_dim threads.
+// pool: [num_blocks_total, num_kv_heads, block_size, head_dim]
+// block_ids: [max_blocks] for this sequence (logical → physical block map).
+__global__ void copy_kv_position_kernel(
+    __nv_bfloat16* __restrict__ pool,
+    const int* __restrict__ block_ids,
+    int src_pos, int dst_pos,
+    int head_dim, int block_size
+) {
+    int h = blockIdx.x;
+    int d = threadIdx.x;
+    if (d >= head_dim) return;
+
+    int num_kv_heads = gridDim.x;
+
+    int src_blk = src_pos / block_size;
+    int src_slot = src_pos % block_size;
+    int src_phys = block_ids[src_blk];
+
+    int dst_blk = dst_pos / block_size;
+    int dst_slot = dst_pos % block_size;
+    int dst_phys = block_ids[dst_blk];
+
+    long long src_off = ((long long)src_phys * num_kv_heads + h) * block_size * head_dim
+                      + src_slot * head_dim + d;
+    long long dst_off = ((long long)dst_phys * num_kv_heads + h) * block_size * head_dim
+                      + dst_slot * head_dim + d;
+
+    pool[dst_off] = pool[src_off];
+}
+
+void launch_copy_kv_position(
+    void* k_pool, void* v_pool,
+    const int* block_ids,
+    int src_pos, int dst_pos,
+    int num_kv_heads, int head_dim, int block_size,
+    void* stream
+) {
+    int threads = head_dim < 32 ? 32 : head_dim;
+    if (threads > 1024) threads = 1024;
+    dim3 grid(num_kv_heads);
+    copy_kv_position_kernel<<<grid, threads, 0, (cudaStream_t)stream>>>(
+        (__nv_bfloat16*)k_pool, block_ids,
+        src_pos, dst_pos, head_dim, block_size
+    );
+    CUDA_CHECK_LAST_ERROR();
+    copy_kv_position_kernel<<<grid, threads, 0, (cudaStream_t)stream>>>(
+        (__nv_bfloat16*)v_pool, block_ids,
+        src_pos, dst_pos, head_dim, block_size
+    );
+    CUDA_CHECK_LAST_ERROR();
+}
+
+}

csrc/common.cuh

@@ -48,3 +48,18 @@ __device__ __forceinline__ float block_reduce_max(float val) {
     if (warp_id == 0) val = warp_reduce_max(val);
     return val;
 }
+
+// --- Launch error checking ---
+// Always on, including release builds. A launch with an invalid config
+// (e.g. 32-bit overflow in grid/index math) is otherwise silent and produces
+// garbage with no clue — the MoE int32-overflow bug was found exactly because
+// release swallowed the launch failure. `cudaGetLastError()` does not
+// synchronize the stream, so the per-launch host cost is negligible.
+#include <cstdio>
+#define CUDA_CHECK_LAST_ERROR() do { \
+    cudaError_t err = cudaGetLastError(); \
+    if (err != cudaSuccess) { \
+        fprintf(stderr, "CUDA kernel launch error at %s:%d: %s\n", \
+                __FILE__, __LINE__, cudaGetErrorString(err)); \
+    } \
+} while(0)

csrc/embedding/embedding.cu

@@ -1,4 +1,5 @@
 #include <cuda_bf16.h>
+#include "../common.cuh"
 
 // Embedding lookup: out[seq_idx] = table[token_ids[seq_idx]]
 // Grid: num_tokens, Block: handles hidden_size elements per token.
@@ -7,10 +8,12 @@ __global__ void embedding_f32(
     const float* __restrict__ table,    // [vocab_size, hidden_size]
     const int* __restrict__ token_ids,  // [num_tokens]
     float* __restrict__ out,            // [num_tokens, hidden_size]
-    int hidden_size
+    int hidden_size,
+    int vocab_size
 ) {
     int token_idx = blockIdx.x;
     int tid = token_ids[token_idx];
+    if (tid < 0 || tid >= vocab_size) return;
     const float* row = table + tid * hidden_size;
     float* dst = out + token_idx * hidden_size;
 
@@ -23,10 +26,12 @@ __global__ void embedding_bf16(
     const __nv_bfloat16* __restrict__ table,
     const int* __restrict__ token_ids,
     __nv_bfloat16* __restrict__ out,
-    int hidden_size
+    int hidden_size,
+    int vocab_size
 ) {
     int token_idx = blockIdx.x;
     int tid = token_ids[token_idx];
+    if (tid < 0 || tid >= vocab_size) return;
     const __nv_bfloat16* row = table + tid * hidden_size;
     __nv_bfloat16* dst = out + token_idx * hidden_size;
 
@@ -38,18 +43,20 @@ __global__ void embedding_bf16(
 extern "C" {
 
 void launch_embedding_f32(const void* table, const void* token_ids, void* out,
-                          int num_tokens, int hidden_size, void* stream) {
+                          int num_tokens, int hidden_size, int vocab_size, void* stream) {
     int block = (hidden_size < 256) ? hidden_size : 256;
     embedding_f32<<<num_tokens, block, 0, (cudaStream_t)stream>>>(
-        (const float*)table, (const int*)token_ids, (float*)out, hidden_size);
+        (const float*)table, (const int*)token_ids, (float*)out, hidden_size, vocab_size);
+    CUDA_CHECK_LAST_ERROR();
 }
 
 void launch_embedding_bf16(const void* table, const void* token_ids, void* out,
-                           int num_tokens, int hidden_size, void* stream) {
+                           int num_tokens, int hidden_size, int vocab_size, void* stream) {
     int block = (hidden_size < 256) ? hidden_size : 256;
     embedding_bf16<<<num_tokens, block, 0, (cudaStream_t)stream>>>(
         (const __nv_bfloat16*)table, (const int*)token_ids,
-        (__nv_bfloat16*)out, hidden_size);
+        (__nv_bfloat16*)out, hidden_size, vocab_size);
+    CUDA_CHECK_LAST_ERROR();
 }
 
 }

csrc/embedding/rope.cu

@@ -1,10 +1,11 @@
 #include <cuda_bf16.h>
 #include <math.h>
+#include "../common.cuh"
 
-// RoPE: Rotary Position Embedding
-// For each pair (x[2i], x[2i+1]) at position `pos`:
-//   y[2i]   = x[2i] * cos - x[2i+1] * sin
-//   y[2i+1] = x[2i] * sin + x[2i+1] * cos
+// RoPE: Rotary Position Embedding, using the Qwen/Llama rotate_half layout.
+// For each dimension i in the first half at position `pos`:
+//   y[i]            = x[i] * cos - x[i + half_dim] * sin
+//   y[i + half_dim] = x[i + half_dim] * cos + x[i] * sin
 // where cos/sin come from precomputed cos_cache/sin_cache.
 //
 // cos_cache[pos][i] = cos(pos * freq[i])
@@ -35,11 +36,11 @@ __global__ void rope_f32(
     float sin_val = sin_cache[pos * half_dim + pair_idx];
 
     int base = (token_idx * num_heads + head_idx) * head_dim;
-    float x0 = x[base + 2 * pair_idx];
-    float x1 = x[base + 2 * pair_idx + 1];
+    float x0 = x[base + pair_idx];
+    float x1 = x[base + pair_idx + half_dim];
 
-    x[base + 2 * pair_idx]     = x0 * cos_val - x1 * sin_val;
-    x[base + 2 * pair_idx + 1] = x0 * sin_val + x1 * cos_val;
+    x[base + pair_idx]            = x0 * cos_val - x1 * sin_val;
+    x[base + pair_idx + half_dim] = x1 * cos_val + x0 * sin_val;
 }
 
 __global__ void rope_bf16(
@@ -61,11 +62,11 @@ __global__ void rope_bf16(
     float sin_val = sin_cache[pos * half_dim + pair_idx];
 
     int base = (token_idx * num_heads + head_idx) * head_dim;
-    float x0 = __bfloat162float(x[base + 2 * pair_idx]);
-    float x1 = __bfloat162float(x[base + 2 * pair_idx + 1]);
+    float x0 = __bfloat162float(x[base + pair_idx]);
+    float x1 = __bfloat162float(x[base + pair_idx + half_dim]);
 
-    x[base + 2 * pair_idx]     = __float2bfloat16(x0 * cos_val - x1 * sin_val);
-    x[base + 2 * pair_idx + 1] = __float2bfloat16(x0 * sin_val + x1 * cos_val);
+    x[base + pair_idx]            = __float2bfloat16(x0 * cos_val - x1 * sin_val);
+    x[base + pair_idx + half_dim] = __float2bfloat16(x1 * cos_val + x0 * sin_val);
 }
 
 // Precompute cos/sin cache on GPU
@@ -94,6 +95,7 @@ void launch_rope_f32(void* x, const void* cos_cache, const void* sin_cache,
     rope_f32<<<grid, block, 0, (cudaStream_t)stream>>>(
         (float*)x, (const float*)cos_cache, (const float*)sin_cache,
         (const int*)positions, num_heads, head_dim);
+    CUDA_CHECK_LAST_ERROR();
 }
 
 void launch_rope_bf16(void* x, const void* cos_cache, const void* sin_cache,
@@ -104,6 +106,7 @@ void launch_rope_bf16(void* x, const void* cos_cache, const void* sin_cache,
     rope_bf16<<<grid, block, 0, (cudaStream_t)stream>>>(
         (__nv_bfloat16*)x, (const float*)cos_cache, (const float*)sin_cache,
         (const int*)positions, num_heads, head_dim);
+    CUDA_CHECK_LAST_ERROR();
 }
 
 void launch_compute_rope_cache(void* cos_cache, void* sin_cache,
@@ -111,6 +114,7 @@ void launch_compute_rope_cache(void* cos_cache, void* sin_cache,
                                void* stream) {
     compute_rope_cache<<<max_seq_len, half_dim, 0, (cudaStream_t)stream>>>(
         (float*)cos_cache, (float*)sin_cache, max_seq_len, half_dim, theta);
+    CUDA_CHECK_LAST_ERROR();
 }
 
 }

csrc/embedding/transpose.cu

@@ -0,0 +1,242 @@
+#include <cuda_bf16.h>
+#include "../common.cuh"
+
+// Transpose between [S, H, D] and [H, S, D] layouts (used for RoPE and attention).
+// Also handles [S, H*D] → [H, S, D] (reshape_heads) and reverse (merge_heads).
+
+// reshape_heads: [S, H*D] → [1, H, S, D]
+// Input layout:  element at [s, h*D + d] = flat[s * H*D + h*D + d]
+// Output layout: element at [0, h, s, d] = flat[h * S*D + s*D + d]
+__global__ void reshape_heads_bf16(
+    const __nv_bfloat16* __restrict__ in,
+    __nv_bfloat16* __restrict__ out,
+    int seq_len, int num_heads, int head_dim
+) {
+    int hidden = num_heads * head_dim;
+    int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    int total = seq_len * hidden;
+    if (idx >= total) return;
+
+    int s = idx / hidden;
+    int rem = idx % hidden;
+    int h = rem / head_dim;
+    int d = rem % head_dim;
+
+    int out_idx = h * seq_len * head_dim + s * head_dim + d;
+    out[out_idx] = in[idx];
+}
+
+// merge_heads: [1, H, S, D] → [S, H*D]
+// Input layout:  element at [0, h, s, d] = flat[h * S*D + s*D + d]
+// Output layout: element at [s, h*D + d] = flat[s * H*D + h*D + d]
+__global__ void merge_heads_bf16(
+    const __nv_bfloat16* __restrict__ in,
+    __nv_bfloat16* __restrict__ out,
+    int seq_len, int num_heads, int head_dim
+) {
+    int hidden = num_heads * head_dim;
+    int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    int total = seq_len * hidden;
+    if (idx >= total) return;
+
+    // idx is output index: [s, h*D + d]
+    int s = idx / hidden;
+    int rem = idx % hidden;
+    int h = rem / head_dim;
+    int d = rem % head_dim;
+
+    int in_idx = h * seq_len * head_dim + s * head_dim + d;
+    out[idx] = in[in_idx];
+}
+
+// transpose_for_rope: [1, H, S, D] → [S, H, D]
+// Input:  [h, s, d] at h*S*D + s*D + d
+// Output: [s, h, d] at s*H*D + h*D + d
+__global__ void transpose_hsd_to_shd_bf16(
+    const __nv_bfloat16* __restrict__ in,
+    __nv_bfloat16* __restrict__ out,
+    int seq_len, int num_heads, int head_dim
+) {
+    int total = seq_len * num_heads * head_dim;
+    int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    if (idx >= total) return;
+
+    // idx = output flat index: s*H*D + h*D + d
+    int s = idx / (num_heads * head_dim);
+    int rem = idx % (num_heads * head_dim);
+    int h = rem / head_dim;
+    int d = rem % head_dim;
+
+    int in_idx = h * seq_len * head_dim + s * head_dim + d;
+    out[idx] = in[in_idx];
+}
+
+// transpose_from_rope: [S, H, D] → [1, H, S, D]
+// Input:  [s, h, d] at s*H*D + h*D + d
+// Output: [h, s, d] at h*S*D + s*D + d
+__global__ void transpose_shd_to_hsd_bf16(
+    const __nv_bfloat16* __restrict__ in,
+    __nv_bfloat16* __restrict__ out,
+    int seq_len, int num_heads, int head_dim
+) {
+    int total = seq_len * num_heads * head_dim;
+    int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    if (idx >= total) return;
+
+    // idx = output flat index: h*S*D + s*D + d
+    int h = idx / (seq_len * head_dim);
+    int rem = idx % (seq_len * head_dim);
+    int s = rem / head_dim;
+    int d = rem % head_dim;
+
+    int in_idx = s * num_heads * head_dim + h * head_dim + d;
+    out[idx] = in[in_idx];
+}
+
+// repeat_kv: [1, KV_H, S, D] → [1, KV_H * n_rep, S, D]
+__global__ void repeat_kv_bf16(
+    const __nv_bfloat16* __restrict__ in,
+    __nv_bfloat16* __restrict__ out,
+    int kv_heads, int n_rep, int seq_len, int head_dim
+) {
+    int total_heads = kv_heads * n_rep;
+    int total = total_heads * seq_len * head_dim;
+    int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    if (idx >= total) return;
+
+    int out_h = idx / (seq_len * head_dim);
+    int rem = idx % (seq_len * head_dim);
+    int kv_h = out_h / n_rep;
+
+    int in_idx = kv_h * seq_len * head_dim + rem;
+    out[idx] = in[in_idx];
+}
+
+// ---- Generic strided copy (up to 4D) ----
+// Each thread copies one element. Maps flat contiguous output index to strided input index.
+// Unused dimensions are padded with shape=1, stride=0.
+
+__global__ void strided_copy_bf16(
+    const __nv_bfloat16* __restrict__ in,
+    __nv_bfloat16* __restrict__ out,
+    int numel,
+    int ndim,
+    int shape0, int shape1, int shape2, int shape3,
+    int in_stride0, int in_stride1, int in_stride2, int in_stride3,
+    int in_offset
+) {
+    int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    if (idx >= numel) return;
+
+    // Decompose flat output index into multi-dim indices (rightmost = fastest)
+    int remaining = idx;
+    int i3 = remaining % shape3; remaining /= shape3;
+    int i2 = remaining % shape2; remaining /= shape2;
+    int i1 = remaining % shape1; remaining /= shape1;
+    int i0 = remaining;
+
+    int in_idx = in_offset + i0 * in_stride0 + i1 * in_stride1 + i2 * in_stride2 + i3 * in_stride3;
+    out[idx] = in[in_idx];
+}
+
+__global__ void strided_copy_f32(
+    const float* __restrict__ in,
+    float* __restrict__ out,
+    int numel,
+    int ndim,
+    int shape0, int shape1, int shape2, int shape3,
+    int in_stride0, int in_stride1, int in_stride2, int in_stride3,
+    int in_offset
+) {
+    int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    if (idx >= numel) return;
+
+    int remaining = idx;
+    int i3 = remaining % shape3; remaining /= shape3;
+    int i2 = remaining % shape2; remaining /= shape2;
+    int i1 = remaining % shape1; remaining /= shape1;
+    int i0 = remaining;
+
+    int in_idx = in_offset + i0 * in_stride0 + i1 * in_stride1 + i2 * in_stride2 + i3 * in_stride3;
+    out[idx] = in[in_idx];
+}
+
+extern "C" {
+
+void launch_reshape_heads_bf16(const void* in, void* out,
+                               int seq_len, int num_heads, int head_dim, void* stream) {
+    int total = seq_len * num_heads * head_dim;
+    int block = 256;
+    int grid = (total + block - 1) / block;
+    reshape_heads_bf16<<<grid, block, 0, (cudaStream_t)stream>>>(
+        (const __nv_bfloat16*)in, (__nv_bfloat16*)out, seq_len, num_heads, head_dim);
+    CUDA_CHECK_LAST_ERROR();
+}
+
+void launch_merge_heads_bf16(const void* in, void* out,
+                             int seq_len, int num_heads, int head_dim, void* stream) {
+    int total = seq_len * num_heads * head_dim;
+    int block = 256;
+    int grid = (total + block - 1) / block;
+    merge_heads_bf16<<<grid, block, 0, (cudaStream_t)stream>>>(
+        (const __nv_bfloat16*)in, (__nv_bfloat16*)out, seq_len, num_heads, head_dim);
+    CUDA_CHECK_LAST_ERROR();
+}
+
+void launch_transpose_hsd_to_shd_bf16(const void* in, void* out,
+                                       int seq_len, int num_heads, int head_dim, void* stream) {
+    int total = seq_len * num_heads * head_dim;
+    int block = 256;
+    int grid = (total + block - 1) / block;
+    transpose_hsd_to_shd_bf16<<<grid, block, 0, (cudaStream_t)stream>>>(
+        (const __nv_bfloat16*)in, (__nv_bfloat16*)out, seq_len, num_heads, head_dim);
+    CUDA_CHECK_LAST_ERROR();
+}
+
+void launch_transpose_shd_to_hsd_bf16(const void* in, void* out,
+                                       int seq_len, int num_heads, int head_dim, void* stream) {
+    int total = seq_len * num_heads * head_dim;
+    int block = 256;
+    int grid = (total + block - 1) / block;
+    transpose_shd_to_hsd_bf16<<<grid, block, 0, (cudaStream_t)stream>>>(
+        (const __nv_bfloat16*)in, (__nv_bfloat16*)out, seq_len, num_heads, head_dim);
+    CUDA_CHECK_LAST_ERROR();
+}
+
+void launch_repeat_kv_bf16(const void* in, void* out,
+                           int kv_heads, int n_rep, int seq_len, int head_dim, void* stream) {
+    int total = kv_heads * n_rep * seq_len * head_dim;
+    int block = 256;
+    int grid = (total + block - 1) / block;
+    repeat_kv_bf16<<<grid, block, 0, (cudaStream_t)stream>>>(
+        (const __nv_bfloat16*)in, (__nv_bfloat16*)out, kv_heads, n_rep, seq_len, head_dim);
+    CUDA_CHECK_LAST_ERROR();
+}
+
+void launch_strided_copy_bf16(const void* in, void* out, int numel, int ndim,
+                              int shape0, int shape1, int shape2, int shape3,
+                              int in_stride0, int in_stride1, int in_stride2, int in_stride3,
+                              int in_offset, void* stream) {
+    int block = 256;
+    int grid = (numel + block - 1) / block;
+    strided_copy_bf16<<<grid, block, 0, (cudaStream_t)stream>>>(
+        (const __nv_bfloat16*)in, (__nv_bfloat16*)out, numel, ndim,
+        shape0, shape1, shape2, shape3,
+        in_stride0, in_stride1, in_stride2, in_stride3, in_offset);
+    CUDA_CHECK_LAST_ERROR();
+}
+
+void launch_strided_copy_f32(const void* in, void* out, int numel, int ndim,
+                             int shape0, int shape1, int shape2, int shape3,
+                             int in_stride0, int in_stride1, int in_stride2, int in_stride3,
+                             int in_offset, void* stream) {
+    int block = 256;
+    int grid = (numel + block - 1) / block;
+    strided_copy_f32<<<grid, block, 0, (cudaStream_t)stream>>>(
+        (const float*)in, (float*)out, numel, ndim,
+        shape0, shape1, shape2, shape3,
+        in_stride0, in_stride1, in_stride2, in_stride3, in_offset);
+    CUDA_CHECK_LAST_ERROR();
+}
+
+}

csrc/gemm/gemv.cu

@@ -0,0 +1,196 @@
+#include <cuda_bf16.h>
+#include <cuda_runtime.h>
+#include "../common.cuh"
+
+// K-split GEMV for M=1 BF16 decode.
+//
+//   y[n] = sum_k x[k] * W[k * N + n]
+//
+// Grid: (N / TILE_N, K / TILE_K) partials, followed by a deterministic
+// fixed-order reduction over K blocks. The previous implementation used
+// atomicAdd into y_fp32[col]; that made BF16 greedy decode sensitive to
+// inter-block scheduling when logits were close.
+
+#define GEMV_TILE_N 128
+#define GEMV_TILE_K 256
+#define GEMV_BLOCK  128
+
+__global__ void gemv_bf16_partial_kernel(
+    const __nv_bfloat16* __restrict__ x,
+    const __nv_bfloat16* __restrict__ W,
+    float* __restrict__ partials,
+    int K, int N
+) {
+    const int block_n = blockIdx.x;
+    const int block_k = blockIdx.y;
+    const int t = threadIdx.x;
+    const int col = block_n * GEMV_TILE_N + t;
+
+    const int k_start = block_k * GEMV_TILE_K;
+    const int k_end = min(k_start + GEMV_TILE_K, K);
+    const int k_len = k_end - k_start;
+
+    // Cooperative load of x into shared memory uses ALL threads in the block
+    // (indexed by t, independent of col). Threads whose column is out of range
+    // must still help load and reach the barrier — returning early here would
+    // leave part of x_shared uninitialized AND make __syncthreads divergent
+    // (UB). So the col>=N check happens only AFTER the load + barrier. This bug
+    // produced intermittent huge/garbage outputs whenever N % GEMV_TILE_N != 0
+    // (e.g. gpt-oss decode o_proj with N=2880), collapsing the forward pass.
+    __shared__ float x_shared[GEMV_TILE_K];
+    for (int i = t; i < k_len; i += GEMV_BLOCK) {
+        x_shared[i] = __bfloat162float(x[k_start + i]);
+    }
+    __syncthreads();
+
+    if (col >= N) return;
+
+    float sum = 0.0f;
+    for (int ki = 0; ki < k_len; ki++) {
+        sum += x_shared[ki] * __bfloat162float(W[(long long)(k_start + ki) * N + col]);
+    }
+
+    partials[(long long)block_k * N + col] = sum;
+}
+
+__global__ void gemv_reduce_to_bf16_kernel(
+    const float* __restrict__ partials,
+    __nv_bfloat16* __restrict__ dst,
+    int n,
+    int num_k_blocks
+) {
+    int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    if (idx < n) {
+        float sum = 0.0f;
+        for (int kb = 0; kb < num_k_blocks; kb++) {
+            sum += partials[(long long)kb * n + idx];
+        }
+        dst[idx] = __float2bfloat16(sum);
+    }
+}
+
+// Batched variant: M rows, same W. Grid.z = batch row index.
+// Numerically identical to calling launch_gemv_bf16 M times in sequence because
+// each z-slice executes the same accumulation order on the same data.
+// partials buffer must be [M * num_k_blocks * N] floats.
+__global__ void gemv_bf16_batched_partial_kernel(
+    const __nv_bfloat16* __restrict__ x,  // [M, K]
+    const __nv_bfloat16* __restrict__ W,  // [K, N]
+    float* __restrict__ partials,         // [M, num_k_blocks, N]
+    int K, int N
+) {
+    const int block_n = blockIdx.x;
+    const int block_k = blockIdx.y;
+    const int row = blockIdx.z;
+    const int t = threadIdx.x;
+    const int col = block_n * GEMV_TILE_N + t;
+
+    const int k_start = block_k * GEMV_TILE_K;
+    const int k_end = min(k_start + GEMV_TILE_K, K);
+    const int k_len = k_end - k_start;
+
+    __shared__ float x_shared[GEMV_TILE_K];
+    const __nv_bfloat16* x_row = x + (long long)row * K;
+    for (int i = t; i < k_len; i += GEMV_BLOCK) {
+        x_shared[i] = __bfloat162float(x_row[k_start + i]);
+    }
+    __syncthreads();
+
+    if (col >= N) return;
+
+    float sum = 0.0f;
+    for (int ki = 0; ki < k_len; ki++) {
+        sum += x_shared[ki] * __bfloat162float(W[(long long)(k_start + ki) * N + col]);
+    }
+
+    int num_k_blocks = (K + GEMV_TILE_K - 1) / GEMV_TILE_K;
+    partials[((long long)row * num_k_blocks + block_k) * N + col] = sum;
+}
+
+__global__ void gemv_batched_reduce_to_bf16_kernel(
+    const float* __restrict__ partials,  // [M, num_k_blocks, N]
+    __nv_bfloat16* __restrict__ dst,     // [M, N]
+    int n,
+    int num_k_blocks
+) {
+    int col = blockIdx.x * blockDim.x + threadIdx.x;
+    int row = blockIdx.y;
+    if (col >= n) return;
+
+    float sum = 0.0f;
+    const float* row_partials = partials + (long long)row * num_k_blocks * n;
+    for (int kb = 0; kb < num_k_blocks; kb++) {
+        sum += row_partials[(long long)kb * n + col];
+    }
+    dst[(long long)row * n + col] = __float2bfloat16(sum);
+}
+
+extern "C" {
+
+void launch_gemv_bf16(
+    const void* x,
+    const void* W,
+    void* y_bf16,
+    void* y_fp32_buf,
+    int K, int N,
+    void* stream
+) {
+    cudaStream_t s = (cudaStream_t)stream;
+
+    int num_k_blocks = (K + GEMV_TILE_K - 1) / GEMV_TILE_K;
+    dim3 grid((N + GEMV_TILE_N - 1) / GEMV_TILE_N, num_k_blocks);
+
+    gemv_bf16_partial_kernel<<<grid, GEMV_BLOCK, 0, s>>>(
+        (const __nv_bfloat16*)x,
+        (const __nv_bfloat16*)W,
+        (float*)y_fp32_buf,
+        K, N
+    );
+    CUDA_CHECK_LAST_ERROR();
+
+    // Fixed-order FP32 reduction over K blocks, then BF16 conversion.
+    int conv_block = 256;
+    int conv_grid = (N + conv_block - 1) / conv_block;
+    gemv_reduce_to_bf16_kernel<<<conv_grid, conv_block, 0, s>>>(
+        (const float*)y_fp32_buf,
+        (__nv_bfloat16*)y_bf16,
+        N,
+        num_k_blocks
+    );
+    CUDA_CHECK_LAST_ERROR();
+}
+
+void launch_gemv_bf16_batched(
+    const void* x,     // [M, K] BF16
+    const void* W,     // [K, N] BF16
+    void* y_bf16,      // [M, N] BF16
+    void* y_fp32_buf,  // [M * num_k_blocks * N] FP32
+    int M, int K, int N,
+    void* stream
+) {
+    cudaStream_t s = (cudaStream_t)stream;
+
+    int num_k_blocks = (K + GEMV_TILE_K - 1) / GEMV_TILE_K;
+    dim3 grid((N + GEMV_TILE_N - 1) / GEMV_TILE_N, num_k_blocks, M);
+
+    gemv_bf16_batched_partial_kernel<<<grid, GEMV_BLOCK, 0, s>>>(
+        (const __nv_bfloat16*)x,
+        (const __nv_bfloat16*)W,
+        (float*)y_fp32_buf,
+        K, N
+    );
+    CUDA_CHECK_LAST_ERROR();
+
+    int conv_block = 256;
+    int conv_grid_x = (N + conv_block - 1) / conv_block;
+    dim3 reduce_grid(conv_grid_x, M);
+    gemv_batched_reduce_to_bf16_kernel<<<reduce_grid, conv_block, 0, s>>>(
+        (const float*)y_fp32_buf,
+        (__nv_bfloat16*)y_bf16,
+        N,
+        num_k_blocks
+    );
+    CUDA_CHECK_LAST_ERROR();
+}
+
+} // extern "C"

csrc/gemm/naive.cu

@@ -1,4 +1,5 @@
 #include <cuda_bf16.h>
+#include "../common.cuh"
 
 // Naive GEMM: each thread computes one element of C.
 // C[i][j] = sum_k A[i][k] * B[k][j]
@@ -46,6 +47,7 @@ void launch_gemm_naive_bf16(
     gemm_naive_bf16<<<grid, block, 0, (cudaStream_t)stream>>>(
         (const __nv_bfloat16*)A, (const __nv_bfloat16*)B, (__nv_bfloat16*)C, M, N, K
     );
+    CUDA_CHECK_LAST_ERROR();
 }
 
 void launch_gemm_naive_f32(
@@ -57,6 +59,7 @@ void launch_gemm_naive_f32(
     gemm_naive_f32<<<grid, block, 0, (cudaStream_t)stream>>>(
         (const float*)A, (const float*)B, (float*)C, M, N, K
     );
+    CUDA_CHECK_LAST_ERROR();
 }
 
 } // extern "C"

csrc/gemm/tiled.cu

@@ -1,4 +1,5 @@
 #include <cuda_bf16.h>
+#include "../common.cuh"
 
 // Tiled GEMM using shared memory.
 // Each thread block loads TILE_SIZE x TILE_SIZE tiles of A and B
@@ -100,6 +101,7 @@ void launch_gemm_tiled_f32(
     gemm_tiled_f32<<<grid, block, 0, (cudaStream_t)stream>>>(
         (const float*)A, (const float*)B, (float*)C, M, N, K
     );
+    CUDA_CHECK_LAST_ERROR();
 }
 
 void launch_gemm_tiled_bf16(
@@ -111,6 +113,7 @@ void launch_gemm_tiled_bf16(
     gemm_tiled_bf16<<<grid, block, 0, (cudaStream_t)stream>>>(
         (const __nv_bfloat16*)A, (const __nv_bfloat16*)B, (__nv_bfloat16*)C, M, N, K
     );
+    CUDA_CHECK_LAST_ERROR();
 }
 
 } // extern "C"

csrc/moe/moe_kernels.cu

@@ -0,0 +1,254 @@
+#include <cuda_bf16.h>
+#include <float.h>
+#include "../common.cuh"
+
+// ============================================================
+// MoE Top-K + Softmax kernel
+//
+// Input:  router_logits [num_tokens, num_experts]  BF16
+// Output: topk_ids      [num_tokens, top_k]        int32
+//         topk_weights  [num_tokens, top_k]         float32
+//
+// One block per token. Threads cooperatively find top-k indices
+// via repeated argmax, then compute softmax over the k winners.
+// num_experts <= 256 (fits in registers / shared memory).
+// ============================================================
+
+#define MOE_MAX_EXPERTS 256
+#define MOE_MAX_TOPK 8
+
+__global__ void moe_topk_softmax_bf16_kernel(
+    const __nv_bfloat16* __restrict__ router_logits,
+    int* __restrict__ topk_ids,
+    float* __restrict__ topk_weights,
+    int num_experts, int top_k
+) {
+    int token = blockIdx.x;
+    int tid = threadIdx.x;
+    const __nv_bfloat16* row = router_logits + token * num_experts;
+
+    // Load logits into shared memory
+    __shared__ float smem_logits[MOE_MAX_EXPERTS];
+    __shared__ int smem_ids[MOE_MAX_TOPK];
+    __shared__ float smem_vals[MOE_MAX_TOPK];
+
+    for (int i = tid; i < num_experts; i += blockDim.x) {
+        smem_logits[i] = __bfloat162float(row[i]);
+    }
+    __syncthreads();
+
+    // Find top-k via repeated argmax (k is small, typically 4)
+    if (tid == 0) {
+        for (int k = 0; k < top_k; k++) {
+            float best_val = -INFINITY;
+            int best_idx = 0;
+            for (int e = 0; e < num_experts; e++) {
+                if (smem_logits[e] > best_val) {
+                    best_val = smem_logits[e];
+                    best_idx = e;
+                }
+            }
+            smem_ids[k] = best_idx;
+            smem_vals[k] = best_val;
+            smem_logits[best_idx] = -INFINITY; // mask out selected
+        }
+
+        // Softmax over top-k values (in FP32)
+        float max_val = smem_vals[0];
+        for (int k = 1; k < top_k; k++)
+            max_val = fmaxf(max_val, smem_vals[k]);
+
+        float exp_sum = 0.0f;
+        for (int k = 0; k < top_k; k++) {
+            smem_vals[k] = expf(smem_vals[k] - max_val);
+            exp_sum += smem_vals[k];
+        }
+        float inv_sum = 1.0f / exp_sum;
+        for (int k = 0; k < top_k; k++)
+            smem_vals[k] *= inv_sum;
+
+        // Write outputs
+        for (int k = 0; k < top_k; k++) {
+            topk_ids[token * top_k + k] = smem_ids[k];
+            topk_weights[token * top_k + k] = smem_vals[k];
+        }
+    }
+}
+
+// ============================================================
+// MoE Replicate kernel
+//
+// Input:  x     [num_tokens, hidden]               BF16
+// Output: x_rep [local_experts, num_tokens, hidden] BF16
+//
+// Copies x into each expert's batch slot.
+// ============================================================
+
+__global__ void moe_replicate_bf16_kernel(
+    const __nv_bfloat16* __restrict__ x,
+    __nv_bfloat16* __restrict__ x_rep,
+    int num_tokens, int hidden, int local_experts
+) {
+    // 64-bit index: local_experts * num_tokens * hidden overflows int32 at
+    // ~2.3k prefill tokens (gpt-oss TP=1, 32 experts), which is inside the
+    // supported context window. A 32-bit `total` silently wraps, the launch
+    // fails, and (in release) the error is invisible — see common.cuh.
+    long long idx = (long long)blockIdx.x * blockDim.x + threadIdx.x;
+    long long total = (long long)local_experts * num_tokens * hidden;
+    if (idx >= total) return;
+
+    // x_rep[expert, token, dim] = x[token, dim]
+    long long row_stride = (long long)num_tokens * hidden;
+    x_rep[idx] = x[idx % row_stride];
+}
+
+// ============================================================
+// MoE Bias Add 3D kernel
+//
+// Input:  x    [batch, num_tokens, dim]  BF16 (in-place output)
+//         bias [batch, dim]              BF16
+//
+// x[b, t, d] += bias[b, d]
+// ============================================================
+
+__global__ void moe_bias_add_3d_bf16_kernel(
+    __nv_bfloat16* __restrict__ x,
+    const __nv_bfloat16* __restrict__ bias,
+    int batch, int num_tokens, int dim
+) {
+    // 64-bit index: batch * num_tokens * dim overflows int32 at ~3.6k prefill
+    // tokens (gpt-oss TP=1, 32 experts, 2*intermediate dim) — see moe_replicate.
+    long long idx = (long long)blockIdx.x * blockDim.x + threadIdx.x;
+    long long total = (long long)batch * num_tokens * dim;
+    if (idx >= total) return;
+
+    long long td = (long long)num_tokens * dim;
+    int b = (int)(idx / td);   // < batch (small)
+    int d = (int)(idx % dim);  // < dim
+    float v = __bfloat162float(x[idx]) + __bfloat162float(bias[(long long)b * dim + d]);
+    x[idx] = __float2bfloat16(v);
+}
+
+// ============================================================
+// MoE Weighted Sum kernel
+//
+// Input:  expert_out  [local_experts, num_tokens, hidden]  BF16
+//         topk_ids    [num_tokens, top_k]                  int32 (global expert ids)
+//         topk_weights[num_tokens, top_k]                  float32
+//         expert_start: first global expert id this rank owns
+//         local_experts: number of experts this rank owns
+//
+// Output: out [num_tokens, hidden] BF16
+//
+// For each (token, dim): accumulate in FP32:
+//   sum = 0
+//   for k in 0..top_k:
+//     global_id = topk_ids[token, k]
+//     if global_id in [expert_start, expert_start + local_experts):
+//       local_id = global_id - expert_start
+//       sum += topk_weights[token, k] * expert_out[local_id, token, dim]
+//   out[token, dim] = bf16(sum)
+// ============================================================
+
+__global__ void moe_weighted_sum_bf16_kernel(
+    const __nv_bfloat16* __restrict__ expert_out,
+    const int* __restrict__ topk_ids,
+    const float* __restrict__ topk_weights,
+    __nv_bfloat16* __restrict__ out,
+    int num_tokens, int hidden, int top_k,
+    int expert_start, int local_experts
+) {
+    // 64-bit index: `local_id * expert_stride` overflows int32 for long prefills
+    // (expert_stride = num_tokens * hidden), reading the wrong expert element.
+    long long idx = (long long)blockIdx.x * blockDim.x + threadIdx.x;
+    long long total = (long long)num_tokens * hidden;
+    if (idx >= total) return;
+
+    long long token = idx / hidden;
+    int dim = (int)(idx % hidden);
+
+    long long expert_stride = (long long)num_tokens * hidden; // stride between experts in expert_out
+
+    float sum = 0.0f;
+    for (int k = 0; k < top_k; k++) {
+        int global_id = topk_ids[token * top_k + k];
+        int local_id = global_id - expert_start;
+        if (local_id >= 0 && local_id < local_experts) {
+            float w = topk_weights[token * top_k + k];
+            float v = __bfloat162float(expert_out[local_id * expert_stride + token * hidden + dim]);
+            sum += w * v;
+        }
+    }
+    out[idx] = __float2bfloat16(sum);
+}
+
+
+extern "C" {
+
+void launch_moe_topk_softmax_bf16(
+    const void* router_logits,
+    void* topk_ids, void* topk_weights,
+    int num_tokens, int num_experts, int top_k,
+    void* stream
+) {
+    int block = 128;
+    moe_topk_softmax_bf16_kernel<<<num_tokens, block, 0, (cudaStream_t)stream>>>(
+        (const __nv_bfloat16*)router_logits,
+        (int*)topk_ids, (float*)topk_weights,
+        num_experts, top_k
+    );
+    CUDA_CHECK_LAST_ERROR();
+}
+
+void launch_moe_replicate_bf16(
+    const void* x, void* x_rep,
+    int num_tokens, int hidden, int local_experts,
+    void* stream
+) {
+    long long total = (long long)local_experts * num_tokens * hidden;
+    int block = 256;
+    int grid = (int)((total + block - 1) / block);
+    moe_replicate_bf16_kernel<<<grid, block, 0, (cudaStream_t)stream>>>(
+        (const __nv_bfloat16*)x, (__nv_bfloat16*)x_rep,
+        num_tokens, hidden, local_experts
+    );
+    CUDA_CHECK_LAST_ERROR();
+}
+
+void launch_moe_bias_add_3d_bf16(
+    void* x, const void* bias,
+    int batch, int num_tokens, int dim,
+    void* stream
+) {
+    long long total = (long long)batch * num_tokens * dim;
+    int block = 256;
+    int grid = (int)((total + block - 1) / block);
+    moe_bias_add_3d_bf16_kernel<<<grid, block, 0, (cudaStream_t)stream>>>(
+        (__nv_bfloat16*)x, (const __nv_bfloat16*)bias,
+        batch, num_tokens, dim
+    );
+    CUDA_CHECK_LAST_ERROR();
+}
+
+void launch_moe_weighted_sum_bf16(
+    const void* expert_out,
+    const void* topk_ids, const void* topk_weights,
+    void* out,
+    int num_tokens, int hidden, int top_k,
+    int expert_start, int local_experts,
+    void* stream
+) {
+    long long total = (long long)num_tokens * hidden;
+    int block = 256;
+    int grid = (int)((total + block - 1) / block);
+    moe_weighted_sum_bf16_kernel<<<grid, block, 0, (cudaStream_t)stream>>>(
+        (const __nv_bfloat16*)expert_out,
+        (const int*)topk_ids, (const float*)topk_weights,
+        (__nv_bfloat16*)out,
+        num_tokens, hidden, top_k,
+        expert_start, local_experts
+    );
+    CUDA_CHECK_LAST_ERROR();
+}
+
+}

csrc/moe/moe_sparse.cu

@@ -0,0 +1,254 @@
+#include <cuda_bf16.h>
+#include <cuda_fp8.h>
+#include <cstdint>
+#include "../common.cuh"
+
+// ============================================================
+// Sparse MoE decode GEMVs — compute ONLY the routed experts.
+//
+// The dense path replicates x across all local experts and runs a
+// batched GEMM, reading every expert's weights per token. Decode is
+// memory-bound, so reading only the top-k routed experts' weights
+// (~2 of 16 local on average at TP=2) is a ~8x byte reduction.
+//
+// Each block handles one (token, slot) pair's tile of output columns.
+// It reads topk_ids[token, slot] from device memory (no host sync),
+// and exits early if the expert is not owned by this rank. The early
+// return is BLOCK-UNIFORM (every thread sees the same topk_ids value
+// and returns before the shared-memory staging + __syncthreads), so
+// it is safe — unlike the divergent-return bug fixed in gemv.cu.
+//
+// Outputs for non-local slots are NEVER written (uninitialized memory,
+// possibly NaN bit patterns). Downstream consumers must SKIP non-local
+// slots rather than multiply by zero (NaN * 0 = NaN).
+//
+// Per-expert weight scale and bias are fused into the epilogue:
+//   y[t, slot, n] = acc * w_scale[lid] + bias[lid, n]
+// which matches the dense path's GEMM -> moe_bias_add_3d sequence.
+//
+// Activation addressing (x_per_slot):
+//   gate_up: all slots of a token share x[token, :]        (x_per_slot=0)
+//   down:    each slot has its own activation row
+//            x[token * top_k + slot, :]                    (x_per_slot=1)
+// ============================================================
+
+#define SPARSE_TILE_N 8  // output columns per block (= warps per block)
+
+// Weights FP8 E4M3 [local_experts, N, K], activations BF16 (W8A16).
+// Decode is memory-bound (~2 FLOP/byte), so dequant-in-registers GEMV
+// loses nothing to tensor cores and skips activation quantization.
+__global__ void moe_sparse_gemv_fp8_bf16_kernel(
+    const __nv_bfloat16* __restrict__ x,     // [T, K] or [T*top_k, K]
+    const __nv_fp8_e4m3* __restrict__ w,     // [local_experts, N, K]
+    const float* __restrict__ w_scales,      // [local_experts]
+    const __nv_bfloat16* __restrict__ bias,  // [local_experts, N]
+    const int* __restrict__ topk_ids,        // [T, top_k] global expert ids
+    __nv_bfloat16* __restrict__ y,           // [T, top_k, N]
+    int N, int K, int top_k,
+    int expert_start, int local_experts,
+    int x_per_slot
+) {
+    int token = blockIdx.z;
+    int slot = blockIdx.y;
+    int eid = topk_ids[token * top_k + slot];
+    int lid = eid - expert_start;
+    if (lid < 0 || lid >= local_experts) return;  // block-uniform: safe
+
+    extern __shared__ float xs[];  // [K] activation row as float
+    const __nv_bfloat16* xrow =
+        x + (long long)(x_per_slot ? token * top_k + slot : token) * K;
+    for (int i = threadIdx.x; i < K; i += blockDim.x) {
+        xs[i] = __bfloat162float(xrow[i]);
+    }
+    __syncthreads();
+
+    int n = blockIdx.x * SPARSE_TILE_N + (threadIdx.x >> 5);
+    if (n >= N) return;  // after __syncthreads: safe
+    int lane = threadIdx.x & 31;
+
+    // One warp per output column; uint4 = 16 FP8 weights per lane, the
+    // warp covers 512 contiguous bytes per iteration (coalesced).
+    const uint8_t* wrow = (const uint8_t*)w + ((long long)lid * N + n) * K;
+    float acc = 0.0f;
+    for (int i = lane; i < (K >> 4); i += 32) {
+        uint4 packed = *(const uint4*)(wrow + (long long)i * 16);
+        const __nv_fp8_e4m3* pw = (const __nv_fp8_e4m3*)&packed;
+        const float* xk = xs + i * 16;
+        #pragma unroll
+        for (int j = 0; j < 16; j++) {
+            acc += xk[j] * float(pw[j]);
+        }
+    }
+
+    #pragma unroll
+    for (int o = 16; o > 0; o >>= 1) {
+        acc += __shfl_down_sync(0xffffffffu, acc, o);
+    }
+    if (lane == 0) {
+        float v = acc * w_scales[lid]
+                + __bfloat162float(bias[(long long)lid * N + n]);
+        y[((long long)token * top_k + slot) * N + n] = __float2bfloat16(v);
+    }
+}
+
+// MXFP4 W4A16 variant: packed E2M1 nibbles + per-32 UE8M0 block scale,
+// same structure as batched_gemv_mxfp4_bf16_kernel but expert-indexed
+// via topk_ids and with fused per-expert bias.
+#define MXFP4_BLOCK 32
+
+__device__ __constant__ float kSparseFp4Levels[8] =
+    {0.f, 0.5f, 1.f, 1.5f, 2.f, 3.f, 4.f, 6.f};
+
+__device__ __forceinline__ float sparse_fp4_to_float(uint8_t code) {
+    float mag = kSparseFp4Levels[code & 0x7];
+    return (code & 0x8) ? -mag : mag;
+}
+
+__global__ void moe_sparse_gemv_mxfp4_bf16_kernel(
+    const __nv_bfloat16* __restrict__ x,     // [T, K] or [T*top_k, K]
+    const uint8_t* __restrict__ w_packed,    // [local_experts, N, K/2]
+    const uint8_t* __restrict__ w_scales,    // [local_experts, N, K/32]
+    const __nv_bfloat16* __restrict__ bias,  // [local_experts, N]
+    const int* __restrict__ topk_ids,        // [T, top_k]
+    __nv_bfloat16* __restrict__ y,           // [T, top_k, N]
+    int N, int K, int top_k,
+    int expert_start, int local_experts,
+    int x_per_slot
+) {
+    int token = blockIdx.z;
+    int slot = blockIdx.y;
+    int eid = topk_ids[token * top_k + slot];
+    int lid = eid - expert_start;
+    if (lid < 0 || lid >= local_experts) return;  // block-uniform: safe
+
+    extern __shared__ float xs[];
+    const __nv_bfloat16* xrow =
+        x + (long long)(x_per_slot ? token * top_k + slot : token) * K;
+    for (int i = threadIdx.x; i < K; i += blockDim.x) {
+        xs[i] = __bfloat162float(xrow[i]);
+    }
+    __syncthreads();
+
+    int n = blockIdx.x * SPARSE_TILE_N + (threadIdx.x >> 5);
+    if (n >= N) return;
+    int lane = threadIdx.x & 31;
+    int nblk = K / MXFP4_BLOCK;
+
+    const uint8_t* wp = w_packed + ((long long)lid * N + n) * (K >> 1);
+    const uint8_t* ws = w_scales + ((long long)lid * N + n) * nblk;
+
+    float acc = 0.0f;
+    for (int blk = lane; blk < nblk; blk += 32) {
+        float scale = exp2f((float)((int)ws[blk] - 127));
+        uint4 packed = *(const uint4*)(wp + (long long)blk * 16);  // 32 nibbles
+        const uint8_t* pb = (const uint8_t*)&packed;
+        const float* xk = xs + blk * MXFP4_BLOCK;
+        #pragma unroll
+        for (int i = 0; i < 16; i++) {
+            uint8_t b = pb[i];
+            acc += xk[2 * i]     * (sparse_fp4_to_float(b & 0xF) * scale);
+            acc += xk[2 * i + 1] * (sparse_fp4_to_float(b >> 4)  * scale);
+        }
+    }
+
+    #pragma unroll
+    for (int o = 16; o > 0; o >>= 1) {
+        acc += __shfl_down_sync(0xffffffffu, acc, o);
+    }
+    if (lane == 0) {
+        float v = acc + __bfloat162float(bias[(long long)lid * N + n]);
+        y[((long long)token * top_k + slot) * N + n] = __float2bfloat16(v);
+    }
+}
+
+// Weighted sum over the slot axis: out[t, d] = sum over local slots of
+// topk_weights[t, k] * down[t, k, d]. Non-local slots hold uninitialized
+// memory and are SKIPPED (not multiplied by zero).
+__global__ void moe_weighted_sum_sparse_bf16_kernel(
+    const __nv_bfloat16* __restrict__ down,  // [T, top_k, hidden]
+    const int* __restrict__ topk_ids,        // [T, top_k]
+    const float* __restrict__ topk_weights,  // [T, top_k]
+    __nv_bfloat16* __restrict__ out,         // [T, hidden]
+    int num_tokens, int hidden, int top_k,
+    int expert_start, int local_experts
+) {
+    int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    int total = num_tokens * hidden;
+    if (idx >= total) return;
+
+    int token = idx / hidden;
+    int dim = idx % hidden;
+
+    float sum = 0.0f;
+    for (int k = 0; k < top_k; k++) {
+        int lid = topk_ids[token * top_k + k] - expert_start;
+        if (lid >= 0 && lid < local_experts) {
+            float w = topk_weights[token * top_k + k];
+            float v = __bfloat162float(
+                down[((long long)token * top_k + k) * hidden + dim]);
+            sum += w * v;
+        }
+    }
+    out[idx] = __float2bfloat16(sum);
+}
+
+extern "C" {
+
+void launch_moe_sparse_gemv_fp8_bf16(
+    const void* x, const void* w, const void* w_scales, const void* bias,
+    const void* topk_ids, void* y,
+    int num_tokens, int N, int K, int top_k,
+    int expert_start, int local_experts, int x_per_slot,
+    void* stream
+) {
+    dim3 grid((N + SPARSE_TILE_N - 1) / SPARSE_TILE_N, top_k, num_tokens);
+    int block = SPARSE_TILE_N * 32;
+    size_t smem = (size_t)K * sizeof(float);
+    moe_sparse_gemv_fp8_bf16_kernel<<<grid, block, smem, (cudaStream_t)stream>>>(
+        (const __nv_bfloat16*)x, (const __nv_fp8_e4m3*)w,
+        (const float*)w_scales, (const __nv_bfloat16*)bias,
+        (const int*)topk_ids, (__nv_bfloat16*)y,
+        N, K, top_k, expert_start, local_experts, x_per_slot
+    );
+    CUDA_CHECK_LAST_ERROR();
+}
+
+void launch_moe_sparse_gemv_mxfp4_bf16(
+    const void* x, const void* w_packed, const void* w_scales, const void* bias,
+    const void* topk_ids, void* y,
+    int num_tokens, int N, int K, int top_k,
+    int expert_start, int local_experts, int x_per_slot,
+    void* stream
+) {
+    dim3 grid((N + SPARSE_TILE_N - 1) / SPARSE_TILE_N, top_k, num_tokens);
+    int block = SPARSE_TILE_N * 32;
+    size_t smem = (size_t)K * sizeof(float);
+    moe_sparse_gemv_mxfp4_bf16_kernel<<<grid, block, smem, (cudaStream_t)stream>>>(
+        (const __nv_bfloat16*)x, (const uint8_t*)w_packed,
+        (const uint8_t*)w_scales, (const __nv_bfloat16*)bias,
+        (const int*)topk_ids, (__nv_bfloat16*)y,
+        N, K, top_k, expert_start, local_experts, x_per_slot
+    );
+    CUDA_CHECK_LAST_ERROR();
+}
+
+void launch_moe_weighted_sum_sparse_bf16(
+    const void* down, const void* topk_ids, const void* topk_weights,
+    void* out,
+    int num_tokens, int hidden, int top_k,
+    int expert_start, int local_experts,
+    void* stream
+) {
+    int total = num_tokens * hidden;
+    int block = 256;
+    int grid = (total + block - 1) / block;
+    moe_weighted_sum_sparse_bf16_kernel<<<grid, block, 0, (cudaStream_t)stream>>>(
+        (const __nv_bfloat16*)down,
+        (const int*)topk_ids, (const float*)topk_weights,
+        (__nv_bfloat16*)out,
+        num_tokens, hidden, top_k, expert_start, local_experts
+    );
+    CUDA_CHECK_LAST_ERROR();
+}
+
+}

csrc/normalization/layernorm.cu

@@ -14,27 +14,34 @@ __global__ void layernorm_f32(
     const float* x_row = x + row * hidden_size;
     float* out_row = out + row * hidden_size;
 
-    // Welford online: compute mean and variance in one pass
+    // Pass 1: compute mean
     float local_sum = 0.0f;
-    float local_sum_sq = 0.0f;
     for (int i = threadIdx.x; i < hidden_size; i += blockDim.x) {
-        float v = x_row[i];
-        local_sum += v;
-        local_sum_sq += v * v;
+        local_sum += x_row[i];
     }
     local_sum = block_reduce_sum(local_sum);
-    local_sum_sq = block_reduce_sum(local_sum_sq);
 
     __shared__ float s_mean, s_inv_std;
     if (threadIdx.x == 0) {
-        float mean = local_sum / hidden_size;
-        float var = local_sum_sq / hidden_size - mean * mean;
-        s_mean = mean;
-        s_inv_std = rsqrtf(var + eps);
+        s_mean = local_sum / hidden_size;
     }
     __syncthreads();
 
     float mean = s_mean;
+
+    // Pass 2: compute variance = sum((x - mean)^2) / N
+    float local_var = 0.0f;
+    for (int i = threadIdx.x; i < hidden_size; i += blockDim.x) {
+        float d = x_row[i] - mean;
+        local_var += d * d;
+    }
+    local_var = block_reduce_sum(local_var);
+
+    if (threadIdx.x == 0) {
+        s_inv_std = rsqrtf(local_var / hidden_size + eps);
+    }
+    __syncthreads();
+
     float inv_std = s_inv_std;
     for (int i = threadIdx.x; i < hidden_size; i += blockDim.x) {
         out_row[i] = gamma[i] * (x_row[i] - mean) * inv_std + beta[i];
@@ -52,26 +59,34 @@ __global__ void layernorm_bf16(
     const __nv_bfloat16* x_row = x + row * hidden_size;
     __nv_bfloat16* out_row = out + row * hidden_size;
 
+    // Pass 1: compute mean
     float local_sum = 0.0f;
-    float local_sum_sq = 0.0f;
     for (int i = threadIdx.x; i < hidden_size; i += blockDim.x) {
-        float v = __bfloat162float(x_row[i]);
-        local_sum += v;
-        local_sum_sq += v * v;
+        local_sum += __bfloat162float(x_row[i]);
     }
     local_sum = block_reduce_sum(local_sum);
-    local_sum_sq = block_reduce_sum(local_sum_sq);
 
     __shared__ float s_mean, s_inv_std;
     if (threadIdx.x == 0) {
-        float mean = local_sum / hidden_size;
-        float var = local_sum_sq / hidden_size - mean * mean;
-        s_mean = mean;
-        s_inv_std = rsqrtf(var + eps);
+        s_mean = local_sum / hidden_size;
     }
     __syncthreads();
 
     float mean = s_mean;
+
+    // Pass 2: compute variance = sum((x - mean)^2) / N
+    float local_var = 0.0f;
+    for (int i = threadIdx.x; i < hidden_size; i += blockDim.x) {
+        float d = __bfloat162float(x_row[i]) - mean;
+        local_var += d * d;
+    }
+    local_var = block_reduce_sum(local_var);
+
+    if (threadIdx.x == 0) {
+        s_inv_std = rsqrtf(local_var / hidden_size + eps);
+    }
+    __syncthreads();
+
     float inv_std = s_inv_std;
     for (int i = threadIdx.x; i < hidden_size; i += blockDim.x) {
         float v = __bfloat162float(x_row[i]);
@@ -86,17 +101,21 @@ extern "C" {
 void launch_layernorm_f32(const void* x, const void* gamma, const void* beta,
                           void* out, int rows, int hidden_size, float eps, void* stream) {
     int block = (hidden_size < 1024) ? hidden_size : 1024;
+    if (block < 32) block = 32;
     layernorm_f32<<<rows, block, 0, (cudaStream_t)stream>>>(
         (const float*)x, (const float*)gamma, (const float*)beta,
         (float*)out, hidden_size, eps);
+    CUDA_CHECK_LAST_ERROR();
 }
 
 void launch_layernorm_bf16(const void* x, const void* gamma, const void* beta,
                            void* out, int rows, int hidden_size, float eps, void* stream) {
     int block = (hidden_size < 1024) ? hidden_size : 1024;
+    if (block < 32) block = 32;
     layernorm_bf16<<<rows, block, 0, (cudaStream_t)stream>>>(
         (const __nv_bfloat16*)x, (const __nv_bfloat16*)gamma, (const __nv_bfloat16*)beta,
         (__nv_bfloat16*)out, hidden_size, eps);
+    CUDA_CHECK_LAST_ERROR();
 }
 
 }

csrc/normalization/rmsnorm.cu

@@ -63,21 +63,78 @@ __global__ void rmsnorm_bf16(
     }
 }
 
+// Fused Add + RMSNorm: sum_out = x + residual, normed_out = rmsnorm(sum_out, gamma, eps)
+// Each block handles one row of [hidden_size].
+__global__ void add_rmsnorm_bf16(
+    const __nv_bfloat16* __restrict__ x,
+    const __nv_bfloat16* __restrict__ residual,
+    const __nv_bfloat16* __restrict__ gamma,
+    __nv_bfloat16* __restrict__ normed_out,
+    __nv_bfloat16* __restrict__ sum_out,
+    int hidden_size, float eps
+) {
+    int row = blockIdx.x;
+    const __nv_bfloat16* x_row = x + row * hidden_size;
+    const __nv_bfloat16* res_row = residual + row * hidden_size;
+    __nv_bfloat16* sum_row = sum_out + row * hidden_size;
+    __nv_bfloat16* norm_row = normed_out + row * hidden_size;
+
+    // Pass 1: compute sum = x + residual, and accumulate sum_sq
+    float sum_sq = 0.0f;
+    for (int i = threadIdx.x; i < hidden_size; i += blockDim.x) {
+        float s = __bfloat162float(x_row[i]) + __bfloat162float(res_row[i]);
+        sum_row[i] = __float2bfloat16(s);
+        sum_sq += s * s;
+    }
+    sum_sq = block_reduce_sum(sum_sq);
+
+    __shared__ float s_rms_inv;
+    if (threadIdx.x == 0) {
+        s_rms_inv = rsqrtf(sum_sq / hidden_size + eps);
+    }
+    __syncthreads();
+
+    // Pass 2: normed_out = sum * rms_inv * gamma
+    float rms_inv = s_rms_inv;
+    for (int i = threadIdx.x; i < hidden_size; i += blockDim.x) {
+        float s = __bfloat162float(sum_row[i]);
+        float g = __bfloat162float(gamma[i]);
+        norm_row[i] = __float2bfloat16(s * rms_inv * g);
+    }
+}
+
 extern "C" {
 
 void launch_rmsnorm_f32(const void* x, const void* gamma, void* out,
                         int rows, int hidden_size, float eps, void* stream) {
     int block = (hidden_size < 1024) ? hidden_size : 1024;
+    if (block < 32) block = 32;
     rmsnorm_f32<<<rows, block, 0, (cudaStream_t)stream>>>(
         (const float*)x, (const float*)gamma, (float*)out, hidden_size, eps);
+    CUDA_CHECK_LAST_ERROR();
 }
 
 void launch_rmsnorm_bf16(const void* x, const void* gamma, void* out,
                          int rows, int hidden_size, float eps, void* stream) {
     int block = (hidden_size < 1024) ? hidden_size : 1024;
+    if (block < 32) block = 32;
     rmsnorm_bf16<<<rows, block, 0, (cudaStream_t)stream>>>(
         (const __nv_bfloat16*)x, (const __nv_bfloat16*)gamma,
         (__nv_bfloat16*)out, hidden_size, eps);
+    CUDA_CHECK_LAST_ERROR();
+}
+
+void launch_add_rmsnorm_bf16(const void* x, const void* residual, const void* gamma,
+                             void* normed_out, void* sum_out,
+                             int rows, int hidden_size, float eps, void* stream) {
+    int block = (hidden_size < 1024) ? hidden_size : 1024;
+    if (block < 32) block = 32;
+    add_rmsnorm_bf16<<<rows, block, 0, (cudaStream_t)stream>>>(
+        (const __nv_bfloat16*)x, (const __nv_bfloat16*)residual,
+        (const __nv_bfloat16*)gamma,
+        (__nv_bfloat16*)normed_out, (__nv_bfloat16*)sum_out,
+        hidden_size, eps);
+    CUDA_CHECK_LAST_ERROR();
 }
 
 }

csrc/quantization/dequant_fp8.cu

@@ -0,0 +1,53 @@
+#include <cuda_bf16.h>
+#include <cuda_fp8.h>
+#include "../common.cuh"
+
+// Dequantize FP8 E4M3 → BF16 with per-expert (per-batch-slice) FP32 scale.
+//
+// Input:  src    [num_experts, rows, cols]  FP8 E4M3 (1 byte each)
+//         scales [num_experts]              FP32
+// Output: dst    [num_experts, rows, cols]  BF16
+//
+// Each element: dst[e, r, c] = bf16( float(src[e, r, c]) * scales[e] )
+
+__global__ void dequant_fp8e4m3_to_bf16_kernel(
+    const __nv_fp8_e4m3* __restrict__ src,
+    const float* __restrict__ scales,
+    __nv_bfloat16* __restrict__ dst,
+    int num_experts, int rows, int cols
+) {
+    // 64-bit index: num_experts * rows * cols overflows int32 for 32 experts
+    // at ~8k*8k weight matrices, same class as the MoE fix in cfbd64d.
+    long long idx = (long long)blockIdx.x * blockDim.x + threadIdx.x;
+    long long total = (long long)num_experts * rows * cols;
+    if (idx >= total) return;
+
+    long long expert_stride = (long long)rows * cols;
+    int expert = (int)(idx / expert_stride);
+    float scale = scales[expert];
+    float val = float(src[idx]) * scale;
+    dst[idx] = __float2bfloat16(val);
+}
+
+extern "C" {
+
+void launch_dequant_fp8e4m3_to_bf16(
+    const void* src,
+    const void* scales,
+    void* dst,
+    int num_experts, int rows, int cols,
+    void* stream
+) {
+    long long total = (long long)num_experts * rows * cols;
+    int block = 256;
+    int grid = (int)((total + block - 1) / block);
+    dequant_fp8e4m3_to_bf16_kernel<<<grid, block, 0, (cudaStream_t)stream>>>(
+        (const __nv_fp8_e4m3*)src,
+        (const float*)scales,
+        (__nv_bfloat16*)dst,
+        num_experts, rows, cols
+    );
+    CUDA_CHECK_LAST_ERROR();
+}
+
+}

csrc/quantization/mxfp4_gemm.cu

@@ -0,0 +1,135 @@
+#include <cuda_bf16.h>
+#include <cstdint>
+#include "../common.cuh"
+
+// MXFP4 W4A16 for MoE experts. Weights stored [E, N, K] with K (reduction)
+// contiguous, blocked by 32: packed 4-bit E2M1 (two nibbles/byte, lo = even k)
+// + one UE8M0 scale byte per 32 elements. The decode win is reading 4-bit
+// weights from HBM (half of FP8) and dequantizing on-chip to BF16.
+
+#define MXFP4_BLOCK 32
+
+// E2M1 magnitude by 3-bit code; bit 3 is the sign.
+__device__ __constant__ float kFp4Levels[8] = {0.f, 0.5f, 1.f, 1.5f, 2.f, 3.f, 4.f, 6.f};
+
+__device__ __forceinline__ float fp4_to_float(uint8_t code) {
+    float mag = kFp4Levels[code & 0x7];
+    return (code & 0x8) ? -mag : mag;
+}
+
+// Decode (M=1) fused GEMV, batched over experts.
+//   y[e, n] = sum_k x[e, k] * dequant(W[e, n, k])
+// Grid: (N/TILE_N, E). Each block loads the activation x[e, :] into shared
+// memory ONCE and computes TILE_N output columns from it (one warp per column),
+// so the activation is read from HBM once per TILE_N outputs instead of once
+// per output. Weights are unique per output and read coalesced as uint4; the
+// UE8M0 block scale is hoisted to once per 32-element block.
+#define MXFP4_TILE_N 8   // output columns per block (= warps per block)
+
+__global__ void batched_gemv_mxfp4_bf16_kernel(
+    const __nv_bfloat16* __restrict__ x,   // [E, K]
+    const uint8_t* __restrict__ w_packed,  // [E, N, K/2]
+    const uint8_t* __restrict__ w_scales,  // [E, N, K/32]
+    __nv_bfloat16* __restrict__ y,         // [E, N]
+    int E, int N, int K
+) {
+    extern __shared__ float xs[];          // [K] activation for this expert
+    int e = blockIdx.y;
+    int n_base = blockIdx.x * MXFP4_TILE_N;
+    int warp = threadIdx.x >> 5;           // 0..TILE_N-1
+    int lane = threadIdx.x & 31;
+    int nthreads = blockDim.x;             // TILE_N * 32
+    int nblk = K / MXFP4_BLOCK;
+
+    // Cooperatively stage x[e, :] into shared memory (converted to float).
+    const __nv_bfloat16* xe = x + (long long)e * K;
+    for (int k = threadIdx.x; k < K; k += nthreads) {
+        xs[k] = __bfloat162float(xe[k]);
+    }
+    __syncthreads();
+
+    int n = n_base + warp;
+    if (n >= N) return;
+
+    const uint8_t* wp = w_packed + ((long long)e * N + n) * (K >> 1);
+    const uint8_t* ws = w_scales + ((long long)e * N + n) * nblk;
+
+    float acc = 0.0f;
+    for (int blk = lane; blk < nblk; blk += 32) {
+        float scale = exp2f((float)((int)ws[blk] - 127));
+        uint4 packed = *(const uint4*)(wp + (long long)blk * 16);  // 16 bytes = 32 nibbles
+        const uint8_t* pb = (const uint8_t*)&packed;
+        const float* xk = xs + blk * MXFP4_BLOCK;
+        #pragma unroll
+        for (int i = 0; i < 16; i++) {
+            uint8_t b = pb[i];
+            acc += xk[2 * i]     * (fp4_to_float(b & 0xF) * scale);
+            acc += xk[2 * i + 1] * (fp4_to_float(b >> 4)  * scale);
+        }
+    }
+
+    // Warp reduction.
+    #pragma unroll
+    for (int o = 16; o > 0; o >>= 1) {
+        acc += __shfl_down_sync(0xffffffffu, acc, o);
+    }
+    if (lane == 0) y[(long long)e * N + n] = __float2bfloat16(acc);
+}
+
+// Prefill fallback: dequant MXFP4 [E, N, K] -> BF16 [E, K, N] (transposed back
+// to the [E, K, N] layout the BF16 batched GEMM expects). Not bandwidth-optimal,
+// but prefill is compute-bound so it is not the decode hot path.
+__global__ void dequant_mxfp4_to_bf16_t_kernel(
+    const uint8_t* __restrict__ w_packed,  // [E, N, K/2]
+    const uint8_t* __restrict__ w_scales,  // [E, N, K/32]
+    __nv_bfloat16* __restrict__ out,       // [E, K, N]
+    int E, int N, int K
+) {
+    long long idx = (long long)blockIdx.x * blockDim.x + threadIdx.x;
+    long long total = (long long)E * N * K;
+    if (idx >= total) return;
+    int k = idx % K;
+    int n = (idx / K) % N;
+    int e = idx / ((long long)N * K);
+
+    int Kh = K >> 1;
+    int Ks = K / MXFP4_BLOCK;
+    uint8_t byte = w_packed[((long long)e * N + n) * Kh + (k >> 1)];
+    uint8_t code = (k & 1) ? (byte >> 4) : (byte & 0xF);
+    float scale = exp2f((float)((int)w_scales[((long long)e * N + n) * Ks + k / MXFP4_BLOCK] - 127));
+    float val = fp4_to_float(code) * scale;
+    // write to out[e, k, n]
+    out[((long long)e * K + k) * N + n] = __float2bfloat16(val);
+}
+
+extern "C" {
+
+void launch_batched_gemv_mxfp4_bf16(
+    const void* x, const void* w_packed, const void* w_scales, void* y,
+    int E, int N, int K, void* stream
+) {
+    dim3 grid((N + MXFP4_TILE_N - 1) / MXFP4_TILE_N, E);
+    int block = MXFP4_TILE_N * 32;  // one warp per output column
+    size_t smem = (size_t)K * sizeof(float);
+    batched_gemv_mxfp4_bf16_kernel<<<grid, block, smem, (cudaStream_t)stream>>>(
+        (const __nv_bfloat16*)x, (const uint8_t*)w_packed, (const uint8_t*)w_scales,
+        (__nv_bfloat16*)y, E, N, K
+    );
+    CUDA_CHECK_LAST_ERROR();
+}
+
+void launch_dequant_mxfp4_to_bf16_t(
+    const void* w_packed, const void* w_scales, void* out,
+    int E, int N, int K, void* stream
+) {
+    long long total = (long long)E * N * K;
+    int block = 256;
+    long long grid = (total + block - 1) / block;
+    dequant_mxfp4_to_bf16_t_kernel<<<(unsigned)grid, block, 0, (cudaStream_t)stream>>>(
+        (const uint8_t*)w_packed, (const uint8_t*)w_scales, (__nv_bfloat16*)out,
+        E, N, K
+    );
+    CUDA_CHECK_LAST_ERROR();
+}
+
+}

csrc/quantization/quantize_fp8.cu

@@ -0,0 +1,160 @@
+#include <cuda_bf16.h>
+#include <cuda_fp8.h>
+#include <float.h>
+#include "../common.cuh"
+
+// Per-row quantize BF16 → FP8 E4M3 with per-row FP32 scale output.
+//
+// Input:  src    [num_rows, cols]  BF16
+// Output: dst    [num_rows, cols]  FP8 E4M3
+//         scales [num_rows]        FP32
+//
+// Algorithm per row:
+//   absmax = max(|src[row, :]|)
+//   scale  = absmax / 448.0       (FP8 E4M3 max representable)
+//   dst[row, i] = fp8(src[row, i] / scale)
+//
+// Grid: one block per row.  Block: 256 threads.
+// Each thread handles ceil(cols / 256) elements.
+
+#define QUANT_BLOCK 256
+#define FP8_E4M3_MAX 448.0f
+
+__global__ void quantize_bf16_to_fp8e4m3_rowwise_kernel(
+    const __nv_bfloat16* __restrict__ src,
+    __nv_fp8_e4m3* __restrict__ dst,
+    float* __restrict__ scales,
+    int num_rows, int cols
+) {
+    int row = blockIdx.x;
+    if (row >= num_rows) return;
+    int tid = threadIdx.x;
+
+    const __nv_bfloat16* row_src = src + (long long)row * cols;
+    __nv_fp8_e4m3* row_dst = dst + (long long)row * cols;
+
+    // Step 1: Compute per-row absmax via shared-memory reduction.
+    __shared__ float smem_max[QUANT_BLOCK];
+    float local_max = 0.0f;
+    for (int i = tid; i < cols; i += QUANT_BLOCK) {
+        float v = fabsf(__bfloat162float(row_src[i]));
+        local_max = fmaxf(local_max, v);
+    }
+    smem_max[tid] = local_max;
+    __syncthreads();
+
+    // Block reduction
+    for (int s = QUANT_BLOCK / 2; s > 0; s >>= 1) {
+        if (tid < s) {
+            smem_max[tid] = fmaxf(smem_max[tid], smem_max[tid + s]);
+        }
+        __syncthreads();
+    }
+
+    float absmax = smem_max[0];
+    float scale = absmax / FP8_E4M3_MAX;
+    // Clamp scale to avoid div-by-zero for all-zero rows
+    if (scale < 1e-12f) scale = 1e-12f;
+    float inv_scale = 1.0f / scale;
+
+    // Thread 0 writes the scale
+    if (tid == 0) {
+        scales[row] = scale;
+    }
+
+    // Step 2: Quantize each element
+    for (int i = tid; i < cols; i += QUANT_BLOCK) {
+        float v = __bfloat162float(row_src[i]) * inv_scale;
+        row_dst[i] = __nv_fp8_e4m3(v);
+    }
+}
+
+// Row-wise scale: data[row, :] *= scales[row]  (in-place, BF16)
+__global__ void rowwise_scale_bf16_kernel(
+    __nv_bfloat16* __restrict__ data,
+    const float* __restrict__ scales,
+    int num_rows, int cols
+) {
+    int row = blockIdx.x;
+    if (row >= num_rows) return;
+    int tid = threadIdx.x;
+    float s = scales[row];
+    __nv_bfloat16* row_data = data + (long long)row * cols;
+    for (int i = tid; i < cols; i += blockDim.x) {
+        float v = __bfloat162float(row_data[i]) * s;
+        row_data[i] = __float2bfloat16(v);
+    }
+}
+
+// Combined dequant scale for batched MoE FP8 GEMM output.
+//   data[row, :] *= a_scales[row] * b_scales[row / tokens]
+// where row = expert * tokens + token. a_scales is the per-token activation
+// scale; b_scales is the per-expert scalar weight scale. Lets a single
+// strided-batched FP8 matmul (alpha=1, scales=1) recover the real result in
+// one pass instead of folding the weight scale into a per-expert GEMM call.
+__global__ void rowwise_scale_moe_bf16_kernel(
+    __nv_bfloat16* __restrict__ data,
+    const float* __restrict__ a_scales,
+    const float* __restrict__ b_scales,
+    int num_rows, int cols, int tokens
+) {
+    int row = blockIdx.x;
+    if (row >= num_rows) return;
+    int tid = threadIdx.x;
+    float s = a_scales[row] * b_scales[row / tokens];
+    __nv_bfloat16* row_data = data + (long long)row * cols;
+    for (int i = tid; i < cols; i += blockDim.x) {
+        float v = __bfloat162float(row_data[i]) * s;
+        row_data[i] = __float2bfloat16(v);
+    }
+}
+
+extern "C" {
+
+void launch_rowwise_scale_bf16(
+    void* data, const void* scales,
+    int num_rows, int cols,
+    void* stream
+) {
+    int block = 256;
+    int grid = num_rows;
+    rowwise_scale_bf16_kernel<<<grid, block, 0, (cudaStream_t)stream>>>(
+        (__nv_bfloat16*)data, (const float*)scales,
+        num_rows, cols
+    );
+    CUDA_CHECK_LAST_ERROR();
+}
+
+void launch_rowwise_scale_moe_bf16(
+    void* data, const void* a_scales, const void* b_scales,
+    int num_rows, int cols, int tokens,
+    void* stream
+) {
+    int block = 256;
+    int grid = num_rows;
+    rowwise_scale_moe_bf16_kernel<<<grid, block, 0, (cudaStream_t)stream>>>(
+        (__nv_bfloat16*)data, (const float*)a_scales, (const float*)b_scales,
+        num_rows, cols, tokens
+    );
+    CUDA_CHECK_LAST_ERROR();
+}
+
+void launch_quantize_bf16_to_fp8e4m3_rowwise(
+    const void* src,
+    void* dst,
+    void* scales,
+    int num_rows, int cols,
+    void* stream
+) {
+    int grid = num_rows;
+    int block = QUANT_BLOCK;
+    quantize_bf16_to_fp8e4m3_rowwise_kernel<<<grid, block, 0, (cudaStream_t)stream>>>(
+        (const __nv_bfloat16*)src,
+        (__nv_fp8_e4m3*)dst,
+        (float*)scales,
+        num_rows, cols
+    );
+    CUDA_CHECK_LAST_ERROR();
+}
+
+}

csrc/reduce/argmax.cu

@@ -0,0 +1,92 @@
+#include <cuda_bf16.h>
+#include <float.h>
+#include "../common.cuh"
+
+// Argmax along the last dim of a [rows, cols] tensor.
+// One block per row; output is [rows] int32 indices of the max element.
+//
+// Reduction: each thread scans a strided slice and tracks the running
+// (value, index) pair, then warp-shuffle reduce, then a single-warp
+// reduce over per-warp leaders. Tie-break: smaller index wins so the
+// result is deterministic across launches.
+//
+// For BF16 logits the comparison happens in FP32 to avoid losing
+// precision near the top of the distribution.
+
+__global__ void argmax_bf16_kernel(
+    const __nv_bfloat16* __restrict__ logits,
+    int* __restrict__ out_idx,
+    int cols
+) {
+    int row = blockIdx.x;
+    const __nv_bfloat16* row_ptr = logits + (long long)row * cols;
+    int tid = threadIdx.x;
+    unsigned mask = 0xffffffff;
+
+    // Strided per-thread max.
+    float local_max = -FLT_MAX;
+    int   local_idx = INT_MAX;
+    for (int i = tid; i < cols; i += blockDim.x) {
+        float v = __bfloat162float(row_ptr[i]);
+        // strict `>` keeps the smallest index on ties, since we scan ascending.
+        if (v > local_max) {
+            local_max = v;
+            local_idx = i;
+        }
+    }
+
+    // Warp-level reduce of (val, idx) pairs.
+    #pragma unroll
+    for (int offset = 16; offset > 0; offset >>= 1) {
+        float other_val = __shfl_down_sync(mask, local_max, offset);
+        int   other_idx = __shfl_down_sync(mask, local_idx, offset);
+        bool take = (other_val > local_max) ||
+                    (other_val == local_max && other_idx < local_idx);
+        if (take) {
+            local_max = other_val;
+            local_idx = other_idx;
+        }
+    }
+
+    // Per-warp leaders → shared memory → single warp final reduce.
+    __shared__ float s_val[32];
+    __shared__ int   s_idx[32];
+    int lane     = tid & 31;
+    int warp_id  = tid >> 5;
+    int num_warps = (blockDim.x + 31) >> 5;
+
+    if (lane == 0) {
+        s_val[warp_id] = local_max;
+        s_idx[warp_id] = local_idx;
+    }
+    __syncthreads();
+
+    if (warp_id == 0) {
+        float v = (tid < num_warps) ? s_val[lane] : -FLT_MAX;
+        int   i = (tid < num_warps) ? s_idx[lane] : INT_MAX;
+        #pragma unroll
+        for (int offset = 16; offset > 0; offset >>= 1) {
+            float ov = __shfl_down_sync(mask, v, offset);
+            int   oi = __shfl_down_sync(mask, i, offset);
+            bool take = (ov > v) || (ov == v && oi < i);
+            if (take) { v = ov; i = oi; }
+        }
+        if (lane == 0) {
+            out_idx[row] = i;
+        }
+    }
+}
+
+extern "C" {
+
+void launch_argmax_bf16(const void* logits, void* out_idx,
+                        int rows, int cols, void* stream) {
+    // 1024 threads/block keeps occupancy high and gives 32 warps for the
+    // final reduce (matches the 32-slot shared arrays above).
+    int block = 1024;
+    argmax_bf16_kernel<<<rows, block, 0, (cudaStream_t)stream>>>(
+        (const __nv_bfloat16*)logits, (int*)out_idx, cols);
+    CUDA_CHECK_LAST_ERROR();
+}
+
+}

csrc/reduce/softmax.cu

@@ -90,17 +90,19 @@ __global__ void softmax_bf16(
 extern "C" {
 
 void launch_softmax_f32(const void* x, void* out, int rows, int cols, void* stream) {
-    int block = (cols < 1024) ? cols : 1024;
+    int block = (cols < 512) ? cols : 512;
     if (block < 32) block = 32;
     softmax_f32<<<rows, block, 0, (cudaStream_t)stream>>>(
         (const float*)x, (float*)out, cols);
+    CUDA_CHECK_LAST_ERROR();
 }
 
 void launch_softmax_bf16(const void* x, void* out, int rows, int cols, void* stream) {
-    int block = (cols < 1024) ? cols : 1024;
+    int block = (cols < 512) ? cols : 512;
     if (block < 32) block = 32;
     softmax_bf16<<<rows, block, 0, (cudaStream_t)stream>>>(
         (const __nv_bfloat16*)x, (__nv_bfloat16*)out, cols);
+    CUDA_CHECK_LAST_ERROR();
 }
 
 }

docs/00-roadmap.md

@@ -9,7 +9,7 @@
 | 抽象层级 | Level 0.5 | 自写 CUDA kernel + cuBLAS 可切换，便于 benchmark 对比 |
 | 硬件 | 8×RTX 5090 (Blackwell, CC 12.0, 32GB GDDR7) | 纯 PCIe Gen5 x16 互联，无 NVLink (详见下方硬件拓扑) |
 | 语言 | Rust + CUDA (C/C++) | Rust FFI 调用 CUDA |
-| 起步模型 | GPT-2 124M → Qwen3-7B | 从简单到实用 |
+| 起步模型 | GPT-2 124M → Qwen3-8B | 从简单到实用 |
 | 精度 | BF16/FP16 | 后期扩展 FP8 |
 | Tensor | 自己实现 | 完整学习 tensor 抽象设计 |
 | Tokenizer | 自己实现 BPE | 学习分词机制 |
@@ -101,7 +101,7 @@ Phase 8: GPT-2 完整推理 ◄──────────── 里程碑
     │
 Phase 9: KV Cache + Autoregressive Generation
     │
-Phase 10: Qwen3-7B 支持 ◄─────────── 里程碑 ② 7B 模型推理
+Phase 10: Qwen3-8B 支持 ◄─────────── 里程碑 ② 8B 模型推理
     │
 Phase 11: Paged Attention + KV Cache Manager
     │
@@ -109,7 +109,7 @@ Phase 12: Continuous Batching + Request Scheduler
     │
 Phase 13: HTTP API + SSE Streaming ◄── 里程碑 ③ 端到端 API 可用
     │
-Phase 14: Flash Attention v2
+Phase 14: Flash Attention (FA2 for SM120)
     │
 Phase 15: 性能优化 ◄──────────────── 里程碑 ④ 50% vLLM throughput
     │
@@ -625,8 +625,8 @@ safetensors file (disk)
 
 - [ ] 加载 GPT-2 124M (`openai-community/gpt2`)，打印所有 tensor name, shape, dtype
 - [ ] 抽查几个 tensor 的前 10 个值，与 PyTorch `from_pretrained` 对比
-- [ ] 加载 Qwen3-7B sharded 权重，验证所有 tensor 都成功加载
-- [ ] 性能: 测量 7B 模型权重加载时间 (mmap → GPU 全流程)
+- [ ] 加载 Qwen3-8B sharded 权重，验证所有 tensor 都成功加载
+- [ ] 性能: 测量 8B 模型权重加载时间 (mmap → GPU 全流程)
 - [ ] 错误处理: 缺少 tensor、dtype 不匹配、文件不存在等情况
 
 ---
@@ -869,15 +869,15 @@ weights × V_cache [B, H, S, D] → output [B, H, 1, D]
 
 ---
 
-## Phase 10: Qwen3-7B 支持 — 里程碑 ②
+## Phase 10: Qwen3-8B 支持 — 里程碑 ②
 
 **Crate**: `xserv-model`
 
-**目标**: 扩展模型定义以支持 Qwen3-7B，验证输出正确性。
+**目标**: 扩展模型定义以支持 Qwen3-8B，验证输出正确性。
 
 ### 架构对比
 
-| 特性 | GPT-2 (124M) | Qwen3-7B |
+| 特性 | GPT-2 (124M) | Qwen3-8B |
 |------|-------------|----------|
 | Normalization | LayerNorm (pre-LN) | RMSNorm (pre-LN) |
 | Position Encoding | Learned absolute (wpe) | RoPE (无单独参数) |
@@ -885,8 +885,8 @@ weights × V_cache [B, H, S, D] → output [B, H, 1, D]
 | Activation | GELU | SwiGLU (SiLU gate) |
 | FFN | Linear(H→4H) → GELU → Linear(4H→H) | gate_proj + up_proj → SiLU gate → down_proj |
 | Vocab Size | 50,257 | ~152,000 |
-| Hidden Size | 768 | 3,584 (7B) |
-| Layers | 12 | 28 |
+| Hidden Size | 768 | 4,096 (8B) |
+| Layers | 12 | 36 |
 | Tied Embeddings | Yes | No |
 
 ### 需要新增/修改的组件
@@ -948,16 +948,16 @@ pub struct Qwen3DecoderLayer {
 ### 显存预算 (BF16, 单卡 5090 32GB)
 
 ```
-模型权重:  7B × 2B = ~14 GB
-KV cache:  28 layers × 2(KV) × 8 heads × 4096 tokens × 128 dim × 2B ≈ 4.5 GB
+模型权重:  8B × 2B = ~16 GB
+KV cache:  36 layers × 2(KV) × 8 heads × 4096 tokens × 128 dim × 2B ≈ 5.6 GB
 Activation (单请求): ~1 GB
 ────────────────────────
-总计: ~19.5 GB (单请求)，剩余 ~12 GB 可用于更多并发
+总计: ~22.6 GB (单请求)，剩余 ~10 GB 可用于更多并发
 ```
 
 ### 测试验收
 
-- [ ] 加载 Qwen3-7B 权重到单张 5090，打印模型结构和参数量
+- [ ] 加载 Qwen3-8B 权重到单张 5090，打印模型结构和参数量
 - [ ] Prefill logits 与 HF transformers 对比: 输入 "你好" → top-5 logits 一致
 - [ ] 英文生成: "What is the capital of France?" → 生成合理回答
 - [ ] 中文生成: "请介绍一下量子计算" → 生成通顺中文
@@ -1196,7 +1196,7 @@ GET  /health                 # 健康检查
 **Chat Completion Request**:
 ```json
 {
-  "model": "qwen3-7b",
+  "model": "qwen3-8b",
   "messages": [
     {"role": "system", "content": "You are a helpful assistant."},
     {"role": "user", "content": "What is 1+1?"}
@@ -1211,13 +1211,13 @@ GET  /health                 # 健康检查
 
 **SSE Streaming Response**:
 ```
-data: {"id":"chatcmpl-xxx","object":"chat.completion.chunk","created":1234567890,"model":"qwen3-7b","choices":[{"index":0,"delta":{"role":"assistant","content":""},"finish_reason":null}]}
+data: {"id":"chatcmpl-xxx","object":"chat.completion.chunk","created":1234567890,"model":"qwen3-8b","choices":[{"index":0,"delta":{"role":"assistant","content":""},"finish_reason":null}]}
 
-data: {"id":"chatcmpl-xxx","object":"chat.completion.chunk","created":1234567890,"model":"qwen3-7b","choices":[{"index":0,"delta":{"content":"The"},"finish_reason":null}]}
+data: {"id":"chatcmpl-xxx","object":"chat.completion.chunk","created":1234567890,"model":"qwen3-8b","choices":[{"index":0,"delta":{"content":"The"},"finish_reason":null}]}
 
-data: {"id":"chatcmpl-xxx","object":"chat.completion.chunk","created":1234567890,"model":"qwen3-7b","choices":[{"index":0,"delta":{"content":" answer"},"finish_reason":null}]}
+data: {"id":"chatcmpl-xxx","object":"chat.completion.chunk","created":1234567890,"model":"qwen3-8b","choices":[{"index":0,"delta":{"content":" answer"},"finish_reason":null}]}
 
-data: {"id":"chatcmpl-xxx","object":"chat.completion.chunk","created":1234567890,"model":"qwen3-7b","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}
+data: {"id":"chatcmpl-xxx","object":"chat.completion.chunk","created":1234567890,"model":"qwen3-8b","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}
 
 data: [DONE]
 ```
@@ -1228,7 +1228,7 @@ data: [DONE]
   "id": "chatcmpl-xxx",
   "object": "chat.completion",
   "created": 1234567890,
-  "model": "qwen3-7b",
+  "model": "qwen3-8b",
   "choices": [{
     "index": 0,
     "message": {"role": "assistant", "content": "The answer is 2."},
@@ -1278,7 +1278,7 @@ Client (curl / Python OpenAI SDK)
   ```bash
   curl http://localhost:8080/v1/chat/completions \
     -H "Content-Type: application/json" \
-    -d '{"model":"qwen3-7b","messages":[{"role":"user","content":"Hello"}],"stream":true}'
+    -d '{"model":"qwen3-8b","messages":[{"role":"user","content":"Hello"}],"stream":true}'
   ```
   看到 SSE 逐 token 输出
 
@@ -1287,7 +1287,7 @@ Client (curl / Python OpenAI SDK)
   from openai import OpenAI
   client = OpenAI(base_url="http://localhost:8080/v1", api_key="unused")
   for chunk in client.chat.completions.create(
-      model="qwen3-7b",
+      model="qwen3-8b",
       messages=[{"role": "user", "content": "What is 1+1?"}],
       stream=True
   ):
@@ -1302,12 +1302,26 @@ Client (curl / Python OpenAI SDK)
 
 ---
 
-## Phase 14: Flash Attention v2
+## Phase 14: Flash Attention (FA2 for SM120)
 
 **Crate**: `xserv-kernels`
 **CUDA 源码**: `csrc/attention/flash_attention.cu`
 
-**目标**: 实现 Flash Attention v2 的 CUDA kernel，大幅降低 attention 的显存占用并提升速度。
+**目标**: 实现 Flash Attention 的 CUDA kernel，大幅降低 attention 的显存占用并提升速度。
+
+### 硬件适配说明
+
+Flash Attention 已发展到第 4 代 (FA4, arxiv 2603.05451)，但各版本有明确的硬件依赖:
+
+| 版本 | 目标架构 | 关键硬件特性 | RTX 5090 兼容 |
+|------|---------|------------|--------------|
+| FA2 | 通用 CUDA (SM75+) | 标准 shared memory + HMMA | **是** ✅ |
+| FA3 | Hopper SM90 (H100) | TMA + WGMMA + warp specialization | 否 |
+| FA4 | Blackwell SM100 (B200/B300) | TMEM + async MMA + 2-CTA mode | 否 |
+
+**RTX 5090 (SM120, CC 12.0) 使用的是消费级 Blackwell 架构 (GB202)，与数据中心 Blackwell (B200, SM100) 是不同的硅片设计。SM120 物理上没有 TMEM (Tensor Memory) 子系统，因此 FA4 的 kernel 无法在 5090 上运行。这不是软件限制，是硬件级差异。**
+
+因此本项目实现 **FA2 算法**，使用标准 CUDA (shared memory + HMMA)。FA2 的核心优化——online softmax tiling、O(1) 显存占用——在任何架构上都有效。
 
 ### 核心思想
 
@@ -1323,16 +1337,18 @@ Flash Attention 的解法:
 - 将 Q, K, V 分成 tiles，在 SRAM (shared memory) 中计算
 - 使用 **online softmax trick**: 边算边更新 running max 和 running sum
 
-### 算法 (Forward Pass)
+### 算法 (Forward Pass, FA2)
+
+FA2 相比 FA1 的改进: 外层循环遍历 Q tiles (而非 K/V)，减少 HBM 读写次数。
 
 ```
 Br, Bc = tile sizes for Q and K/V respectively
 
-for each Q tile (q_start..q_start+Br):
+for each Q tile (q_start..q_start+Br):                    ← 外层: Q tiles
     load Q_tile [Br, D] to shared memory
-    initialize: O_tile = 0, l = 0, m = -inf    // running sum and max
+    initialize: O_tile = 0, l = 0, m = -inf               // running sum and max
 
-    for each K,V tile (kv_start..kv_start+Bc):
+    for each K,V tile (kv_start..kv_start+Bc):             ← 内层: K/V tiles
         load K_tile [Bc, D], V_tile [Bc, D] to shared memory
 
         // Compute attention scores for this tile pair
@@ -1345,6 +1361,8 @@ for each Q tile (q_start..q_start+Br):
         m_new = max(m, rowmax(S_tile))          // new running max
         P_tile = exp(S_tile - m_new)            // safe exp
         l_new = exp(m - m_new) * l + rowsum(P_tile)  // update running sum
+
+        // Rescale and accumulate output
         O_tile = diag(exp(m - m_new)) * O_tile + P_tile @ V_tile
         m = m_new
         l = l_new
@@ -1356,9 +1374,12 @@ for each Q tile (q_start..q_start+Br):
 ### 实现要点
 
 1. **Tile 大小选择**:
-   - 受限于 shared memory (5090 Blackwell CC 12.0: 需要实测确认 per-SM shared memory 上限)
-   - 需要同时存 Q_tile, K_tile, V_tile, S_tile
-   - 典型值: Br=Bc=128 for D=128, BF16
+   - 5090 SM120: shared memory per SM = 100 KB (需实测确认)
+   - 需同时存 Q_tile, K_tile, V_tile, S_tile
+   - BF16: Q_tile [Br, D] = Br × 128 × 2B; K_tile [Bc, D] = Bc × 128 × 2B
+   - S_tile [Br, Bc] 保持 FP32 = Br × Bc × 4B
+   - 推荐起步: Br=Bc=64, head_dim=128 → 共需 ~100KB shared memory
+   - 优化版: Br=Bc=128 需要更多 shared memory, 可能需要拆分
 
 2. **Causal mask 优化**:
    - 如果 K/V tile 完全在 Q tile 的"未来"（kv_start > q_end）→ 跳过整个 tile
@@ -1369,10 +1390,14 @@ for each Q tile (q_start..q_start+Br):
    - Q, K, V 的加载用 BF16（节省 bandwidth）
    - 最终 O 转回 BF16 写出
 
-4. **与 Paged Attention 的结合**:
-   - Flash Attention 的 K/V tile 遍历逻辑需要适配间接寻址
-   - 每个 tile 查 block_table 得到物理地址
-   - 这是 "Flash-Decoding" / "FlashInfer" 的核心
+4. **GQA 支持**:
+   - K/V heads 数量 < Q heads 时，kernel 中做 `kv_head = q_head / num_groups` 索引
+   - 不需要 repeat_kv 操作，直接在 kernel 内部解决
+
+5. **Decode attention 特化**:
+   - Decode 时 Q 只有 1 行 (Br=1)，退化为 vector-matrix attention
+   - 可以写一个专门的 decode attention kernel (类似 FlashDecoding)
+   - 沿 KV sequence 维度做 parallel reduction
 
 ### 测试验收
 
@@ -1386,8 +1411,9 @@ for each Q tile (q_start..q_start+Br):
 | 8192 | OOM? | MB | OOM? | ms |
 | 32768 | OOM | MB | OOM | ms |
 
-- [ ] 集成到 Qwen3-7B，端到端 decode latency 对比
+- [ ] 集成到 Qwen3-8B，端到端 decode latency 对比
 - [ ] Profile: `ncu` 分析 compute utilization, memory throughput
+- [ ] GQA 支持: 无 repeat_kv 开销
 
 ---
 
@@ -1441,7 +1467,7 @@ ncu --target-processes all --set full ./target/release/xserv-server
 
 ### 测试验收
 
-- [ ] 安装 vLLM，同一台机器跑 Qwen3-7B
+- [ ] 安装 vLLM，同一台机器跑 Qwen3-8B
 - [ ] Benchmark 对比:
 
 | Metric | vLLM | xserv | Ratio |
@@ -1488,7 +1514,7 @@ ncu --target-processes all --set full ./target/release/xserv-server
 
 - **无损**: rejection sampling 保证输出分布与纯 target model 一致
 - **加速条件**: draft model 足够快且与 target 分布接近
-- **Draft model 选择**: Qwen3-0.5B / Qwen3-1.5B 作为 Qwen3-7B 的 draft
+- **Draft model 选择**: Qwen3-0.5B / Qwen3-1.5B 作为 Qwen3-8B 的 draft
 
 ### KV Cache 处理
 
@@ -1578,7 +1604,7 @@ Row Parallel: down_proj 按行切分
 
 ### 测试验收
 
-- [ ] TP=2: Qwen3-7B 输出与单卡 (TP=1) 完全一致
+- [ ] TP=2: Qwen3-8B 输出与单卡 (TP=1) 完全一致
 - [ ] TP=4: 每卡权重显存占用约 1/4
 - [ ] Scaling benchmark (同组 GPU 0-3):
 
@@ -1646,7 +1672,7 @@ tensor_fp8 = cast_to_fp8(tensor / scale)
 | FP8 E4M3 | X.XX | +0.XX |
 | INT8 weight-only | X.XX | +0.XX |
 
-- [ ] 显存: FP8 权重占用约 BF16 的一半 (~7 GB for 7B model)
+- [ ] 显存: FP8 权重占用约 BF16 的一半 (~8 GB for 8B model)
 - [ ] 性能: FP8 GEMM throughput vs BF16 GEMM
 
 ---
@@ -1722,16 +1748,39 @@ Text  → Tokenizer   → Text Tokens   ────────────→
 
 ---
 
+## 实际进展记录(与原计划的分叉,2026-06 更新)
+
+Phase 0–17 按计划完成。Phase 18 起实际路线偏离了上面的原计划
+(speculative decoding 与多模态推迟),实际走向是 MoE + 量化 + 稀疏化:
+
+| 实际 Phase | 内容 | 文档 |
+|---|---|---|
+| 18 | Pipeline Parallelism(PP=2/4) | `18-pipeline-parallelism.md`、`benchmarks/pp-sweep.md` |
+| 19 | **gpt-oss-20b MoE**:harmony 格式、attention sinks + sliding window、YaRN;两个 CUDA bug 实战(prefill sinks NaN、GEMV 未初始化 smem);GSM8K 94.5% 对齐 llama.cpp;FP8 W8A8 / MXFP4 W4A16 量化 | `19-gpt-oss-moe.md`、`benchmarks/{fp8-quantization,mxfp4-and-llama-decode}.md` |
+| 20 | **稀疏 top-k MoE decode**:只算被路由的专家,decode 13.9→7.0ms,TP=2 下 decode/TTFT 全面快于 llama.cpp 同配置;gpt-oss 单卡 serving | `20-sparse-moe.md`、`benchmarks/sparse-moe.md` |
+| 21 | **decode CUDA Graph + GPU argmax**:整个 decode step 录成一个图回放(thread-local launch stream、retained-warmup 分配策略、NCCL capture);greedy 采样换 GPU argmax。TPOT 7.5→5.9ms(TP=1)/ 5.8ms(TP=2);TP=2 全面领先 llama(1.26-1.47×),TP=1 差距 2.5×→2.0× | `21-cuda-graph-decode.md` |
+
+**下一步候选(按预期收益排序):**
+
+| 候选 Phase | 内容 | 预期 |
+|---|---|---|
+| 22 | **非专家权重量化**:qkv/o + lm_head(1.16GB/token)仍是 BF16 | TPOT 再省 ~1.5ms |
+| 23 | **稀疏 prefill**(按专家 permute + grouped GEMM) | 长 prompt TTFT 51-75 → ~30ms |
+| 24 | server 侧 harmony channel 分离(`reasoning_content` 流式输出,对齐 llama-server 行为) | API 易用性 |
+| — | Speculative Decoding、多模态(原 16/19) | 推迟 |
+
 ## 里程碑总结
 
 | 里程碑 | Phase | 验收标准 |
 |--------|-------|---------|
 | ① GPT-2 推理 | 8 | CLI 输入 prompt, GPT-2 生成连贯文本, logits 与 PyTorch 一致 |
-| ② Qwen3-7B 推理 | 10 | 7B 模型中英文对话, 多轮 chat template 正确 |
+| ② Qwen3-8B 推理 | 10 | 8B 模型中英文对话, 多轮 chat template 正确 |
 | ③ E2E API | 13 | HTTP streaming API, Python OpenAI SDK 可调用, 10 并发正确 |
 | ④ 性能达标 | 15 | throughput >= 50% vLLM, profiling 报告完成 |
 | ⑤ 多卡推理 | 17 | TP=2/4 同组 GPU 推理正确, scaling benchmark 完成 |
-| ⑥ 多模态 | 19 | 图片输入 → 文字回答, API 端到端 |
+| ⑥ MoE 模型(实际) | 19 | gpt-oss-20b 端到端正确, GSM8K 与 llama.cpp 持平 ✅ |
+| ⑦ 性能反超(实际) | 20 | 同配置 decode 快于 llama.cpp(TP=2 达成;单卡是 Phase 21+ 目标) ✅ |
+| ⑧ 多模态 | 推迟 | 图片输入 → 文字回答, API 端到端 |
 
 ## 外部依赖清单
 

docs/05-attention.md

@@ -0,0 +1,92 @@
+# Phase 5: Naive Attention Kernel — Design Document
+
+## Goal
+
+实现标准 Multi-Head Attention（不做 Flash/Paged 优化），用组合式方法（GEMM + Softmax）完成。这是理解 attention 计算流程的基础，也是后续 Flash Attention 的 baseline。
+
+## 计算流程
+
+```
+Input: Q [B, H, S, D], K [B, H, S, D], V [B, H, S, D]
+       B=batch, H=num_heads, S=seq_len, D=head_dim
+
+1. scores = Q @ K^T / sqrt(D)           → [B, H, S, S]
+2. scores += causal_mask                 → 上三角置为 -inf
+3. weights = softmax(scores, dim=-1)     → [B, H, S, S]
+4. output = weights @ V                  → [B, H, S, D]
+```
+
+## 设计选择
+
+### 组合式实现（Phase 3 GEMM + Phase 4 Softmax）
+
+不写新的 fused CUDA kernel，而是复用已有的 matmul 和 softmax：
+- `scores = batched_matmul(Q, K^T)` — 需要支持 batched GEMM
+- `masked_fill(scores, causal_mask, -inf)` — 新的逐元素 kernel
+- `softmax(scores)` — 复用 Phase 4
+- `output = batched_matmul(weights, V)` — 复用 batched GEMM
+
+这意味着需要先扩展 matmul 支持 batched GEMM（cublasGemmStridedBatchedEx）。
+
+### Causal Mask
+
+不显式构造 mask 矩阵。写一个 kernel：
+```
+if (col > row + offset) score = -infinity
+```
+其中 offset 用于支持 KV cache 场景（decode 时 query 的 row 偏移）。
+
+### Batched GEMM via cuBLAS
+
+`cublasGemmStridedBatchedEx` 在一个 batch 维度上并行执行多个 GEMM：
+```
+C[b] = A[b] @ B[b]  for b = 0..batch_count
+stride_a = M * K, stride_b = K * N, stride_c = M * N
+```
+
+Attention 中 batch 维度 = B * H（batch_size × num_heads）。
+
+## 文件布局
+
+```
+csrc/attention/
+└── causal_mask.cu      # causal mask fill kernel
+
+crates/xserv-kernels/src/
+├── gemm.rs             # 扩展: batched_matmul
+├── attention.rs        # NEW: multi_head_attention()
+└── causal_mask.rs      # NEW: causal mask apply
+```
+
+## API 设计
+
+```rust
+/// Multi-head attention (naive, materializes S×S scores).
+/// q, k, v: [batch, num_heads, seq_len, head_dim]
+/// Returns: [batch, num_heads, seq_len, head_dim]
+pub fn attention(q: &Tensor, k: &Tensor, v: &Tensor, causal: bool) -> Tensor;
+
+/// Batched matmul: A[b] @ B[b] for all b.
+/// a: [..., M, K], b: [..., K, N] → [..., M, N]
+pub fn batched_matmul(a: &Tensor, b: &Tensor) -> Tensor;
+```
+
+## Test Plan
+
+- [x] batched_matmul: [4,8,32,64]×[4,8,64,32] → max_err 2.7e-7
+- [x] attention (non-causal): B=1,H=2,S=8,D=16 → max_err 4.5e-8
+- [x] attention (causal): B=1,H=2,S=16,D=32 → max_err 3.0e-8
+- [x] attention (causal, larger): B=2,H=4,S=64,D=64 → max_err 6.0e-8
+- [x] causal mask 语义: position 0 只能看到 token 0，output[0] == V[0] → exact
+
+## Takeaways
+
+1. **`to_device` 不应强制 contiguous**：最初 `to_device()` 会先调 `contiguous()`，而 GPU 的 `contiguous()` 又调 `to_device(Cpu)`，导致无限递归栈溢出。修复：`to_device()` 直接传输 raw storage，保留 strides/offset，用户需要时自己调 `contiguous()`。GPU `contiguous()` 现在走 GPU→CPU→CPU contiguous→CPU→GPU 路径——正确但低效，Phase 15 需要写 GPU contiguous kernel。
+
+2. **Batched GEMM via `cublasGemmStridedBatchedEx`**：row-major trick 同 Phase 3，额外参数是 stride（元素数，不是字节）。stride_a = M×K, stride_b = K×N, stride_c = M×N。注意初始版本错误地乘了 `elem_size`，cuBLAS 的 stride 单位是元素。
+
+3. **Attention 的组合式实现足够验证正确性**：没有写 fused kernel，而是复用 `batched_matmul` + `scale` + `causal_mask` + `softmax`。精度极好（max_err < 1e-7），因为每步都在 FP32 中完成。缺点是 S×S score 矩阵完全 materialize（O(S²) 显存），Flash Attention 会解决。
+
+4. **Scale kernel 的必要性**：原本想在 CPU 上做 scale（round-trip），但那太慢了。加了 `scale_f32/bf16` 逐元素 CUDA kernel。未来可以把 scale 合进 GEMM 的 alpha 参数，省一次 kernel launch。
+
+5. **Causal mask 的 offset 设计**：`col > row + offset` 中的 offset 为 KV cache 场景预留。Decode 时 Q 只有 1 行但 KV cache 有前 S 行，offset = kv_len - q_len 确保 decode query 能看到所有 cached tokens。