tools: single-stream decode benchmark vs llama.cpp

xserv_vs_llama.py runs each server one at a time on the same GPUs (drains VRAM
between), streams identical prompts through /v1/chat/completions, and reports
median TTFT/TPOT/throughput. Counts llama's reasoning_content as real decode
tokens so the gpt-oss CoT is measured fairly.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

.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

@@ -6,6 +6,8 @@ members = [
     "crates/xserv-kernels",
     "crates/xserv-model",
     "crates/xserv-tokenizer",
+    "crates/xserv-server",
+    "crates/xserv-distributed",
 ]
 
 [workspace.package]
@@ -16,7 +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 +1,197 @@
+# 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。当前在单张 RTX 5090 上可以
+跑通 **Qwen3-8B**（BF16），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）
+- **性能**（RTX 5090，Qwen3-8B BF16，贪心解码，单流）：约 **56 tok/s**，约为 HF transformers 的 1.4×、llama.cpp 的 ~0.6×
+- **精度**：在 AIME 2025 / GSM8K 上与 llama.cpp 同权重对比基本持平（数值保真度验证通过）
+- **服务**：OpenAI 兼容 `/v1/chat/completions`，支持 SSE 流式输出
+- **关键能力**：自写 GEMM / Flash-Attention 2(SM120) / Paged-Attention kernel、
+  分页 KV cache（含 **CPU 换出/换入** 弹性显存）、连续批处理（continuous batching）、
+  CUDA Graph 解码、按显存自适应的 KV 池
+
+> 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
+│   ├── embedding/         #   embedding lookup / RoPE / transpose
+│   └── 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-model/       # 模型定义（GPT-2 / Qwen3）、权重加载、KV cache、采样
+│   └── xserv-server/      # tokio + axum HTTP 服务、调度器
+├── 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 0–18：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）**；
+并加入了 **llama.cpp 对比基准** 与 **KV CPU 换出** 等基础设施。
+
+后续方向：PP microbatch/1F1B 流水线重叠（吞吐收益）、2D TP×PP、投机解码、量化（FP8 / INT8）、多模态。
+
+## 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 +85,40 @@ 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) {
+    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));
+    });
+}
+
 /// 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 +45,14 @@ pub fn current_device() -> Result<u32> {
 }
 
 pub fn device_info(device: u32) -> Result<DeviceInfo> {
-    // Get device name from cudaGetDeviceProperties (only use the name field).
+    // Heap-allocate oversized buffer for cudaDeviceProp (layout varies by CUDA version).
-    let mut prop = unsafe { std::mem::zeroed::<ffi::CudaDeviceProp>() };
+    // CUDA 12.x struct is ~5-6 KB; use 32 KB to guard against future growth.
-    error::check(unsafe { ffi::cudaGetDeviceProperties(&mut prop, device as i32) })?;
+    let mut prop_buf = vec![0u8; 32768];
-    let name = unsafe { CStr::from_ptr(prop.name.as_ptr()) }
+    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 +13,16 @@ pub const CUDA_MEMCPY_D2D: i32 = 3;
 pub const CUDA_SUCCESS: i32 = 0;
 pub const CUDA_ERROR_OUT_OF_MEMORY: i32 = 2;
 
-#[repr(C)]
+/// cudaStreamCaptureMode::cudaStreamCaptureModeGlobal
-pub struct CudaDeviceProp {
+pub const CUDA_STREAM_CAPTURE_MODE_GLOBAL: i32 = 0;
-    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],
-}
 
 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 +39,7 @@ 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;
@@ -62,6 +50,18 @@ 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,

crates/xserv-cuda/src/graph.rs

@@ -0,0 +1,98 @@
+//! 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();
+        error::check(unsafe {
+            ffi::cudaStreamBeginCapture(
+                stream.as_raw(),
+                ffi::CUDA_STREAM_CAPTURE_MODE_GLOBAL,
+            )
+        })
+    }
+
+    /// 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;

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 +22,13 @@ 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 {
@@ -87,6 +102,70 @@ 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 +178,29 @@ 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() {
+        if self.owned && !self.ptr.is_null() {
-            unsafe { ffi::cudaFree(self.ptr) };
+            if self.pooled {
+                crate::allocator::return_to_pool(self.ptr, self.len);
+            } else {
+                unsafe { ffi::cudaFree(self.ptr) };
+            }
         }
     }
 }

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 +1,82 @@
+//! 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 +1,161 @@
+//! 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::device;
+use xserv_cuda::GpuBuffer;
+
+pub use ffi::NcclUniqueId as UniqueId;
+
+/// The CUDA "null" (default) stream. The model's kernels and cuBLAS calls run
+/// on it, so issuing NCCL on the same stream keeps AllReduce correctly ordered
+/// after the producing matmul and before the consuming kernel — no extra sync.
+const NULL_STREAM: xserv_cuda::ffi::CudaStream = std::ptr::null_mut();
+
+/// 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,
+                    NULL_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, NULL_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, NULL_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 +1,50 @@
+//! 2-GPU AllReduce smoke test. Skips if fewer than 2 GPUs are present.
+
+use half::bf16;
+use std::thread;
+use xserv_cuda::{device, GpuBuffer};
+use xserv_distributed::{get_unique_id, TpContext};
+
+#[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 +1,62 @@
+//! 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::{device, GpuBuffer};
+use xserv_distributed::{get_unique_id, PpContext};
+
+#[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)
@@ -16,13 +17,23 @@ 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/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

@@ -12,13 +12,18 @@ unsafe extern "C" {
     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 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::zeros(x.shape(), x.dtype(), x.device());
+    let out = Tensor::empty(x.shape(), x.dtype(), x.device());
-    let n = x.numel() as i32;
+    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, std::ptr::null_mut()),
@@ -26,7 +31,6 @@ fn dispatch_unary(x: &Tensor, f32_fn: unsafe extern "C" fn(*const c_void, *mut c
             _ => panic!("unsupported dtype"),
         }
     }
-    xserv_cuda::device::synchronize().unwrap();
     out
 }
 
@@ -37,8 +41,10 @@ fn dispatch_binary(a: &Tensor, b: &Tensor,
     assert!(a.is_contiguous() && b.is_contiguous());
     assert!(matches!(a.device(), Device::Cuda(_)));
     assert_eq!(a.dtype(), b.dtype());
-    let out = Tensor::zeros(a.shape(), a.dtype(), a.device());
+    let out = Tensor::empty(a.shape(), a.dtype(), a.device());
-    let n = a.numel() as i32;
+    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, std::ptr::null_mut()),
@@ -46,7 +52,6 @@ fn dispatch_binary(a: &Tensor, b: &Tensor,
             _ => panic!("unsupported dtype"),
         }
     }
-    xserv_cuda::device::synchronize().unwrap();
     out
 }
 
@@ -55,8 +60,10 @@ pub fn silu(x: &Tensor) -> Tensor { dispatch_unary(x, launch_silu_f32, launch_si
 
 pub fn scale(x: &Tensor, scale_val: f32) -> Tensor {
     assert!(x.is_contiguous() && matches!(x.device(), Device::Cuda(_)));
-    let out = Tensor::zeros(x.shape(), x.dtype(), x.device());
+    let out = Tensor::empty(x.shape(), x.dtype(), x.device());
-    let n = x.numel() as i32;
+    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_scale_f32(x.data_ptr() as _, out.data_ptr() as *mut c_void, scale_val, n, std::ptr::null_mut()),
@@ -64,9 +71,59 @@ pub fn scale(x: &Tensor, scale_val: f32) -> Tensor {
             _ => panic!("unsupported dtype for scale"),
         }
     }
-    xserv_cuda::device::synchronize().unwrap();
     out
 }
 
 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) }
+
+/// 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,
+            std::ptr::null_mut(),
+        );
+    }
+    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,
+            std::ptr::null_mut(),
+        );
+    }
+    out
+}

crates/xserv-kernels/src/argmax.rs

@@ -0,0 +1,65 @@
+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,
+            std::ptr::null_mut(),
+        );
+    }
+
+    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

@@ -10,6 +10,127 @@ unsafe extern "C" {
                               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_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,
+    );
+}
+
+/// 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,
+        );
+    }
 }
 
 fn apply_causal_mask(scores: &Tensor, offset: usize) {
@@ -33,7 +154,6 @@ fn apply_causal_mask(scores: &Tensor, offset: usize) {
             _ => panic!("unsupported dtype for causal mask"),
         }
     }
-    xserv_cuda::device::synchronize().unwrap();
 }
 
 /// Multi-head attention (naive, materializes S×S score matrix).
@@ -75,3 +195,278 @@ pub fn attention(q: &Tensor, k: &Tensor, v: &Tensor, causal: bool) -> Tensor {
     // 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)
+            std::ptr::null_mut(),
+        );
+    }
+
+    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 },
+            std::ptr::null_mut(),
+        );
+    }
+
+    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,
+            std::ptr::null_mut(),
+        );
+    }
+
+    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,
+            std::ptr::null_mut(),
+        );
+    }
+
+    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,
+            std::ptr::null_mut(),
+        );
+    }
+
+    output
+}

crates/xserv-kernels/src/dispatch.rs

@@ -0,0 +1,118 @@
+//! 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

@@ -4,9 +4,9 @@ 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);
+                            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, stream: *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 +18,9 @@ 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 +29,29 @@ 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})");
+    }
+
+    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 _,
                 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, std::ptr::null_mut(),
             ),
             DType::BF16 => launch_embedding_bf16(
                 table.data_ptr() as _, ids_gpu.as_ptr() as _,
                 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, std::ptr::null_mut(),
             ),
             _ => panic!("unsupported dtype for embedding"),
         }
     }
-    xserv_cuda::device::synchronize().unwrap();
     out
 }

crates/xserv-kernels/src/gemm.rs

@@ -1,7 +1,16 @@
+use std::cell::RefCell;
 use std::ffi::c_void;
 use xserv_cuda::error::{self, Result};
+use xserv_cuda::GpuBuffer;
 use xserv_tensor::{DType, Device, Tensor};
 
+const CUBLAS_WORKSPACE_BYTES: usize = 32 * 1024 * 1024;
+
+// 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);
+}
+
 #[derive(Debug, Clone, Copy)]
 pub enum GemmBackend {
     Naive,
@@ -18,7 +27,7 @@ unsafe extern "C" {
 }
 
 // --- FFI: cuBLAS ---
-type CublasHandle = *mut c_void;
+pub type CublasHandle = *mut c_void;
 
 #[allow(non_upper_case_globals)]
 const CUBLAS_OP_N: i32 = 0;
@@ -34,6 +43,7 @@ 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,
@@ -63,13 +73,25 @@ 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) })
     }
 }
 
@@ -78,9 +100,34 @@ 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.
@@ -97,7 +144,9 @@ pub fn matmul(a: &Tensor, b: &Tensor, backend: GemmBackend) -> Tensor {
     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;
@@ -113,7 +162,6 @@ pub fn matmul(a: &Tensor, b: &Tensor, backend: GemmBackend) -> Tensor {
                     _ => panic!("unsupported dtype for naive GEMM"),
                 }
             }
-            xserv_cuda::device::synchronize().unwrap();
         }
         GemmBackend::Tiled => {
             unsafe {
@@ -123,40 +171,44 @@ pub fn matmul(a: &Tensor, b: &Tensor, backend: GemmBackend) -> Tensor {
                     _ => panic!("unsupported dtype for tiled GEMM"),
                 }
             }
-            xserv_cuda::device::synchronize().unwrap();
         }
         GemmBackend::CuBlas => {
-            // cuBLAS uses column-major, but we have row-major tensors.
+            if m == 1 && dtype == DType::BF16 && n >= 256 {
-            // Trick: compute C^T = B^T @ A^T, which gives us C in row-major.
+                let mut fp32_buf = xserv_cuda::allocator::cached_alloc(n * 4).unwrap();
-            // cuBLAS sees our row-major data as column-major transposed.
+                unsafe {
-            let ctx = CublasContext::new().unwrap();
+                    launch_gemv_bf16(
-            let alpha = 1.0f32;
+                        a_ptr, b_ptr, c_ptr,
-            let beta = 0.0f32;
+                        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 {
+                let (a_type, b_type, c_type) = match dtype {
-                DType::F32 => (CUDA_R_32F, CUDA_R_32F, CUDA_R_32F),
+                    DType::F32 => (CUDA_R_32F, CUDA_R_32F, CUDA_R_32F),
-                DType::BF16 => (CUDA_R_16BF, CUDA_R_16BF, CUDA_R_16BF),
+                    DType::BF16 => (CUDA_R_16BF, CUDA_R_16BF, CUDA_R_16BF),
-                _ => panic!("unsupported dtype for cuBLAS GEMM"),
+                    _ => panic!("unsupported dtype for cuBLAS GEMM"),
-            };
+                };
 
-            unsafe {
+                with_cublas(|handle| unsafe {
-                cublasSetStream_v2(ctx.handle, null_stream);
+                    cublasSetStream_v2(handle, null_stream);
-                // Row-major trick: swap A/B and transpose flags
+                    error::check(cublasGemmEx(
-                // C(row-major) = A @ B  <=>  C^T(col-major) = B^T @ A^T
+                        handle,
-                error::check(cublasGemmEx(
+                        CUBLAS_OP_N, CUBLAS_OP_N,
-                    ctx.handle,
+                        n as i32, m as i32, k as i32,
-                    CUBLAS_OP_N, CUBLAS_OP_N,
+                        &alpha as *const f32 as *const c_void,
-                    n as i32, m as i32, k as i32,
+                        b_ptr, b_type, n as i32,
-                    &alpha as *const f32 as *const c_void,
+                        a_ptr, a_type, k as i32,
-                    b_ptr, b_type, n as i32,  // B as col-major = B^T
+                        &beta as *const f32 as *const c_void,
-                    a_ptr, a_type, k as i32,  // A as col-major = A^T
+                        c_ptr, c_type, n as i32,
-                    &beta as *const f32 as *const c_void,
+                        CUBLAS_COMPUTE_32F,
-                    c_ptr, c_type, n as i32,  // C as col-major = C^T
+                        -1,
-                    CUBLAS_COMPUTE_32F,
+                    )).expect("cuBLAS GEMM failed");
-                    -1, // default algo
+                });
-                )).expect("cuBLAS GEMM failed");
             }
-            xserv_cuda::device::synchronize().unwrap();
         }
     }
 
@@ -190,7 +242,8 @@ pub fn batched_matmul(a: &Tensor, b: &Tensor) -> Tensor {
     let mut out_shape: Vec<usize> = a.shape()[..ndim - 2].to_vec();
     out_shape.push(m);
     out_shape.push(n);
-    let c = Tensor::zeros(&out_shape, a.dtype(), a.device());
+    // 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 {
@@ -206,12 +259,11 @@ pub fn batched_matmul(a: &Tensor, b: &Tensor) -> Tensor {
     let stride_b = (k * n) as i64;
     let stride_c = (m * n) as i64;
 
-    let ctx = CublasContext::new().unwrap();
+    with_cublas(|handle| unsafe {
-    unsafe {
+        cublasSetStream_v2(handle, std::ptr::null_mut());
-        cublasSetStream_v2(ctx.handle, std::ptr::null_mut());
         // Row-major trick: C = A @ B ⟺ C^T = B^T @ A^T (col-major)
         error::check(cublasGemmStridedBatchedEx(
-            ctx.handle,
+            handle,
             CUBLAS_OP_N, CUBLAS_OP_N,
             n as i32, m as i32, k as i32,
             &alpha as *const f32 as *const c_void,
@@ -223,7 +275,6 @@ pub fn batched_matmul(a: &Tensor, b: &Tensor) -> Tensor {
             CUBLAS_COMPUTE_32F,
             -1,
         )).expect("cuBLAS batched GEMM failed");
-    }
+    });
-    xserv_cuda::device::synchronize().unwrap();
     c
 }

crates/xserv-kernels/src/layernorm.rs

@@ -17,7 +17,9 @@ 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() {
@@ -34,6 +36,5 @@ pub fn layernorm(x: &Tensor, gamma: &Tensor, beta: &Tensor, eps: f32) -> Tensor
             _ => panic!("unsupported dtype for layernorm"),
         }
     }
-    xserv_cuda::device::synchronize().unwrap();
     out
 }

crates/xserv-kernels/src/lib.rs

@@ -1,17 +1,29 @@
 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::{add, gelu, mul, scale, silu};
+pub use activation::{add, gelu, gpt_oss_glu, mul, scale, silu, silu_mul};
-pub use attention::attention;
+pub use argmax::{argmax_bf16_single, argmax_bf16_to_host};
+pub use transpose::{merge_heads_gpu, repeat_kv_gpu, reshape_heads_gpu, strided_to_contiguous_gpu, transpose_for_rope_gpu, transpose_from_rope_gpu};
+pub use attention::{attention, decode_attention, flash_attention, flash_attention_sinks, paged_decode_attention, paged_decode_attention_sinks, reshape_and_cache_bf16, reshape_and_cache_batched_bf16};
 pub use embedding::embedding;
 pub use gemm::{batched_matmul, matmul, GemmBackend};
 pub use layernorm::layernorm;
-pub use rmsnorm::rmsnorm;
+pub use rmsnorm::{add_rmsnorm, rmsnorm};
 pub use rope::{rope_inplace, RopeCache};
 pub use softmax::softmax;
+
+/// 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 +1,223 @@
+use std::ffi::c_void;
+use xserv_tensor::{DType, Device, Tensor};
+
+use crate::gemm::{cublas_handle, CublasHandle};
+
+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 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);
+
+    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,
+            std::ptr::null_mut(),
+        );
+    }
+
+    (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,
+            std::ptr::null_mut(),
+        );
+    }
+
+    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,
+            std::ptr::null_mut(),
+        );
+    }
+}
+
+/// 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,
+            std::ptr::null_mut(),
+        );
+    }
+
+    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, std::ptr::null_mut());
+        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 +1,485 @@
+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;
+const CUBLASLT_MATMUL_DESC_A_SCALE_MODE: i32 = 31;
+const CUBLASLT_MATMUL_DESC_B_SCALE_MODE: i32 = 32;
+
+// MatrixLayout attributes
+const CUBLASLT_MATRIX_LAYOUT_BATCH_COUNT: i32 = 5;
+const CUBLASLT_MATRIX_LAYOUT_STRIDED_BATCH_OFFSET: i32 = 6;
+
+// Scale modes
+const CUBLASLT_MATMUL_MATRIX_SCALE_SCALAR: i32 = 0;
+const CUBLASLT_MATMUL_MATRIX_SCALE_OUTER_VEC_32F: i32 = 3;
+
+// 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,
+            std::ptr::null_mut(),
+        );
+    }
+
+    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,
+            std::ptr::null_mut(),
+        );
+    }
+
+    (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,
+                std::ptr::null_mut(),
+            );
+            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,
+            std::ptr::null_mut(),
+        );
+    }
+
+    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,
+            std::ptr::null_mut(),
+        );
+    }
+    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,
+            std::ptr::null_mut(),
+        );
+    }
+    out
+}

crates/xserv-kernels/src/rmsnorm.rs

@@ -6,6 +6,9 @@ unsafe extern "C" {
                           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 {
@@ -17,7 +20,9 @@ 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() {
@@ -32,6 +37,43 @@ pub fn rmsnorm(x: &Tensor, gamma: &Tensor, eps: f32) -> Tensor {
             _ => 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,
+            std::ptr::null_mut(),
+        );
+    }
+
+    (normed_out, sum_out)
+}

crates/xserv-kernels/src/rope.rs

@@ -34,7 +34,81 @@ impl RopeCache {
                 max_seq_len as i32, half_dim as i32, theta, std::ptr::null_mut(),
             );
         }
-        xserv_cuda::device::synchronize().unwrap();
+
+        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 }
     }
@@ -59,7 +133,7 @@ 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();
 
     unsafe {
@@ -81,5 +155,4 @@ pub fn rope_inplace(x: &Tensor, cache: &RopeCache, positions: &[u32]) {
             _ => panic!("unsupported dtype for rope"),
         }
     }
-    xserv_cuda::device::synchronize().unwrap();
 }

crates/xserv-kernels/src/softmax.rs

@@ -14,7 +14,9 @@ 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() {
@@ -29,6 +31,5 @@ pub fn softmax(x: &Tensor) -> Tensor {
             _ => panic!("unsupported dtype for softmax"),
         }
     }
-    xserv_cuda::device::synchronize().unwrap();
     out
 }

crates/xserv-kernels/src/transpose.rs

@@ -0,0 +1,142 @@
+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, std::ptr::null_mut(),
+        );
+    }
+    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, std::ptr::null_mut(),
+        );
+    }
+    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, std::ptr::null_mut(),
+        );
+    }
+    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, std::ptr::null_mut(),
+        );
+    }
+    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, std::ptr::null_mut(),
+        );
+    }
+    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, std::ptr::null_mut(),
+            ),
+            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, std::ptr::null_mut(),
+            ),
+            _ => panic!("strided_to_contiguous_gpu: unsupported dtype {:?}", x.dtype()),
+        }
+    }
+    out
+}

crates/xserv-kernels/tests/gemm_test.rs

@@ -121,6 +121,20 @@ fn test_gemm_cublas_bf16_small() { run_gemm_test_bf16(GemmBackend::CuBlas, 4, 4,
 #[test]
 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]

crates/xserv-kernels/tests/ops_test.rs

@@ -74,10 +74,10 @@ 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 x0 = x[base + i];
-                let x1 = x[base + 2 * i + 1];
+                let x1 = x[base + i + half_dim];
-                x[base + 2 * i]     = x0 * cos_val - x1 * sin_val;
+                x[base + i] = x0 * cos_val - x1 * sin_val;
-                x[base + 2 * i + 1] = x0 * sin_val + x1 * cos_val;
+                x[base + i + half_dim] = x1 * cos_val + x0 * sin_val;
             }
         }
     }

crates/xserv-model/Cargo.toml

@@ -8,7 +8,11 @@ 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 +1,310 @@
+use std::path::PathBuf;
+use std::sync::Arc;
+use std::time::Instant;
+
+use xserv_distributed::{TpContext, UniqueId, get_unique_id};
+use xserv_model::{loader, GptOss, ModelConfig, PagedKVCache, BLOCK_SIZE};
+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 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 = model.forward_decode_paged(&[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();
+
+    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 _ = model.forward_decode_paged(&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);
+    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-qwen3.rs

@@ -1,14 +1,14 @@
 use std::path::PathBuf;
 use std::time::Instant;
 use xserv_model::qwen3::sample_greedy;
-use xserv_model::{loader, KVCache, ModelConfig, Qwen3};
+use xserv_model::{loader, DecodeGraphState, GpuKVCache, ModelConfig, Qwen3};
 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]");
+        eprintln!("Usage: bench-qwen3 <model-dir> [--gen-tokens N] [--cuda-graph]");
         std::process::exit(1);
     }
     let model_dir = PathBuf::from(&args[1]);
@@ -18,6 +18,7 @@ fn main() {
         .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();
 
@@ -31,12 +32,21 @@ fn main() {
     // Warmup
     {
         let ids = tokenizer.encode("warmup");
-        let mut cache = KVCache::new(
+        let mut cache = GpuKVCache::new(&config, 256, DType::BF16, 0);
-            config.num_layers(), config.num_kv_heads(), config.head_dim(),
+        let _ = model.forward_gpu_cache(&ids, &mut cache);
-            DType::BF16, Device::Cuda(0),
-        );
-        let _ = model.forward_with_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![
@@ -97,14 +107,17 @@ fn main() {
         let input_ids = tokenizer.encode(prompt);
         let input_len = input_ids.len();
 
-        let mut cache = KVCache::new(
+        let mut cache = GpuKVCache::new(&config, 256, DType::BF16, 0);
-            config.num_layers(), config.num_kv_heads(), config.head_dim(),
+
-            DType::BF16, Device::Cuda(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_with_cache(&input_ids, &mut cache);
+        let logits = model.forward_gpu_cache(&input_ids, &mut cache);
         let first_token = sample_greedy(&logits);
         let ttft_us = t0.elapsed().as_micros();
 
@@ -115,8 +128,35 @@ fn main() {
         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);
+            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; }
@@ -148,12 +188,14 @@ fn main() {
         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,
-            &generated_text.replace('\n', " ")[..generated_text.len().min(60)]
+            truncated
         );
     }
     println!("]");

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

@@ -0,0 +1,194 @@
+//! 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::{loader, ModelConfig, PagedKVCache, Qwen3, BLOCK_SIZE};
+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/xserv-chat.rs

@@ -0,0 +1,937 @@
+use std::io::{self, IsTerminal, Read, Write};
+use std::path::PathBuf;
+
+use std::sync::{mpsc, Arc};
+use std::thread;
+
+use xserv_model::{loader, sample, sample_greedy_penalized, GptOss, ModelConfig, PagedKVCache, Qwen3, SamplingParams, BLOCK_SIZE};
+use xserv_tensor::{DType, Device};
+use xserv_tokenizer::Tokenizer;
+
+enum ChatModel {
+    Qwen3(Qwen3),
+    GptOss(GptOss),
+}
+
+impl ChatModel {
+    fn forward_prefill_paged(&self, tokens: &[u32], slot: usize, cache: &mut PagedKVCache) -> xserv_tensor::Tensor {
+        match self {
+            ChatModel::Qwen3(m) => m.forward_prefill_paged(tokens, slot, cache),
+            ChatModel::GptOss(m) => m.forward_prefill_paged(tokens, slot, cache),
+        }
+    }
+    fn forward_decode_paged(&self, tokens: &[u32], positions: &[usize], slots: &[usize], cache: &mut PagedKVCache) -> xserv_tensor::Tensor {
+        match self {
+            ChatModel::Qwen3(m) => m.forward_decode_paged(tokens, positions, slots, cache),
+            ChatModel::GptOss(m) => m.forward_decode_paged(tokens, positions, slots, cache),
+        }
+    }
+}
+
+// TP worker infrastructure (reused from tp_engine pattern)
+#[derive(Clone)]
+enum TpCommand {
+    Register(usize),
+    Free(usize),
+    Prefill { tokens: Vec<u32>, slot: usize },
+    Decode { tokens: Vec<u32>, positions: Vec<usize>, slots: Vec<usize> },
+}
+
+struct TpHandle {
+    cmd_txs: Vec<mpsc::Sender<TpCommand>>,
+    ack_rx: mpsc::Receiver<()>,
+}
+
+impl TpHandle {
+    fn send(&self, cmd: TpCommand) {
+        for tx in &self.cmd_txs {
+            tx.send(cmd.clone()).ok();
+        }
+    }
+    fn wait(&self) {
+        for _ in 0..self.cmd_txs.len() {
+            self.ack_rx.recv().ok();
+        }
+    }
+}
+
+fn tp_worker_loop(
+    rank: usize, world: usize,
+    id: xserv_distributed::UniqueId,
+    model_dir: std::path::PathBuf,
+    config: ModelConfig,
+    max_seq_len: usize,
+    cmd_rx: mpsc::Receiver<TpCommand>,
+    ack_tx: mpsc::Sender<()>,
+) {
+    let tp = Arc::new(xserv_distributed::TpContext::init(rank, world, id, rank as u32));
+    let weights = loader::load_model_dir(&model_dir, Device::Cpu);
+    let model = if config.is_moe() {
+        ChatModel::GptOss(GptOss::from_weights_tp(config.clone(), weights, rank, world, rank as u32, Some(tp)))
+    } else {
+        ChatModel::Qwen3(Qwen3::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 + 8;
+    let mut cache = PagedKVCache::new_tp(
+        &config, local_kv, total_blocks, 0, 1, max_blocks_per_seq, DType::BF16, rank as u32,
+    );
+    while let Ok(cmd) = cmd_rx.recv() {
+        match cmd {
+            TpCommand::Register(slot) => { let _ = cache.register_sequence(slot); }
+            TpCommand::Free(slot) => cache.free_sequence(slot),
+            TpCommand::Prefill { tokens, slot } => {
+                let _ = model.forward_prefill_paged(&tokens, slot, &mut cache);
+            }
+            TpCommand::Decode { tokens, positions, slots } => {
+                let _ = model.forward_decode_paged(&tokens, &positions, &slots, &mut cache);
+            }
+        }
+        let _ = ack_tx.send(());
+    }
+}
+
+const SLOT: usize = 0;
+
+struct CliOptions {
+    model_dir: PathBuf,
+    max_tokens: usize,
+    max_seq_len: usize,
+    tp: usize,
+    sampling: SamplingParams,
+    system_prompt: Option<String>,
+    enable_thinking: bool,
+    color: bool,
+}
+
+enum Finish {
+    Stop { token_id: u32 },
+    Length,
+}
+
+enum Line {
+    Text(String),
+    Eof,
+}
+
+/// RAII terminal raw-mode guard. Disables canonical mode + echo (keeps output
+/// post-processing and signals), so we read keystrokes ourselves and edit the
+/// line UTF-8-aware. Restores the original termios on drop.
+struct RawMode {
+    orig: libc::termios,
+}
+
+impl RawMode {
+    fn enable() -> Option<Self> {
+        unsafe {
+            let mut orig: libc::termios = std::mem::zeroed();
+            if libc::tcgetattr(libc::STDIN_FILENO, &mut orig) != 0 {
+                return None;
+            }
+            let mut raw = orig;
+            raw.c_lflag &= !(libc::ICANON | libc::ECHO);
+            raw.c_cc[libc::VMIN as usize] = 1;
+            raw.c_cc[libc::VTIME as usize] = 0;
+            if libc::tcsetattr(libc::STDIN_FILENO, libc::TCSANOW, &raw) != 0 {
+                return None;
+            }
+            Some(RawMode { orig })
+        }
+    }
+}
+
+impl Drop for RawMode {
+    fn drop(&mut self) {
+        unsafe {
+            libc::tcsetattr(libc::STDIN_FILENO, libc::TCSANOW, &self.orig);
+        }
+    }
+}
+
+/// Read one line with UTF-8/CJK-aware editing. In a TTY this enters raw mode and
+/// handles keystrokes so Backspace deletes a whole character (not a byte), and
+/// multi-byte input (汉字/日本語/한글) renders correctly. Non-TTY (piped) input
+/// falls back to a plain cooked read.
+fn read_line_edited(prompt: &str) -> Line {
+    let cooked = || -> Line {
+        print!("{prompt}");
+        io::stdout().flush().ok();
+        let mut s = String::new();
+        match io::stdin().read_line(&mut s) {
+            Ok(0) | Err(_) => Line::Eof,
+            Ok(_) => Line::Text(s),
+        }
+    };
+
+    if !io::stdin().is_terminal() {
+        return cooked();
+    }
+    let Some(_raw) = RawMode::enable() else {
+        return cooked();
+    };
+
+    // Single-line editor: on every edit, rewrite the whole line so the terminal
+    // renders correct (incl. double-width CJK) glyphs; \x1b[K clears leftovers.
+    let redraw = |buf: &str| {
+        print!("\r{prompt}{buf}\x1b[K");
+        io::stdout().flush().ok();
+    };
+
+    let mut buf = String::new();
+    redraw(&buf);
+    let mut stdin = io::stdin().lock();
+    let mut byte = [0u8; 1];
+
+    loop {
+        if stdin.read(&mut byte).unwrap_or(0) == 0 {
+            // EOF on the stream.
+            if buf.is_empty() {
+                return Line::Eof;
+            }
+            break;
+        }
+        match byte[0] {
+            b'\r' | b'\n' => {
+                println!();
+                break;
+            }
+            0x7f | 0x08 => {
+                // Backspace: drop one whole char (String::pop is char-aware).
+                buf.pop();
+                redraw(&buf);
+            }
+            0x04 => {
+                // Ctrl-D: EOF only when the line is empty.
+                if buf.is_empty() {
+                    return Line::Eof;
+                }
+            }
+            0x1b => {
+                // Escape sequence (arrows, etc.): consume and ignore the 2 bytes
+                // of a typical CSI sequence so they don't land in the buffer.
+                let mut seq = [0u8; 2];
+                let _ = stdin.read(&mut seq);
+            }
+            b if b < 0x20 => { /* other control bytes: ignore */ }
+            b if b < 0x80 => {
+                buf.push(b as char);
+                redraw(&buf);
+            }
+            b => {
+                // UTF-8 multi-byte: read the continuation bytes for this char.
+                let extra = if b >= 0xF0 { 3 } else if b >= 0xE0 { 2 } else { 1 };
+                let mut bytes = vec![b];
+                let mut cont = [0u8; 1];
+                let mut ok = true;
+                for _ in 0..extra {
+                    if stdin.read(&mut cont).unwrap_or(0) == 0 {
+                        ok = false;
+                        break;
+                    }
+                    bytes.push(cont[0]);
+                }
+                if ok {
+                    if let Ok(s) = std::str::from_utf8(&bytes) {
+                        buf.push_str(s);
+                        redraw(&buf);
+                    }
+                }
+            }
+        }
+    }
+    Line::Text(buf)
+}
+
+fn main() {
+    let opts = parse_args();
+
+    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(&opts.model_dir.join("config.json"));
+    let model_type = config.model_type.as_deref().unwrap_or("unknown");
+    let is_moe = config.is_moe();
+
+    let max_seq_len = opts.max_seq_len.min(config.max_seq_len()).max(1);
+    eprintln!(
+        "Model: {model_type}{}, layers={}, hidden={}, heads={}/{} kv, vocab={}, max_seq_len={}",
+        if is_moe { " (MoE)" } else { "" },
+        config.num_layers(),
+        config.hidden(),
+        config.num_heads(),
+        config.num_kv_heads(),
+        config.vocab_size,
+        max_seq_len
+    );
+
+    let world = opts.tp;
+    if world > 1 {
+        assert!(
+            config.num_kv_heads() % world == 0,
+            "num_kv_heads {} not divisible by tp {world}", config.num_kv_heads()
+        );
+    }
+
+    let (model, mut cache, tp_handle) = if world > 1 {
+        let id = xserv_distributed::get_unique_id();
+        let (ack_tx, ack_rx) = mpsc::channel::<()>();
+        let mut cmd_txs = 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 = opts.model_dir.clone();
+            let config = config.clone();
+            thread::spawn(move || {
+                tp_worker_loop(rank, world, id, model_dir, config, max_seq_len, ctx_rx, ack_tx);
+            });
+        }
+        eprintln!("Loading weights (tp={world})...");
+        let tp = Arc::new(xserv_distributed::TpContext::init(0, world, id, 0));
+        let weights = loader::load_model_dir(&opts.model_dir, Device::Cpu);
+        eprintln!("Loaded {} tensors", weights.len());
+        let m = if is_moe {
+            ChatModel::GptOss(GptOss::from_weights_tp(config.clone(), weights, 0, world, 0, Some(tp)))
+        } else {
+            ChatModel::Qwen3(Qwen3::from_weights_tp(config.clone(), weights, 0, world, 0, 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 + 8;
+        let c = PagedKVCache::new_tp(&config, local_kv, total_blocks, 0, 1, max_blocks_per_seq, DType::BF16, 0);
+        let h = TpHandle { cmd_txs, ack_rx };
+        (m, c, Some(h))
+    } else {
+        eprintln!("Loading weights...");
+        let weights = loader::load_model_dir(&opts.model_dir, Device::Cuda(0));
+        eprintln!("Loaded {} tensors", weights.len());
+        let m = if is_moe {
+            ChatModel::GptOss(GptOss::from_weights(config.clone(), weights))
+        } else {
+            ChatModel::Qwen3(Qwen3::from_weights(config.clone(), weights))
+        };
+        let c = new_paged_cache(&config, max_seq_len);
+        (m, c, None)
+    };
+
+    let tokenizer = Tokenizer::from_file(&opts.model_dir.join("tokenizer.json"));
+    if let Some(h) = &tp_handle { h.send(TpCommand::Register(SLOT)); h.wait(); }
+    cache.register_sequence(SLOT).expect("register chat slot");
+    let use_color = opts.color && io::stdout().is_terminal();
+
+    eprintln!("Ready (paged KV cache, tp={world}).");
+    eprintln!("Commands: /exit, /quit, /clear\n");
+
+    // gpt-oss multi-turn history of (user, assistant-final) text. Harmony
+    // requires re-rendering the conversation each turn with prior analysis
+    // dropped, so the moe path re-prefills from this rather than reusing an
+    // incremental KV cache (which would accumulate CoT + <|return|> and collapse
+    // at longer context). Qwen3 ignores this and keeps the incremental cache.
+    let mut moe_history: Vec<(String, String)> = Vec::new();
+
+    loop {
+        let line = match read_line_edited("user> ") {
+            Line::Eof => break,
+            Line::Text(s) => s,
+        };
+        let input = line.trim();
+        if input.is_empty() {
+            continue;
+        }
+
+        match input {
+            "/exit" | "/quit" | "exit" | "quit" => break,
+            "/clear" => {
+                reset_slot(&mut cache, &tp_handle);
+                moe_history.clear();
+                eprintln!("history and KV cache cleared");
+                continue;
+            }
+            "/help" => {
+                print_help();
+                continue;
+            }
+            _ => {}
+        }
+
+        if is_moe {
+            // Harmony multi-turn: re-render the whole conversation (prior
+            // analysis dropped) and re-prefill into a freshly cleared slot.
+            let prompt = build_conversation_gpt_oss(
+                opts.system_prompt.as_deref(),
+                &moe_history,
+                input,
+            );
+            let prompt_tokens = tokenizer.encode(&prompt);
+            if prompt_tokens.is_empty() {
+                continue;
+            }
+            if prompt_tokens.len() >= max_seq_len {
+                eprintln!(
+                    "context full: conversation needs {} tokens >= max_seq_len {max_seq_len}; use /clear",
+                    prompt_tokens.len()
+                );
+                continue;
+            }
+            let max_new_tokens = opts.max_tokens.min(max_seq_len - prompt_tokens.len());
+            reset_slot(&mut cache, &tp_handle);
+            print!("assistant> ");
+            io::stdout().flush().unwrap();
+            let (_finish, answer) = generate_with_paged_cache(
+                &model, &mut cache, &tokenizer, &prompt_tokens, &opts.sampling,
+                max_new_tokens, use_color, &tp_handle, is_moe, opts.enable_thinking,
+            );
+            moe_history.push((input.to_string(), answer));
+            println!();
+            continue;
+        }
+
+        // Qwen3: incremental KV cache — only the new turn is prefilled and the
+        // assistant's tokens stay cached for the next turn.
+        let include_system = cache.seq_len(SLOT) == 0;
+        let prompt = build_turn_prompt(
+            opts.system_prompt.as_deref(),
+            include_system,
+            input,
+            opts.enable_thinking,
+        );
+        let prompt_tokens = tokenizer.encode(&prompt);
+        if prompt_tokens.is_empty() {
+            continue;
+        }
+
+        let used = cache.seq_len(SLOT);
+        let remaining = max_seq_len.saturating_sub(used);
+        if prompt_tokens.len() >= remaining {
+            eprintln!(
+                "context full: {used}/{max_seq_len} tokens used, new turn needs {} tokens; use /clear",
+                prompt_tokens.len()
+            );
+            continue;
+        }
+        let max_new_tokens = opts.max_tokens.min(remaining - prompt_tokens.len());
+
+        print!("assistant> ");
+        io::stdout().flush().unwrap();
+        let (finish, _answer) = generate_with_paged_cache(
+            &model,
+            &mut cache,
+            &tokenizer,
+            &prompt_tokens,
+            &opts.sampling,
+            max_new_tokens,
+            use_color,
+            &tp_handle,
+            is_moe,
+            opts.enable_thinking,
+        );
+        match finish {
+            Finish::Stop { token_id } => {
+                append_after_stop(&model, &mut cache, &tokenizer, max_seq_len, token_id, &tp_handle);
+            }
+            Finish::Length => {
+                append_text_to_cache(&model, &mut cache, &tokenizer, max_seq_len, "<|im_end|>\n", &tp_handle);
+            }
+        }
+        println!();
+    }
+}
+
+/// Free and re-register the single chat KV slot (clears all cached context).
+fn reset_slot(cache: &mut PagedKVCache, tp: &Option<TpHandle>) {
+    if let Some(h) = tp { h.send(TpCommand::Free(SLOT)); h.wait(); }
+    cache.free_sequence(SLOT);
+    if let Some(h) = tp { h.send(TpCommand::Register(SLOT)); h.wait(); }
+    cache.register_sequence(SLOT).expect("register chat slot");
+}
+
+fn parse_args() -> CliOptions {
+    let args: Vec<String> = std::env::args().skip(1).collect();
+    if args.is_empty() || args.iter().any(|a| a == "--help" || a == "-h") {
+        print_usage_and_exit(0);
+    }
+
+    let mut model_dir = None;
+    let mut max_tokens = 256usize;
+    let mut max_seq_len = 2048usize;
+    let mut tp = 1usize;
+    let mut temperature = 0.0f32;
+    let mut top_k = 0usize;
+    let mut top_p = 1.0f32;
+    let mut system_prompt = None;
+    let mut enable_thinking = false;
+    let mut color = true;
+
+    let mut i = 0;
+    while i < args.len() {
+        match args[i].as_str() {
+            "-m" | "--model" => {
+                i += 1;
+                model_dir = args.get(i).map(PathBuf::from);
+            }
+            "--max-tokens" => {
+                i += 1;
+                max_tokens = parse_value(&args, i, "--max-tokens");
+            }
+            "--max-seq-len" => {
+                i += 1;
+                max_seq_len = parse_value(&args, i, "--max-seq-len");
+            }
+            "--tp" => {
+                i += 1;
+                tp = parse_value(&args, i, "--tp");
+            }
+            "--temperature" => {
+                i += 1;
+                temperature = parse_value(&args, i, "--temperature");
+            }
+            "--top-k" => {
+                i += 1;
+                top_k = parse_value(&args, i, "--top-k");
+            }
+            "--top-p" => {
+                i += 1;
+                top_p = parse_value(&args, i, "--top-p");
+            }
+            "--system" => {
+                i += 1;
+                system_prompt = args.get(i).cloned();
+                if system_prompt.is_none() {
+                    eprintln!("missing value for --system");
+                    std::process::exit(2);
+                }
+            }
+            "--think" => {
+                enable_thinking = true;
+            }
+            "--no-color" => {
+                color = false;
+            }
+            arg if arg.starts_with('-') => {
+                eprintln!("unknown option: {arg}");
+                print_usage_and_exit(2);
+            }
+            arg => {
+                if model_dir.is_some() {
+                    eprintln!("unexpected extra argument: {arg}");
+                    print_usage_and_exit(2);
+                }
+                model_dir = Some(PathBuf::from(arg));
+            }
+        }
+        i += 1;
+    }
+
+    CliOptions {
+        model_dir: model_dir.unwrap_or_else(|| {
+            eprintln!("missing model directory");
+            print_usage_and_exit(2);
+        }),
+        max_tokens: max_tokens.max(1),
+        max_seq_len: max_seq_len.max(1),
+        tp: tp.max(1),
+        sampling: SamplingParams {
+            temperature,
+            top_k,
+            top_p,
+        },
+        system_prompt,
+        enable_thinking,
+        color,
+    }
+}
+
+fn parse_value<T: std::str::FromStr>(args: &[String], i: usize, name: &str) -> T {
+    args.get(i).and_then(|s| s.parse().ok()).unwrap_or_else(|| {
+        eprintln!("invalid or missing value for {name}");
+        std::process::exit(2);
+    })
+}
+
+fn print_usage_and_exit(code: i32) -> ! {
+    eprintln!(
+        "Usage: xserv-chat <model-dir> [options]\n\
+         \n\
+         Options:\n\
+         \t-m, --model DIR       Model directory\n\
+         \t--max-tokens N        Max generated tokens per turn (default: 256)\n\
+         \t--max-seq-len N       Persistent KV context length (default: 2048)\n\
+         \t--tp N                Tensor parallelism degree (default: 1)\n\
+         \t--temperature F       Sampling temperature, 0 = greedy (default: 0)\n\
+         \t--top-k N             Top-k sampling, 0 = disabled (default: 0)\n\
+         \t--top-p F             Top-p sampling (default: 1.0)\n\
+         \t--system TEXT         System prompt for the first turn after start or /clear\n\
+         \t--think               Let Qwen3 emit thinking; rendered gray on terminals\n\
+         \t--no-color            Disable ANSI color for thinking output\n\
+         \t-h, --help            Show this help"
+    );
+    std::process::exit(code);
+}
+
+fn print_help() {
+    eprintln!("Commands:");
+    eprintln!("  /clear  clear chat history and free/recreate the paged KV slot");
+    eprintln!("  /exit   quit");
+    eprintln!("  /quit   quit");
+}
+
+fn new_paged_cache(config: &ModelConfig, max_seq_len: usize) -> PagedKVCache {
+    let max_blocks_per_seq = (max_seq_len + BLOCK_SIZE - 1) / BLOCK_SIZE;
+    let total_blocks = (max_blocks_per_seq + 1).max(2);
+    // Single-slot interactive CLI: no swap pool (cpu_total_blocks = 0).
+    PagedKVCache::new(config, total_blocks, 0, 1, max_blocks_per_seq, DType::BF16, 0)
+}
+
+fn build_turn_prompt(
+    system: Option<&str>,
+    include_system: bool,
+    user_input: &str,
+    enable_thinking: bool,
+) -> String {
+    let mut prompt = String::new();
+    if include_system {
+        if let Some(system) = system {
+            if !system.trim().is_empty() {
+                prompt.push_str("<|im_start|>system\n");
+                prompt.push_str(system.trim());
+                prompt.push_str("<|im_end|>\n");
+            }
+        }
+    }
+    prompt.push_str("<|im_start|>user\n");
+    prompt.push_str(user_input);
+    prompt.push_str("<|im_end|>\n");
+    prompt.push_str("<|im_start|>assistant\n");
+    if !enable_thinking {
+        prompt.push_str("<think>\n\n</think>\n\n");
+    }
+    prompt
+}
+
+/// Render the full gpt-oss harmony conversation for re-prefill. gpt-oss was
+/// trained on this exact system-message structure (identity / knowledge cutoff
+/// / current date / Reasoning level / channels — see the model's
+/// chat_template.jinja `build_system_message`). A hand-rolled substitute puts
+/// the model out of distribution and destabilizes channel selection.
+///
+/// Harmony multi-turn drops prior chain-of-thought: past assistant messages are
+/// rendered as completed `final` channels ending in `<|end|>` (not the
+/// `<|return|>` stop token). Keeping the analysis + `<|return|>` of every turn
+/// in context — as an incremental KV cache does — is out of distribution and
+/// makes the model collapse at longer context. "Reasoning: low" keeps the
+/// analysis channel short for an interactive chat.
+fn build_conversation_gpt_oss(
+    system: Option<&str>,
+    history: &[(String, String)],
+    current_user: &str,
+) -> String {
+    let mut prompt = String::new();
+    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", today_ymd()));
+    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|>");
+    if let Some(sys) = system {
+        if !sys.trim().is_empty() {
+            prompt.push_str("<|start|>developer<|message|># Instructions\n\n");
+            prompt.push_str(sys.trim());
+            prompt.push_str("<|end|>");
+        }
+    }
+    for (user, assistant) in history {
+        prompt.push_str("<|start|>user<|message|>");
+        prompt.push_str(user);
+        prompt.push_str("<|end|>");
+        prompt.push_str("<|start|>assistant<|channel|>final<|message|>");
+        prompt.push_str(assistant.trim());
+        prompt.push_str("<|end|>");
+    }
+    prompt.push_str("<|start|>user<|message|>");
+    prompt.push_str(current_user);
+    prompt.push_str("<|end|>");
+    prompt.push_str("<|start|>assistant");
+    prompt
+}
+
+/// Current UTC date as "YYYY-MM-DD" for the harmony system message. Rata Die
+/// civil-calendar conversion (same algorithm the server uses for strftime_now).
+fn today_ymd() -> String {
+    use std::time::{SystemTime, UNIX_EPOCH};
+    let secs = SystemTime::now().duration_since(UNIX_EPOCH).unwrap().as_secs();
+    let z = (secs / 86400) as i64 + 719468;
+    let era = (if z >= 0 { z } else { z - 146096 }) / 146097;
+    let doe = z - era * 146097;
+    let yoe = (doe - doe / 1460 + doe / 36524 - doe / 146096) / 365;
+    let y = yoe + 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 };
+    format!("{y:04}-{m:02}-{d:02}")
+}
+
+fn generate_with_paged_cache(
+    model: &ChatModel,
+    cache: &mut PagedKVCache,
+    tokenizer: &Tokenizer,
+    prompt_tokens: &[u32],
+    sampling: &SamplingParams,
+    max_tokens: usize,
+    use_color: bool,
+    tp: &Option<TpHandle>,
+    is_moe: bool,
+    enable_thinking: bool,
+) -> (Finish, String) {
+    let harmony_end_id = if is_moe { tokenizer.special_token_id("<|end|>") } else { None };
+    let harmony_channel_id = if is_moe { tokenizer.special_token_id("<|channel|>") } else { None };
+    let harmony_message_id = if is_moe { tokenizer.special_token_id("<|message|>") } else { None };
+    let harmony_special: Vec<u32> = if is_moe {
+        ["<|channel|>", "<|start|>", "<|end|>", "<|message|>", "<|return|>"]
+            .iter().filter_map(|s| tokenizer.special_token_id(s)).collect()
+    } else {
+        Vec::new()
+    };
+    // Harmony channel state: "final" channel text is printed normally,
+    // "analysis" channel is rendered as thinking (gray). After <|channel|>
+    // we read the channel name tokens until <|message|>.
+    #[derive(PartialEq, Clone, Copy)]
+    enum HarmonyState { Normal, ReadingChannel, InAnalysis, InFinal }
+    let mut hstate = if is_moe { HarmonyState::InFinal } else { HarmonyState::Normal };
+
+    // Off by default. A repetition penalty over a harmony stream penalizes the
+    // control tokens (<|channel|>, <|message|>, <|start|>) that MUST repeat to
+    // open the final channel — so a non-1.0 default makes gpt-oss stop right
+    // after the analysis block, before emitting any answer. Opt in via the env
+    // var if you want it for plain (non-harmony) generation.
+    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(512);
+    let mut history: Vec<u32> = Vec::new();
+
+    let pick = |logits: &xserv_tensor::Tensor, sp: &SamplingParams, hist: &[u32]| -> u32 {
+        if rep_penalty > 1.0 && sp.temperature == 0.0 {
+            let start = hist.len().saturating_sub(rep_window);
+            sample_greedy_penalized(logits, &hist[start..], rep_penalty)
+        } else {
+            sample(logits, sp)
+        }
+    };
+
+    if let Some(h) = tp { h.send(TpCommand::Prefill { tokens: prompt_tokens.to_vec(), slot: SLOT }); }
+    let logits = model.forward_prefill_paged(prompt_tokens, SLOT, cache);
+    if let Some(h) = tp { h.wait(); }
+    let mut next = pick(&logits, sampling, &history);
+    let mut decode_buffer = Vec::new();
+    let mut in_thinking = false;
+    let show_thinking = is_moe && enable_thinking;
+    // Visible answer tokens, returned for multi-turn history. For moe this is
+    // the final-channel content only (analysis is suppressed/gray); for Qwen3
+    // it is everything printed. The caller decodes these into the assistant
+    // message it re-renders into the next prompt.
+    let mut answer_ids: Vec<u32> = Vec::new();
+
+    for _ in 0..max_tokens {
+        let position = cache.seq_len(SLOT);
+        if let Some(h) = tp { h.send(TpCommand::Decode { tokens: vec![next], positions: vec![position], slots: vec![SLOT] }); }
+        let logits = model.forward_decode_paged(&[next], &[position], &[SLOT], cache);
+        if let Some(h) = tp { h.wait(); }
+        if tokenizer.is_eos(next) {
+            print_stream_text(
+                &tokenizer.flush_decode_stream(&mut decode_buffer),
+                in_thinking,
+                use_color,
+            );
+            if show_thinking && in_thinking {
+                print_stream_text("\n</think>\n\n", true, use_color);
+            }
+            io::stdout().flush().unwrap();
+            return (Finish::Stop { token_id: next }, tokenizer.decode(&answer_ids));
+        }
+        if harmony_end_id == Some(next) {
+            // <|end|> closes current segment; if in final channel, we're done
+            print_stream_text(
+                &tokenizer.flush_decode_stream(&mut decode_buffer),
+                in_thinking,
+                use_color,
+            );
+            if hstate == HarmonyState::InFinal {
+                io::stdout().flush().unwrap();
+                return (Finish::Stop { token_id: next }, tokenizer.decode(&answer_ids));
+            }
+            // Closing a thinking (analysis/commentary) channel: emit the </think>
+            // marker so it renders like Qwen3's thinking block.
+            if show_thinking && hstate == HarmonyState::InAnalysis {
+                print_stream_text("\n</think>\n\n", true, use_color);
+                in_thinking = false;
+            }
+            hstate = HarmonyState::Normal;
+            next = pick(&logits, sampling, &history);
+            continue;
+        }
+
+        history.push(next);
+
+        // Harmony channel routing state machine
+        if harmony_channel_id == Some(next) {
+            decode_buffer.clear();
+            hstate = HarmonyState::ReadingChannel;
+            next = pick(&logits, sampling, &history);
+            continue;
+        }
+        if harmony_message_id == Some(next) {
+            if hstate == HarmonyState::ReadingChannel {
+                // Channel name was accumulated but we don't need to parse it —
+                // we just check via the channel_name buffer below
+            }
+            decode_buffer.clear();
+            next = pick(&logits, sampling, &history);
+            continue;
+        }
+        if hstate == HarmonyState::ReadingChannel {
+            // Reading channel name tokens (e.g. "final", "analysis")
+            let tok_text = tokenizer.decode(&[next]);
+            if tok_text.contains("final") {
+                hstate = HarmonyState::InFinal;
+                in_thinking = false;
+            } else {
+                hstate = HarmonyState::InAnalysis;
+                // Open a Qwen3-style thinking block for the analysis channel.
+                if show_thinking {
+                    print_stream_text("<think>\n", true, use_color);
+                    in_thinking = true;
+                }
+            }
+            next = pick(&logits, sampling, &history);
+            continue;
+        }
+        if harmony_special.contains(&next) {
+            next = pick(&logits, sampling, &history);
+            continue;
+        }
+        if hstate == HarmonyState::InAnalysis {
+            // Analysis channel = the model's reasoning. With --think, show it as a
+            // thinking block (gray if color); otherwise suppress it (answer only).
+            if show_thinking {
+                print_generated_token(tokenizer, next, &mut decode_buffer, &mut in_thinking, use_color);
+                io::stdout().flush().unwrap();
+            }
+            next = pick(&logits, sampling, &history);
+            continue;
+        }
+        if is_moe && hstate != HarmonyState::InFinal {
+            // Between harmony messages (after a channel's <|end|>, before the
+            // next <|channel|>): the model emits a role header like "assistant".
+            // That's structural, not user-visible content — suppress it. Only
+            // for moe/harmony; non-moe (Qwen3) stays in Normal and prints here.
+            next = pick(&logits, sampling, &history);
+            continue;
+        }
+
+        answer_ids.push(next);
+        print_generated_token(
+            tokenizer,
+            next,
+            &mut decode_buffer,
+            &mut in_thinking,
+            use_color,
+        );
+        io::stdout().flush().unwrap();
+        next = pick(&logits, sampling, &history);
+    }
+
+    print_stream_text(
+        &tokenizer.flush_decode_stream(&mut decode_buffer),
+        in_thinking,
+        use_color,
+    );
+    if show_thinking && in_thinking {
+        print_stream_text("\n</think>\n\n", true, use_color);
+    }
+    io::stdout().flush().unwrap();
+    (Finish::Length, tokenizer.decode(&answer_ids))
+}
+
+fn append_after_stop(
+    model: &ChatModel,
+    cache: &mut PagedKVCache,
+    tokenizer: &Tokenizer,
+    max_seq_len: usize,
+    _stop_token_id: u32,
+    tp: &Option<TpHandle>,
+) {
+    append_text_to_cache(model, cache, tokenizer, max_seq_len, "\n", tp);
+}
+
+fn append_text_to_cache(
+    model: &ChatModel,
+    cache: &mut PagedKVCache,
+    tokenizer: &Tokenizer,
+    max_seq_len: usize,
+    text: &str,
+    tp: &Option<TpHandle>,
+) {
+    let tokens = tokenizer.encode(text);
+    if tokens.is_empty() || cache.seq_len(SLOT) + tokens.len() > max_seq_len {
+        return;
+    }
+    if let Some(h) = tp { h.send(TpCommand::Prefill { tokens: tokens.clone(), slot: SLOT }); }
+    let _ = model.forward_prefill_paged(&tokens, SLOT, cache);
+    if let Some(h) = tp { h.wait(); }
+}
+
+fn print_generated_token(
+    tokenizer: &Tokenizer,
+    token_id: u32,
+    decode_buffer: &mut Vec<u8>,
+    in_thinking: &mut bool,
+    use_color: bool,
+) {
+    if tokenizer.special_token_id("<think>") == Some(token_id) {
+        print_stream_text(
+            &tokenizer.flush_decode_stream(decode_buffer),
+            *in_thinking,
+            use_color,
+        );
+        *in_thinking = true;
+        print_stream_text("<think>", true, use_color);
+        return;
+    }
+
+    if tokenizer.special_token_id("</think>") == Some(token_id) {
+        print_stream_text(
+            &tokenizer.flush_decode_stream(decode_buffer),
+            *in_thinking,
+            use_color,
+        );
+        print_stream_text("</think>", true, use_color);
+        *in_thinking = false;
+        return;
+    }
+
+    let text = tokenizer.decode_token_stream(token_id, decode_buffer);
+    print_stream_text(&text, *in_thinking, use_color);
+}
+
+fn print_stream_text(text: &str, in_thinking: bool, use_color: bool) {
+    if text.is_empty() {
+        return;
+    }
+    if in_thinking && use_color {
+        print!("\x1b[90m{text}\x1b[0m");
+    } else {
+        print!("{text}");
+    }
+}
+

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

@@ -1,6 +1,6 @@
 use std::io::{self, Write};
 use std::path::PathBuf;
-use xserv_model::{loader, KVCache, ModelConfig};
+use xserv_model::{loader, KVCache, ModelConfig, PagedKVCache, BLOCK_SIZE};
 use xserv_tensor::{DType, Device};
 use xserv_tokenizer::Tokenizer;
 
@@ -36,14 +36,18 @@ fn main() {
     eprintln!("Loaded {} tensors", weights.len());
 
     let is_qwen3 = model_type.contains("qwen");
-    let dtype = if is_qwen3 { DType::BF16 } else { DType::F32 };
+    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_qwen3 {
+    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))
@@ -62,40 +66,92 @@ fn main() {
         if input == "quit" || input == "exit" { break; }
 
         let token_ids = tokenizer.encode(input);
-        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),
-        );
 
-        // Prefill + decode
+        if is_gpt_oss {
-        let logits = match &model {
+            // GptOss uses paged KV cache
-            Model::GPT2(m) => m.forward_with_cache(&token_ids, &mut cache),
+            let max_seq = 2048;
-            Model::Qwen3(m) => m.forward_with_cache(&token_ids, &mut cache),
+            let max_blocks_per_seq = (max_seq + BLOCK_SIZE - 1) / BLOCK_SIZE;
-        };
+            let total_blocks = max_blocks_per_seq + 64;
-        let mut next = match &model {
+            let mut paged_cache = PagedKVCache::new(
-            Model::GPT2(_) => xserv_model::gpt2::sample_greedy(&logits),
+                &config, total_blocks, 0, 4, max_blocks_per_seq, DType::BF16, 0,
-            Model::Qwen3(_) => xserv_model::qwen3::sample_greedy(&logits),
+            );
-        };
+            let slot = 0;
+            paged_cache.register_sequence(slot).expect("register slot");
 
-        print!("{input}");
+            let model = match &model { Model::GptOss(m) => m, _ => unreachable!() };
-        io::stdout().flush().unwrap();
+            let logits = model.forward_prefill_paged(&token_ids, slot, &mut paged_cache);
+            let mut next = sample_greedy_last(&logits);
 
-        for _ in 0..max_tokens {
+            print!("{input}");
-            let text = tokenizer.decode(&[next]);
-            print!("{text}");
             io::stdout().flush().unwrap();
 
-            if tokenizer.eos_token_id() == Some(next) { break; }
+            for _ in 0..max_tokens {
+                let text = tokenizer.decode(&[next]);
+                print!("{text}");
+                io::stdout().flush().unwrap();
+
+                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,
+                );
+                next = sample_greedy_last(&logits);
+            }
+            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(&[next], &mut cache),
+                Model::GPT2(m) => m.forward_with_cache(&token_ids, &mut cache),
-                Model::Qwen3(m) => m.forward_with_cache(&[next], &mut cache),
+                Model::Qwen3(m) => m.forward_with_cache(&token_ids, &mut cache),
+                Model::GptOss(_) => unreachable!(),
             };
-            next = match &model {
+            let mut next = match &model {
                 Model::GPT2(_) => xserv_model::gpt2::sample_greedy(&logits),
                 Model::Qwen3(_) => xserv_model::qwen3::sample_greedy(&logits),
+                Model::GptOss(_) => unreachable!(),
             };
+
+            print!("{input}");
+            io::stdout().flush().unwrap();
+
+            for _ in 0..max_tokens {
+                let text = tokenizer.decode(&[next]);
+                print!("{text}");
+                io::stdout().flush().unwrap();
+
+                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!(),
+                };
+                next = match &model {
+                    Model::GPT2(_) => xserv_model::gpt2::sample_greedy(&logits),
+                    Model::Qwen3(_) => xserv_model::qwen3::sample_greedy(&logits),
+                    Model::GptOss(_) => unreachable!(),
+                };
+            }
+            println!();
         }
-        println!();
     }
 }
+
+fn sample_greedy_last(logits: &xserv_tensor::Tensor) -> u32 {
+    use half::bf16;
+    assert_eq!(logits.ndim(), 2);
+    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| a.1.to_f32().partial_cmp(&b.1.to_f32()).unwrap())
+        .map(|(i, _)| i as u32).unwrap()
+}

crates/xserv-model/src/config.rs

@@ -1,6 +1,15 @@
 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>>,
@@ -46,6 +55,28 @@ pub struct ModelConfig {
     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 {
@@ -81,7 +112,7 @@ impl ModelConfig {
     }
 
     pub fn head_dim(&self) -> usize {
-        self.hidden() / self.num_heads()
+        self.explicit_head_dim.unwrap_or_else(|| self.hidden() / self.num_heads())
     }
 
     pub fn ln_eps(&self) -> f32 {
@@ -93,4 +124,32 @@ impl ModelConfig {
     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 +1,458 @@
+//! 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;
+
+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 accumulators (separate per output dimension)
+    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(hidden * 4),
+            fp32_q: alloc(num_heads * head_dim * 4),
+            fp32_kv: alloc(num_kv_heads * head_dim * 4),
+            fp32_intermediate: alloc(intermediate * 4),
+            fp32_vocab: alloc(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/gpt2.rs

@@ -116,6 +116,7 @@ fn tensor_from_raw_bytes(bytes: &[u8], shape: &[usize], dtype: DType) -> Tensor
 
 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}"))
         };
@@ -279,45 +280,88 @@ fn add_bias(x: &Tensor, bias: &Tensor) -> Tensor {
 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 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 device = qkv.device();
-    let q = Tensor::from_slice(&q_data, &[1, num_heads, seq_len, head_dim]).to_device(device);
+    let dtype = qkv.dtype();
-    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);
+    match dtype {
-    (q, k, v)
+        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 src = x_cpu.as_slice::<f32>();
+    let device = x.device();
+    let dtype = x.dtype();
 
-    let mut out = vec![0.0f32; seq_len * hidden];
+    match dtype {
-    for s in 0..seq_len {
+        DType::F32 => {
-        for h in 0..num_heads {
+            let src = x_cpu.as_slice::<f32>();
-            let src_off = (h * seq_len + s) * head_dim;
+            let mut out = vec![0.0f32; seq_len * hidden];
-            let dst_off = s * hidden + h * head_dim;
+            for s in 0..seq_len {
-            out[dst_off..dst_off + head_dim].copy_from_slice(&src[src_off..src_off + head_dim]);
+                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),
     }
-    Tensor::from_slice(&out, &[seq_len, hidden]).to_device(x.device())
 }
 
 /// Greedy sampling: return the argmax token ID from the last position's logits.

crates/xserv-model/src/gpt_oss.rs

@@ -0,0 +1,769 @@
+use std::collections::HashMap;
+use std::ffi::c_void;
+use half::bf16;
+use xserv_kernels::*;
+use xserv_tensor::{Device, Tensor};
+
+use crate::config::ModelConfig;
+use crate::paged_kv_cache::PagedKVCache;
+
+pub struct GptOss {
+    pub config: ModelConfig,
+    embed_tokens: Tensor,
+    layers: Vec<GptOssBlock>,
+    norm: Tensor,
+    norm_bias: Option<Tensor>,
+    lm_head_t: Tensor,
+    rope_cache: RopeCache,
+    tp: Option<std::sync::Arc<xserv_distributed::TpContext>>,
+    local_num_heads: usize,
+    local_num_kv_heads: usize,
+    has_norm_bias: bool,
+}
+
+struct GptOssBlock {
+    input_norm: Tensor,
+    input_norm_bias: Option<Tensor>,
+    // Attention (with bias)
+    q_proj_wt: Tensor,
+    q_proj_bias: Tensor,
+    k_proj_wt: Tensor,
+    k_proj_bias: Tensor,
+    v_proj_wt: Tensor,
+    v_proj_bias: Tensor,
+    o_proj_wt: Tensor,
+    o_proj_bias: Tensor,
+    sinks: Tensor,
+    #[allow(dead_code)]
+    is_sliding: bool,
+    window_size: usize,
+    // MoE MLP
+    post_norm: Tensor,
+    post_norm_bias: Option<Tensor>,
+    router_wt: Tensor,
+    router_bias: Tensor,
+    // 3D expert weights for batched GEMM (contiguous on GPU)
+    expert_gate_up_wt: Tensor,   // [local_experts, hidden, 2*inter] BF16
+    expert_gate_up_bias: Tensor, // [local_experts, 2*inter]
+    expert_down_wt: Tensor,      // [local_experts, inter, hidden] BF16
+    expert_down_bias: Tensor,    // [local_experts, hidden]
+    // FP8 quantized expert weights (Some when running FP8 W8A8)
+    // Transposed layout [E, N, K] for cuBLASLt FP8 (Blackwell requires transA=T)
+    expert_gate_up_fp8: Option<Tensor>,  // [local_experts, 2*inter, hidden] FP8E4M3
+    expert_gate_up_scale: Option<Tensor>,// [local_experts] F32
+    expert_down_fp8: Option<Tensor>,     // [local_experts, hidden, inter] FP8E4M3
+    expert_down_scale: Option<Tensor>,   // [local_experts] F32
+    // MXFP4 W4A16 expert weights (Some when running 4-bit weight-only).
+    // (packed [E, N, K/2] u8, scales [E, N, K/32] u8) in [E, N, K] layout.
+    expert_gate_up_mxfp4: Option<(Tensor, Tensor)>,
+    expert_down_mxfp4: Option<(Tensor, Tensor)>,
+    local_experts: usize,
+    // Activation params
+    glu_alpha: f32,
+    glu_limit: f32,
+}
+
+impl GptOss {
+    pub fn from_weights(config: ModelConfig, w: HashMap<String, Tensor>) -> Self {
+        Self::from_weights_tp(config, w, 0, 1, 0, None)
+    }
+
+    pub fn from_weights_tp(
+        config: ModelConfig,
+        mut w: HashMap<String, Tensor>,
+        rank: usize,
+        world: usize,
+        device: u32,
+        tp: Option<std::sync::Arc<xserv_distributed::TpContext>>,
+    ) -> Self {
+        crate::init_kernels();
+        let dev = Device::Cuda(device);
+        let take = |w: &mut HashMap<String, Tensor>, name: &str| -> Tensor {
+            w.remove(name).unwrap_or_else(|| panic!("missing weight: {name}"))
+        };
+        let repl = |t: Tensor| -> Tensor { t.to_device(dev) };
+        // column-parallel: shard rows of [out, in], transpose → [in, out/world]
+        let col = |t: Tensor| -> Tensor {
+            shard_rows(&t, rank, world).to_device(dev).transpose(0, 1).contiguous()
+        };
+        // row-parallel: shard cols of [out, in], transpose → [in/world, out]
+        let row = |t: Tensor| -> Tensor {
+            shard_cols(&t, rank, world).to_device(dev).transpose(0, 1).contiguous()
+        };
+        // Bias sharding helpers
+        let col_bias = |t: Tensor| -> Tensor { shard_1d(&t, rank, world).to_device(dev) };
+        let repl_bias = |t: Tensor| -> Tensor { t.to_device(dev) };
+
+        let embed_tokens = repl(take(&mut w, "model.embed_tokens.weight"));
+        let norm = repl(take(&mut w, "model.norm.weight"));
+        let norm_bias = w.remove("model.norm.bias").map(|t| repl(t));
+        let lm_head_t = repl(take(&mut w, "lm_head.weight")).transpose(0, 1).contiguous();
+
+        let head_dim = config.head_dim();
+        let rope_theta = config.rope_theta.unwrap_or(150000.0);
+        let max_seq_len = config.max_seq_len().min(8192); // cap for memory
+
+        let rope_cache = if let Some(ref rs) = config.rope_scaling {
+            if rs.rope_type.as_deref() == Some("yarn") {
+                RopeCache::new_yarn(
+                    max_seq_len,
+                    head_dim,
+                    rope_theta,
+                    rs.factor.unwrap_or(1.0),
+                    rs.original_max_position_embeddings.unwrap_or(4096),
+                    rs.beta_fast.unwrap_or(32.0),
+                    rs.beta_slow.unwrap_or(1.0),
+                )
+            } else {
+                RopeCache::new(max_seq_len, head_dim, rope_theta as f32)
+            }
+        } else {
+            RopeCache::new(max_seq_len, head_dim, rope_theta as f32)
+        };
+
+        let num_layers = config.num_layers();
+        let num_experts = config.num_experts();
+        let glu_alpha = config.geglu_alpha();
+        let glu_limit = config.swiglu_limit.unwrap_or(7.0) as f32;
+
+        let mut layers = Vec::with_capacity(num_layers);
+        if rank == 0 {
+            eprintln!(
+                "Loading gpt-oss weights: {} layers, {} experts, world={world}, glu_alpha={glu_alpha}...",
+                num_layers, num_experts
+            );
+        }
+
+        for i in 0..num_layers {
+            let p = format!("model.layers.{i}");
+
+            // Attention weights — column-parallel for Q/K/V, row-parallel for O
+            let q_proj_wt = col(take(&mut w, &format!("{p}.self_attn.q_proj.weight")));
+            let q_proj_bias = col_bias(take(&mut w, &format!("{p}.self_attn.q_proj.bias")));
+            let k_proj_wt = col(take(&mut w, &format!("{p}.self_attn.k_proj.weight")));
+            let k_proj_bias = col_bias(take(&mut w, &format!("{p}.self_attn.k_proj.bias")));
+            let v_proj_wt = col(take(&mut w, &format!("{p}.self_attn.v_proj.weight")));
+            let v_proj_bias = col_bias(take(&mut w, &format!("{p}.self_attn.v_proj.bias")));
+            let o_proj_wt = row(take(&mut w, &format!("{p}.self_attn.o_proj.weight")));
+            let o_proj_bias = repl_bias(take(&mut w, &format!("{p}.self_attn.o_proj.bias")));
+
+            // Sinks: shard per-head across TP ranks
+            let sinks_full = take(&mut w, &format!("{p}.self_attn.sinks"));
+            let sinks = shard_1d(&sinks_full, rank, world).to_device(dev);
+
+            let is_sliding = config.is_sliding_layer(i);
+            let window_size = if is_sliding { config.window_size() } else { 0 };
+
+            // MoE weights — router replicated, experts split across TP ranks
+            let router_wt_raw = take(&mut w, &format!("{p}.mlp.router.weight"));
+            let router_wt = router_wt_raw.to_device(dev).transpose(0, 1).contiguous();
+            let router_bias = repl_bias(take(&mut w, &format!("{p}.mlp.router.bias")));
+
+            // Expert weights: [num_experts, hidden, 2*inter] — stored as 3D tensors
+            // Expert parallelism: rank owns experts [rank*E/world .. (rank+1)*E/world)
+            let gate_up_3d = take(&mut w, &format!("{p}.mlp.experts.gate_up_proj"));
+            let gate_up_bias_2d = take(&mut w, &format!("{p}.mlp.experts.gate_up_proj_bias"));
+            let down_3d = take(&mut w, &format!("{p}.mlp.experts.down_proj"));
+            let down_bias_2d = take(&mut w, &format!("{p}.mlp.experts.down_proj_bias"));
+
+            // FP8 scale tensors (present only in FP8-quantized models)
+            let gate_up_scale = w.remove(&format!("{p}.mlp.experts.gate_up_proj_scale"));
+            let down_scale = w.remove(&format!("{p}.mlp.experts.down_proj_scale"));
+
+            let local_experts = num_experts / world;
+            let expert_start = rank * local_experts;
+
+            // MXFP4 stores 4-bit weights in an FP8E4M3 byte container (same dtype
+            // as FP8), so distinguish by the scale rank: FP8 scale is 1-D [E],
+            // MXFP4 scale is 3-D [E, N, K/32].
+            let is_mxfp4 = gate_up_scale.as_ref().map(|s| s.ndim() == 3).unwrap_or(false);
+            let is_fp8 = !is_mxfp4 && gate_up_3d.dtype() == xserv_tensor::DType::FP8E4M3;
+
+            let mut expert_gate_up_mxfp4: Option<(Tensor, Tensor)> = None;
+            let mut expert_down_mxfp4: Option<(Tensor, Tensor)> = None;
+
+            let inter2 = if is_mxfp4 { gate_up_3d.shape()[1] } else { gate_up_3d.shape()[2] }; // 2*inter (N)
+            let hidden = if is_mxfp4 { gate_up_3d.shape()[2] * 2 } else { gate_up_3d.shape()[1] };
+            let inter = if is_mxfp4 { down_3d.shape()[2] * 2 } else { down_3d.shape()[1] };
+
+            // Slice the rank's range of experts as contiguous 3D tensors on GPU
+            let expert_gate_up_wt;
+            let expert_down_wt;
+            let expert_gate_up_fp8;
+            let expert_gate_up_scale_gpu;
+            let expert_down_fp8;
+            let expert_down_scale_gpu;
+
+            if is_mxfp4 {
+                // MXFP4 W4A16: weights already [E, N, K] packed ([E, N, K/2] bytes)
+                // + scales [E, N, K/32]. Slice this rank's experts (raw bytes).
+                let gu_s = gate_up_scale.expect("MXFP4 model missing gate_up_proj_scale");
+                let d_s = down_scale.expect("MXFP4 model missing down_proj_scale");
+                let gu_packed = slice_expert_range_3d_raw(&gate_up_3d, expert_start, local_experts, inter2, hidden / 2).to_device(dev);
+                let gu_scl = slice_expert_range_3d_raw(&gu_s, expert_start, local_experts, inter2, hidden / 32).to_device(dev);
+                let dn_packed = slice_expert_range_3d_raw(&down_3d, expert_start, local_experts, hidden, inter / 2).to_device(dev);
+                let dn_scl = slice_expert_range_3d_raw(&d_s, expert_start, local_experts, hidden, inter / 32).to_device(dev);
+                expert_gate_up_mxfp4 = Some((gu_packed, gu_scl));
+                expert_down_mxfp4 = Some((dn_packed, dn_scl));
+                expert_gate_up_fp8 = None;
+                expert_gate_up_scale_gpu = None;
+                expert_down_fp8 = None;
+                expert_down_scale_gpu = None;
+                expert_gate_up_wt = Tensor::empty(&[1, 1, 1], xserv_tensor::DType::BF16, dev);
+                expert_down_wt = Tensor::empty(&[1, 1, 1], xserv_tensor::DType::BF16, dev);
+            } else if is_fp8 {
+                // FP8 W8A8 path: load and TRANSPOSE weights for cuBLASLt (requires transA=T on Blackwell).
+                // Original: [E, K, N] → Transposed: [E, N, K]
+                let gu_sliced = slice_expert_range_3d_raw(&gate_up_3d, expert_start, local_experts, hidden, inter2);
+                let dn_sliced = slice_expert_range_3d_raw(&down_3d, expert_start, local_experts, inter, hidden);
+                expert_gate_up_fp8 = Some(transpose_3d_inner_raw(&gu_sliced, local_experts, hidden, inter2).to_device(dev));
+                expert_down_fp8 = Some(transpose_3d_inner_raw(&dn_sliced, local_experts, inter, hidden).to_device(dev));
+                // Scales: [num_experts] F32 → slice to [local_experts]
+                let gu_s = gate_up_scale.expect("FP8 model missing gate_up_proj_scale");
+                let d_s = down_scale.expect("FP8 model missing down_proj_scale");
+                expert_gate_up_scale_gpu = Some(slice_scale_range(&gu_s, expert_start, local_experts).to_device(dev));
+                expert_down_scale_gpu = Some(slice_scale_range(&d_s, expert_start, local_experts).to_device(dev));
+                // Dummy BF16 tensors (never read in FP8 path)
+                expert_gate_up_wt = Tensor::empty(&[1, 1, 1], xserv_tensor::DType::BF16, dev);
+                expert_down_wt = Tensor::empty(&[1, 1, 1], xserv_tensor::DType::BF16, dev);
+            } else {
+                // BF16 path: existing behavior
+                expert_gate_up_wt = slice_expert_range_3d(&gate_up_3d, expert_start, local_experts, hidden, inter2).to_device(dev);
+                expert_down_wt = slice_expert_range_3d(&down_3d, expert_start, local_experts, inter, hidden).to_device(dev);
+                expert_gate_up_fp8 = None;
+                expert_gate_up_scale_gpu = None;
+                expert_down_fp8 = None;
+                expert_down_scale_gpu = None;
+            }
+            let expert_gate_up_bias = slice_expert_range_2d(&gate_up_bias_2d, expert_start, local_experts, inter2).to_device(dev);
+            let expert_down_bias = slice_expert_range_2d(&down_bias_2d, expert_start, local_experts, hidden).to_device(dev);
+
+            xserv_cuda::allocator::cached_trim();
+
+            let input_norm = repl(take(&mut w, &format!("{p}.input_layernorm.weight")));
+            let input_norm_bias = w.remove(&format!("{p}.input_layernorm.bias")).map(|t| repl(t));
+            let post_norm = repl(take(&mut w, &format!("{p}.post_attention_layernorm.weight")));
+            let post_norm_bias = w.remove(&format!("{p}.post_attention_layernorm.bias")).map(|t| repl(t));
+
+            layers.push(GptOssBlock {
+                input_norm,
+                input_norm_bias,
+                q_proj_wt,
+                q_proj_bias,
+                k_proj_wt,
+                k_proj_bias,
+                v_proj_wt,
+                v_proj_bias,
+                o_proj_wt,
+                o_proj_bias,
+                sinks,
+                is_sliding,
+                window_size,
+                post_norm,
+                post_norm_bias,
+                router_wt,
+                router_bias,
+                expert_gate_up_wt,
+                expert_gate_up_bias,
+                expert_down_wt,
+                expert_down_bias,
+                expert_gate_up_fp8,
+                expert_gate_up_scale: expert_gate_up_scale_gpu,
+                expert_down_fp8,
+                expert_down_scale: expert_down_scale_gpu,
+                expert_gate_up_mxfp4,
+                expert_down_mxfp4,
+                local_experts,
+                glu_alpha,
+                glu_limit,
+            });
+        }
+
+        let local_num_heads = config.num_heads() / world;
+        let local_num_kv_heads = config.num_kv_heads() / world;
+
+        let has_norm_bias = norm_bias.is_some();
+        let is_fp8 = layers.first().map(|l| l.expert_gate_up_fp8.is_some()).unwrap_or(false);
+        let is_mxfp4 = layers.first().map(|l| l.expert_gate_up_mxfp4.is_some()).unwrap_or(false);
+        if rank == 0 {
+            if has_norm_bias {
+                eprintln!("gpt-oss: detected LayerNorm bias — using LayerNorm instead of RMSNorm");
+            }
+            if is_fp8 {
+                eprintln!("gpt-oss: FP8 E4M3 quantized expert weights detected (W8A8 cuBLASLt mode)");
+            }
+            if is_mxfp4 {
+                eprintln!("gpt-oss: MXFP4 quantized expert weights detected (W4A16 fused-GEMV mode)");
+            }
+        }
+
+        // Warn about unused weights that the model didn't consume
+        if rank == 0 && !w.is_empty() {
+            eprintln!("WARNING: {} unused weight(s) in model:", w.len());
+            let mut keys: Vec<_> = w.keys().collect();
+            keys.sort();
+            for k in &keys {
+                eprintln!("  {k}");
+            }
+        }
+
+        Self {
+            config,
+            embed_tokens,
+            layers,
+            norm,
+            norm_bias,
+            lm_head_t,
+            rope_cache,
+            tp,
+            local_num_heads,
+            local_num_kv_heads,
+            has_norm_bias,
+        }
+    }
+
+    #[inline]
+    fn all_reduce(&self, t: &Tensor) {
+        if let Some(tp) = &self.tp {
+            if tp.world > 1 {
+                let ptr = t.storage().gpu_buffer().as_ptr() as *mut c_void;
+                tp.all_reduce_sum_bf16_ptr(ptr, t.numel());
+            }
+        }
+    }
+
+    #[inline]
+    fn norm(x: &Tensor, weight: &Tensor, bias: &Option<Tensor>, eps: f32) -> Tensor {
+        match bias {
+            Some(b) => layernorm(x, weight, b, eps),
+            None => rmsnorm(x, weight, eps),
+        }
+    }
+
+    #[inline]
+    fn add_norm(x: &Tensor, residual: &Tensor, weight: &Tensor, bias: &Option<Tensor>, eps: f32) -> (Tensor, Tensor) {
+        match bias {
+            Some(b) => {
+                let sum = xserv_kernels::add(x, residual);
+                let normed = layernorm(&sum, weight, b, eps);
+                (normed, sum)
+            }
+            None => add_rmsnorm(x, residual, weight, eps),
+        }
+    }
+
+    fn norm_eps(&self) -> f32 {
+        if self.has_norm_bias {
+            self.config.ln_eps()
+        } else {
+            self.config.rms_norm_eps.unwrap_or(1e-5) as f32
+        }
+    }
+
+    /// Paged decode: process one token per sequence using paged KV cache.
+    pub fn forward_decode_paged(
+        &self,
+        tokens: &[u32],
+        positions: &[usize],
+        seq_slots: &[usize],
+        paged_cache: &mut PagedKVCache,
+    ) -> Tensor {
+        let batch = tokens.len();
+        assert_eq!(positions.len(), batch);
+        assert_eq!(seq_slots.len(), batch);
+        assert!(batch > 0);
+
+        let num_heads = self.local_num_heads;
+        let num_kv_heads = self.local_num_kv_heads;
+        let head_dim = self.config.head_dim();
+        let eps = self.norm_eps();
+
+        let kv_lens: Vec<i32> = positions.iter().map(|&p| (p + 1) as i32).collect();
+        for (b, &slot) in seq_slots.iter().enumerate() {
+            paged_cache.ensure_capacity(slot, positions[b] + 1);
+        }
+        paged_cache.sync_active_batch_with_lens(seq_slots, &kv_lens);
+
+        let bt_ptr = paged_cache.block_table_gpu().as_ptr() as *const i32;
+        let cl_ptr = paged_cache.context_lens_gpu().as_ptr() as *const i32;
+        let max_blocks = paged_cache.max_blocks_per_seq();
+
+        let positions_u32: Vec<u32> = positions.iter().map(|&p| p as u32).collect();
+
+        let mut x = embedding(&self.embed_tokens, tokens);
+
+        for (layer_idx, layer) in self.layers.iter().enumerate() {
+            let residual = x.clone();
+            let normed = Self::norm(&x, &layer.input_norm, &layer.input_norm_bias, eps);
+
+            // Q/K/V projections with bias
+            let q_all = add_bias(&matmul_2d(&normed, &layer.q_proj_wt), &layer.q_proj_bias);
+            let k_all = add_bias(&matmul_2d(&normed, &layer.k_proj_wt), &layer.k_proj_bias);
+            let v_all = add_bias(&matmul_2d(&normed, &layer.v_proj_wt), &layer.v_proj_bias);
+
+
+            // Reshape for RoPE: [B, H*D] → [B, H, D]
+            let q_3d = q_all.reshape(&[batch, num_heads, head_dim]);
+            let k_3d = k_all.reshape(&[batch, num_kv_heads, head_dim]);
+
+            // RoPE (no QK-norm for gpt-oss)
+            rope_inplace(&q_3d, &self.rope_cache, &positions_u32);
+            rope_inplace(&k_3d, &self.rope_cache, &positions_u32);
+
+            let v_3d = v_all.reshape(&[batch, num_kv_heads, head_dim]);
+
+            // KV cache scatter
+            paged_cache.append_tokens_batched(layer_idx, &k_3d, &v_3d, batch);
+
+            // Paged attention with sinks + sliding window
+            let q_4d = q_3d.reshape(&[batch, num_heads, 1, head_dim]);
+            let k_pool_ptr = paged_cache.k_pool(layer_idx).as_ptr() as *const c_void;
+            let v_pool_ptr = paged_cache.v_pool(layer_idx).as_ptr() as *const c_void;
+            let sinks_ptr = layer.sinks.data_ptr() as *const c_void;
+
+            let attn_out = paged_decode_attention_sinks(
+                &q_4d, k_pool_ptr, v_pool_ptr, bt_ptr, cl_ptr,
+                sinks_ptr,
+                batch, num_heads, num_kv_heads, head_dim, max_blocks,
+                layer.window_size,
+            );
+
+            let attn_merged = attn_out.reshape(&[batch, num_heads * head_dim]);
+            let attn_proj = matmul_2d(&attn_merged, &layer.o_proj_wt);
+            self.all_reduce(&attn_proj);
+            let attn_proj = add_bias(&attn_proj, &layer.o_proj_bias);
+
+
+            // Residual + post-norm
+            let (normed, x_new) = Self::add_norm(&attn_proj, &residual, &layer.post_norm, &layer.post_norm_bias, eps);
+
+            let residual = x_new;
+            let normed = normed.contiguous();
+
+            // MoE MLP
+            let moe_out = self.moe_forward(&normed, layer, batch);
+            x = xserv_kernels::add(&residual, &moe_out);
+        }
+
+        // Advance KV cache
+        for &slot in seq_slots {
+            paged_cache.advance_seq_len(slot, 1);
+        }
+
+        unsafe { xserv_cuda::ffi::cudaDeviceSynchronize(); }
+        let x = Self::norm(&x, &self.norm, &self.norm_bias, eps);
+        unsafe { xserv_cuda::ffi::cudaDeviceSynchronize(); }
+        let logits = matmul_2d(&x, &self.lm_head_t);
+        unsafe { xserv_cuda::ffi::cudaDeviceSynchronize(); }
+        logits
+    }
+
+    /// Paged prefill: process full prompt tokens.
+    pub fn forward_prefill_paged(
+        &self,
+        token_ids: &[u32],
+        slot: usize,
+        paged_cache: &mut PagedKVCache,
+    ) -> Tensor {
+        let new_tokens = token_ids.len();
+        let pos_offset = paged_cache.seq_len(slot);
+        let num_heads = self.local_num_heads;
+        let num_kv_heads = self.local_num_kv_heads;
+        let head_dim = self.config.head_dim();
+        let eps = self.norm_eps();
+
+        paged_cache.ensure_capacity(slot, pos_offset + new_tokens);
+        paged_cache.advance_seq_len(slot, new_tokens);
+
+        let mut x = embedding(&self.embed_tokens, token_ids);
+        let positions: Vec<u32> = (pos_offset..pos_offset + new_tokens).map(|p| p as u32).collect();
+
+        for (layer_idx, layer) in self.layers.iter().enumerate() {
+            let residual = x.clone();
+            let normed = Self::norm(&x, &layer.input_norm, &layer.input_norm_bias, eps);
+
+            let q = add_bias(&matmul_2d(&normed, &layer.q_proj_wt), &layer.q_proj_bias);
+            let k = add_bias(&matmul_2d(&normed, &layer.k_proj_wt), &layer.k_proj_bias);
+            let v = add_bias(&matmul_2d(&normed, &layer.v_proj_wt), &layer.v_proj_bias);
+
+            let q = reshape_heads_gpu(&q, new_tokens, num_heads, head_dim);
+            let k = reshape_heads_gpu(&k, new_tokens, num_kv_heads, head_dim);
+            let v = reshape_heads_gpu(&v, new_tokens, num_kv_heads, head_dim);
+
+            // RoPE
+            let q = transpose_for_rope_gpu(&q, new_tokens, num_heads, head_dim);
+            let k = transpose_for_rope_gpu(&k, new_tokens, num_kv_heads, head_dim);
+            rope_inplace(&q, &self.rope_cache, &positions);
+            rope_inplace(&k, &self.rope_cache, &positions);
+            let q = transpose_from_rope_gpu(&q, new_tokens, num_heads, head_dim);
+            let k = transpose_from_rope_gpu(&k, new_tokens, num_kv_heads, head_dim);
+
+            // KV cache
+            paged_cache.append_tokens(slot, layer_idx, &k, &v, new_tokens, pos_offset);
+            let (k_full, v_full) = paged_cache.gather_kv_contiguous(slot, layer_idx);
+
+            // Flash attention with gpt-oss sinks + (per-layer) sliding window.
+            let attn_out = flash_attention_sinks(&q, &k_full, &v_full, &layer.sinks, layer.window_size);
+
+            let attn_merged = merge_heads_gpu(&attn_out, new_tokens, num_heads, head_dim);
+            let attn_proj = matmul_2d(&attn_merged, &layer.o_proj_wt);
+            self.all_reduce(&attn_proj);
+            let attn_proj = add_bias(&attn_proj, &layer.o_proj_bias);
+
+            let (normed, x_new) = Self::add_norm(&attn_proj, &residual, &layer.post_norm, &layer.post_norm_bias, eps);
+            let residual = x_new;
+
+            // MoE MLP
+            let moe_out = self.moe_forward(&normed, layer, new_tokens);
+            x = xserv_kernels::add(&residual, &moe_out);
+        }
+
+        let x = Self::norm(&x, &self.norm, &self.norm_bias, eps);
+        let logits = matmul_2d(&x, &self.lm_head_t);
+        unsafe { xserv_cuda::ffi::cudaDeviceSynchronize(); }
+        logits
+    }
+
+    /// MoE forward pass — fully on GPU via batched GEMM.
+    ///
+    /// Each rank owns `local_experts` experts. The input is replicated across all
+    /// local experts, processed via two batched cuBLAS GEMMs (gate_up and down),
+    /// and the selected experts' outputs are weighted-summed on GPU. Non-selected
+    /// experts contribute zero (via the routing weights), so no scatter/gather is
+    /// needed. AllReduce sums partial results across TP ranks.
+    fn moe_forward(&self, x: &Tensor, layer: &GptOssBlock, num_tokens: usize) -> Tensor {
+        let num_experts = self.config.num_experts();
+        let top_k = self.config.experts_per_token();
+        let rank = self.tp.as_ref().map(|tp| tp.rank).unwrap_or(0);
+        let local_experts = layer.local_experts;
+        let expert_start = rank * local_experts;
+
+        // 1. Router: [tokens, hidden] @ [hidden, num_experts] + bias → [tokens, num_experts]
+        let router_logits = add_bias(
+            &matmul_2d(x, &layer.router_wt),
+            &layer.router_bias,
+        );
+
+        // 2. GPU top-k + softmax
+        let (topk_ids, topk_weights) = xserv_kernels::moe::moe_topk_softmax(
+            &router_logits, num_experts, top_k,
+        );
+
+        // 3. Replicate input: [tokens, hidden] → [local_experts, tokens, hidden]
+        let x_rep = xserv_kernels::moe::moe_replicate(x, local_experts);
+
+        // 4. Batched GEMM gate_up: [E, tokens, hidden] @ [E, hidden, 2*inter] → [E, tokens, 2*inter]
+        let gate_up = if let Some((ref packed, ref scales)) = layer.expert_gate_up_mxfp4 {
+            // MXFP4 W4A16: decode (M=1) uses the fused 4-bit dequant GEMV; prefill
+            // dequantizes to BF16 then reuses the batched GEMM.
+            let n = packed.shape()[1];
+            let k = packed.shape()[2] * 2;
+            if num_tokens == 1 {
+                let x2 = x_rep.reshape(&[local_experts, k]);
+                xserv_kernels::quantization::batched_gemv_mxfp4(&x2, packed, scales, n, k)
+                    .reshape(&[local_experts, 1, n])
+            } else {
+                let w_bf16 = xserv_kernels::quantization::dequant_mxfp4_to_bf16_t(packed, scales, local_experts, n, k);
+                xserv_kernels::moe::batched_gemm_strided(&x_rep, &w_bf16)
+            }
+        } else if let Some(ref wt_fp8_t) = layer.expert_gate_up_fp8 {
+            // W8A8: quantize activations with per-expert scalar scale, use cuBLASLt FP8 GEMM
+            let (x_fp8, x_scales) = xserv_kernels::quantization::quantize_bf16_to_fp8_rowwise(&x_rep);
+            xserv_kernels::quantization::batched_gemm_fp8(
+                &x_fp8, &x_scales, wt_fp8_t, layer.expert_gate_up_scale.as_ref().unwrap(),
+            )
+        } else {
+            xserv_kernels::moe::batched_gemm_strided(&x_rep, &layer.expert_gate_up_wt)
+        };
+
+        // 5. Bias add: gate_up += expert_gate_up_bias  (in-place)
+        xserv_kernels::moe::moe_bias_add_3d(&gate_up, &layer.expert_gate_up_bias);
+
+        // 6. GLU activation: treat [E * tokens, 2*inter] → [E * tokens, inter]
+        let inter2 = gate_up.shape()[2];
+        let flat_rows = local_experts * num_tokens;
+        let gate_up_flat = gate_up.reshape(&[flat_rows, inter2]);
+        let activated_flat = gpt_oss_glu(&gate_up_flat, layer.glu_alpha, layer.glu_limit);
+        let inter = inter2 / 2;
+        let activated = activated_flat.reshape(&[local_experts, num_tokens, inter]);
+
+        // 7. Batched GEMM down: [E, tokens, inter] @ [E, inter, hidden] → [E, tokens, hidden]
+        let down = if let Some((ref packed, ref scales)) = layer.expert_down_mxfp4 {
+            let n = packed.shape()[1];
+            let k = packed.shape()[2] * 2;
+            if num_tokens == 1 {
+                let a2 = activated.reshape(&[local_experts, k]);
+                xserv_kernels::quantization::batched_gemv_mxfp4(&a2, packed, scales, n, k)
+                    .reshape(&[local_experts, 1, n])
+            } else {
+                let w_bf16 = xserv_kernels::quantization::dequant_mxfp4_to_bf16_t(packed, scales, local_experts, n, k);
+                xserv_kernels::moe::batched_gemm_strided(&activated, &w_bf16)
+            }
+        } else if let Some(ref wt_fp8) = layer.expert_down_fp8 {
+            // W8A8: quantize post-GLU activations to FP8, use cuBLASLt FP8 GEMM
+            let (act_fp8, act_scales) = xserv_kernels::quantization::quantize_bf16_to_fp8_rowwise(&activated);
+            xserv_kernels::quantization::batched_gemm_fp8(
+                &act_fp8, &act_scales, wt_fp8, layer.expert_down_scale.as_ref().unwrap(),
+            )
+        } else {
+            xserv_kernels::moe::batched_gemm_strided(&activated, &layer.expert_down_wt)
+        };
+
+        // 8. Bias add: down += expert_down_bias  (in-place)
+        xserv_kernels::moe::moe_bias_add_3d(&down, &layer.expert_down_bias);
+
+        // 9. Weighted sum across experts → [tokens, hidden]
+        let moe_out = xserv_kernels::moe::moe_weighted_sum(
+            &down, &topk_ids, &topk_weights,
+            expert_start, local_experts, top_k,
+        );
+
+        self.all_reduce(&moe_out);
+        moe_out
+    }
+}
+
+// --- Helpers ---
+
+fn matmul_2d(a: &Tensor, b: &Tensor) -> Tensor {
+    assert_eq!(a.ndim(), 2);
+    assert_eq!(b.ndim(), 2);
+    matmul(a, b, GemmBackend::CuBlas)
+}
+
+/// Add bias to a 2D tensor: [rows, cols] + [cols] → [rows, cols]
+fn add_bias(x: &Tensor, bias: &Tensor) -> Tensor {
+    assert_eq!(x.ndim(), 2);
+    assert_eq!(bias.ndim(), 1);
+    let rows = x.shape()[0];
+    let cols = x.shape()[1];
+    assert_eq!(bias.shape()[0], cols, "bias size {} != cols {}", bias.shape()[0], cols);
+
+    let x_c = x.contiguous();
+
+    if rows == 1 {
+        // Fast path: reshape bias [cols] → [1, cols] (zero-copy), add directly on GPU
+        let bias_2d = bias.reshape(&[1, cols]);
+        return xserv_kernels::add(&x_c, &bias_2d);
+    }
+
+    // General path: tile bias to [rows, cols] via CPU, then add on GPU
+    let bias_cpu = bias.to_device(Device::Cpu);
+    let bias_data = bias_cpu.as_slice::<bf16>();
+    let mut tiled = Vec::with_capacity(rows * cols);
+    for _ in 0..rows {
+        tiled.extend_from_slice(bias_data);
+    }
+    let bias_tiled = Tensor::from_slice(&tiled, &[rows, cols]).to_device(x.device());
+    xserv_kernels::add(&x_c, &bias_tiled)
+}
+
+fn shard_rows(t: &Tensor, rank: usize, world: usize) -> Tensor {
+    if world == 1 { return t.clone(); }
+    let shape = t.shape();
+    assert_eq!(shape.len(), 2);
+    let (rows, cols) = (shape[0], shape[1]);
+    assert!(rows % world == 0, "rows {rows} not divisible by world {world}");
+    let local = rows / world;
+    let host = t.to_device(Device::Cpu);
+    let data = host.as_slice::<bf16>();
+    let start = rank * local * cols;
+    let shard = data[start..start + local * cols].to_vec();
+    Tensor::from_slice(&shard, &[local, cols])
+}
+
+fn shard_cols(t: &Tensor, rank: usize, world: usize) -> Tensor {
+    if world == 1 { return t.clone(); }
+    let shape = t.shape();
+    assert_eq!(shape.len(), 2);
+    let (rows, cols) = (shape[0], shape[1]);
+    assert!(cols % world == 0, "cols {cols} not divisible by world {world}");
+    let local = cols / world;
+    let c0 = rank * local;
+    let host = t.to_device(Device::Cpu);
+    let data = host.as_slice::<bf16>();
+    let mut shard = Vec::with_capacity(rows * local);
+    for r in 0..rows {
+        let base = r * cols + c0;
+        shard.extend_from_slice(&data[base..base + local]);
+    }
+    Tensor::from_slice(&shard, &[rows, local])
+}
+
+fn shard_1d(t: &Tensor, rank: usize, world: usize) -> Tensor {
+    if world == 1 { return t.clone(); }
+    let shape = t.shape();
+    assert_eq!(shape.len(), 1);
+    let total = shape[0];
+    assert!(total % world == 0, "dim {total} not divisible by world {world}");
+    let local = total / world;
+    let host = t.to_device(Device::Cpu);
+    let data = host.as_slice::<bf16>();
+    let start = rank * local;
+    let shard = data[start..start + local].to_vec();
+    Tensor::from_slice(&shard, &[local])
+}
+
+/// Transpose the inner two dimensions of a [batch, rows, cols] tensor → [batch, cols, rows].
+/// Works on raw bytes (any dtype). CPU-only.
+fn transpose_3d_inner_raw(t: &Tensor, batch: usize, rows: usize, cols: usize) -> Tensor {
+    assert_eq!(t.ndim(), 3);
+    assert_eq!(t.shape(), &[batch, rows, cols]);
+    let host = t.to_device(Device::Cpu);
+    let es = t.dtype().size_bytes();
+    let raw = host.as_raw_bytes();
+    let mut out = vec![0u8; batch * cols * rows * es];
+    for b in 0..batch {
+        for r in 0..rows {
+            for c in 0..cols {
+                let src_off = (b * rows * cols + r * cols + c) * es;
+                let dst_off = (b * cols * rows + c * rows + r) * es;
+                out[dst_off..dst_off + es].copy_from_slice(&raw[src_off..src_off + es]);
+            }
+        }
+    }
+    Tensor::from_raw_bytes(&out, &[batch, cols, rows], t.dtype())
+}
+
+/// Extract experts [start..start+count) from a [num_experts, rows, cols] 3D tensor (any dtype, raw bytes).
+fn slice_expert_range_3d_raw(t: &Tensor, start: usize, count: usize, rows: usize, cols: usize) -> Tensor {
+    assert_eq!(t.ndim(), 3);
+    let host = t.to_device(Device::Cpu);
+    let elem_size = t.dtype().size_bytes();
+    let raw = host.as_raw_bytes();
+    let stride = rows * cols * elem_size;
+    let offset = start * stride;
+    let slice = &raw[offset..offset + count * stride];
+    Tensor::from_raw_bytes(slice, &[count, rows, cols], t.dtype())
+}
+
+/// Slice scale tensor [num_experts] F32 → [count] starting at `start`.
+fn slice_scale_range(t: &Tensor, start: usize, count: usize) -> Tensor {
+    assert_eq!(t.ndim(), 1);
+    assert_eq!(t.dtype(), xserv_tensor::DType::F32);
+    let host = t.to_device(Device::Cpu);
+    let data = host.as_slice::<f32>();
+    let slice = data[start..start + count].to_vec();
+    Tensor::from_slice(&slice, &[count])
+}
+
+/// Extract experts [start..start+count) from a [num_experts, rows, cols] 3D tensor
+fn slice_expert_range_3d(t: &Tensor, start: usize, count: usize, rows: usize, cols: usize) -> Tensor {
+    assert_eq!(t.ndim(), 3);
+    let host = t.to_device(Device::Cpu);
+    let data = host.as_slice::<bf16>();
+    let stride = rows * cols;
+    let offset = start * stride;
+    let slice = data[offset..offset + count * stride].to_vec();
+    Tensor::from_slice(&slice, &[count, rows, cols])
+}
+
+/// Extract experts [start..start+count) from a [num_experts, dim] 2D tensor
+fn slice_expert_range_2d(t: &Tensor, start: usize, count: usize, dim: usize) -> Tensor {
+    assert_eq!(t.ndim(), 2);
+    let host = t.to_device(Device::Cpu);
+    let data = host.as_slice::<bf16>();
+    let offset = start * dim;
+    let slice = data[offset..offset + count * dim].to_vec();
+    Tensor::from_slice(&slice, &[count, dim])
+}

crates/xserv-model/src/kv_cache.rs

@@ -0,0 +1,151 @@
+use xserv_cuda::GpuBuffer;
+use xserv_tensor::{DType, Device, Tensor};
+use crate::config::ModelConfig;
+
+/// 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 xserv_tensor::storage::Storage;
+    use xserv_tensor::shape::contiguous_strides;
+    use smallvec::SmallVec;
+
+    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

@@ -1,8 +1,24 @@
 pub mod config;
+pub mod decode_graph;
 pub mod gpt2;
+pub mod gpt_oss;
+pub mod kv_cache;
 pub mod loader;
+pub mod paged_kv_cache;
 pub mod qwen3;
+pub mod sampling;
 
 pub use config::ModelConfig;
+pub use decode_graph::{DecodeGraphState, LayerWeightPtrs};
 pub use gpt2::{GPT2, KVCache};
+pub use gpt_oss::GptOss;
+pub use kv_cache::GpuKVCache;
+pub use paged_kv_cache::{BlockAllocator, Location, PagedKVCache, BLOCK_SIZE};
 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

@@ -19,6 +19,7 @@ pub fn load_safetensors(path: &Path, device: Device) -> HashMap<String, Tensor>
             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;
@@ -83,5 +84,8 @@ fn make_tensor(raw_bytes: &[u8], shape: &[usize], dtype: DType) -> Tensor {
             };
             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 +1,639 @@
+//! 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,
+                std::ptr::null_mut(),
+            );
+        }
+        // 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,
+                std::ptr::null_mut(),
+            );
+        }
+    }
+
+    /// 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;
+    }
+
+    /// 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(())
+    }
+}
+
+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/sampling.rs

@@ -0,0 +1,162 @@
+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);
+    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 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);
+    }
+
+    // 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 +1,516 @@
+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;
+            }
+        }
+    }
+
+    Json(serde_json::json!({
+        "id": id,
+        "object": "chat.completion",
+        "created": created,
+        "model": model_name,
+        "choices": [{
+            "index": 0,
+            "message": { "role": "assistant", "content": content },
+            "finish_reason": finish_reason,
+        }],
+        "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;
+                    }
+                    let chunk =
+                        make_chunk(&id, &model_name, created, None, None, Some(&finish_reason));
+                    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"));
+    }
+
+    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.send(req).map_err(|_| 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),
+    }
+}

crates/xserv-server/src/engine.rs

@@ -0,0 +1,379 @@
+use std::collections::VecDeque;
+use std::path::Path;
+use std::sync::mpsc;
+use std::sync::Once;
+use std::time::Instant;
+use xserv_model::{ModelConfig, PagedKVCache, Qwen3, SamplingParams, sample, BLOCK_SIZE};
+use xserv_model::loader;
+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,
+    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,
+            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);
+        let _ = seq.sender.blocking_send(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);
+        let _ = seq.sender.blocking_send(GenerateEvent::Done {
+            finish_reason: "length".to_string(),
+        });
+    } else {
+        send_token_if_nonempty(seq, text);
+    }
+}
+
+fn send_token_if_nonempty(seq: &Sequence, text: String) {
+    if !text.is_empty() {
+        let id = *seq.generated_tokens.last().unwrap_or(&0);
+        let _ = seq.sender.blocking_send(GenerateEvent::Token { id, text });
+    }
+}
+
+fn is_finished(seq: &Sequence) -> bool {
+    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 +1,123 @@
+mod api;
+mod engine;
+mod pp_engine;
+mod tp_engine;
+
+use axum::{routing::{get, post}, Extension, Router};
+use std::path::PathBuf;
+use std::sync::{mpsc, Arc, Mutex};
+use engine::GenerateRequest;
+use xserv_model::ModelConfig;
+
+pub struct AppState {
+    pub model_name: String,
+    pub chat_template: api::ChatTemplate,
+    pub engine_sender: Mutex<mpsc::Sender<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"));
+    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"));
+
+    // Unbounded channel: allows multiple requests to queue up
+    let (tx, rx) = mpsc::channel::<GenerateRequest>();
+
+    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 {
+            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(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 +1,263 @@
+//! 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::mpsc;
+use std::sync::Arc;
+use std::thread;
+
+use half::bf16;
+use xserv_distributed::{PpContext, UniqueId};
+use xserv_model::loader;
+use xserv_model::sampling::SamplingParams;
+use xserv_model::{sample, ModelConfig, PagedKVCache, Qwen3, BLOCK_SIZE};
+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;
+        emit_text(&tokenizer, &req, next, &mut decode_buf);
+
+        let finish = loop {
+            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;
+            emit_text(&tokenizer, &req, next, &mut decode_buf);
+        };
+
+        let tail = tokenizer.flush_decode_stream(&mut decode_buf);
+        if !tail.is_empty() {
+            let _ = req.sender.blocking_send(GenerateEvent::Token { id: next, text: tail });
+        }
+        let _ = req.sender.blocking_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).
+fn emit_text(tokenizer: &Tokenizer, req: &GenerateRequest, token_id: u32, buf: &mut Vec<u8>) {
+    if tokenizer.is_eos(token_id) {
+        return;
+    }
+    let text = tokenizer.decode_token_stream(token_id, buf);
+    if !text.is_empty() {
+        let _ = req.sender.blocking_send(GenerateEvent::Token { id: token_id, text });
+    }
+}

crates/xserv-server/src/tp_engine.rs

@@ -0,0 +1,239 @@
+//! 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::mpsc;
+use std::sync::Arc;
+use std::thread;
+
+use xserv_distributed::{TpContext, UniqueId};
+use xserv_model::loader;
+use xserv_model::{sample, sample_greedy_penalized, GptOss, ModelConfig, PagedKVCache, Qwen3, BLOCK_SIZE};
+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,
+}
+
+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 }
+}
+
+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 _ = rc.model.forward_decode_paged(&tokens, &positions, &slots, &mut rc.cache);
+            }
+            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>) {
+    assert!(world >= 2, "run_tp requires world >= 2");
+    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;
+        emit_text(&tokenizer, &req, next, &mut decode_buf);
+
+        let finish = loop {
+            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 = rc.model.forward_decode_paged(&[next], &[pos], &[slot], &mut rc.cache);
+            wait_acks(&ack_rx);
+            next = pick(&logits, &req.sampling, &gen_ids);
+            gen_ids.push(next);
+            generated += 1;
+            emit_text(&tokenizer, &req, next, &mut decode_buf);
+        };
+
+        let tail = tokenizer.flush_decode_stream(&mut decode_buf);
+        if !tail.is_empty() {
+            let _ = req.sender.blocking_send(GenerateEvent::Token { id: next, text: tail });
+        }
+        let _ = req.sender.blocking_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).
+fn emit_text(tokenizer: &Tokenizer, req: &GenerateRequest, token_id: u32, buf: &mut Vec<u8>) {
+    if tokenizer.is_eos(token_id) {
+        return;
+    }
+    let text = tokenizer.decode_token_stream(token_id, buf);
+    if !text.is_empty() {
+        let _ = req.sender.blocking_send(GenerateEvent::Token { id: token_id, text });
+    }
+}

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",
         }
     }
 }

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 shape::Dims;
-pub use tensor::Tensor;
+pub use storage::{Device, Storage};
+pub use tensor::{register_gpu_contiguous, Tensor};

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);
+    let ndim = shape.len();
-    strides == expected.as_slice()
+    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.

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 {
+    pub fn cuda(buffer: GpuBuffer, device: u32) -> Self {
-        Self(Arc::new(StorageInner::Cuda { buffer }))
+        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 } => {
+            StorageInner::Cuda { buffer, device } => {
-                let mut dst = GpuBuffer::alloc(buffer.len())?;
+                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(_) => {
+            Device::Cuda(dev) => {
-                let mut buf = GpuBuffer::alloc(len_bytes)?;
+                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 +32,11 @@ 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 +52,25 @@ 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();
@@ -46,12 +84,29 @@ 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"),
         }
     }
 
@@ -85,6 +140,21 @@ 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 +188,15 @@ 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 };
+        let new_strides = if self.is_contiguous() {
-        new_strides.insert(dim, stride_val);
+            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,
@@ -137,9 +212,12 @@ impl Tensor {
         if self.is_contiguous() {
             return self.clone();
         }
-        // For GPU tensors: round-trip through CPU (correct but slow).
+        // For GPU tensors: use the registered GPU kernel if available,
-        // TODO: write a GPU contiguous-copy kernel for performance.
+        // 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());
@@ -207,6 +285,17 @@ 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() {
@@ -232,3 +321,58 @@ impl std::fmt::Debug for Tensor {
         )
     }
 }
+
+#[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-tokenizer/src/bpe.rs

@@ -12,6 +12,7 @@ pub struct Tokenizer {
     special_token_ids: HashMap<u32, String>,
     pre_tokenize_re: Regex,
     eos_token_id: Option<u32>,
+    eos_token_ids: Vec<u32>,
     byte_fallback: bool,
 }
 
@@ -20,6 +21,24 @@ 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)]
@@ -41,6 +60,7 @@ enum MergeEntry {
 struct AddedToken {
     id: u32,
     content: String,
+    #[allow(dead_code)]
     special: bool,
 }
 
@@ -51,8 +71,6 @@ impl Tokenizer {
         let tj: TokenizerJson = serde_json::from_str(&data)
             .unwrap_or_else(|e| panic!("failed to parse tokenizer.json: {e}"));
 
-        let byte_fallback = tj.model.byte_fallback;
-
         // Build encoder: token bytes → ID
         // All HF tokenizers use GPT-2 byte-to-unicode mapping for vocab keys.
         let mut encoder = HashMap::new();
@@ -92,28 +110,77 @@ impl Tokenizer {
             }
         }
 
-        // Special tokens
+        // 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();
-        let mut eos_token_id = None;
         for at in &tj.added_tokens {
-            if at.special {
+            special_tokens.insert(at.content.clone(), at.id);
-                special_tokens.insert(at.content.clone(), at.id);
+            special_token_ids.insert(at.id, at.content.clone());
-                special_token_ids.insert(at.id, at.content.clone());
+            decoder.resize(decoder.len().max(at.id as usize + 1), vec![]);
-                decoder.resize(decoder.len().max(at.id as usize + 1), vec![]);
+            decoder[at.id as usize] = at.content.as_bytes().to_vec();
-                decoder[at.id as usize] = at.content.as_bytes().to_vec();
+        }
-                if at.content == "<|endoftext|>" || at.content == "<|end_of_text|>" {
+        // End-of-generation tokens, in priority order. Families differ:
-                    eos_token_id = Some(at.id);
+        // 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
+        // Pre-tokenization regex: prefer the model's own regex from tokenizer.json,
-        let pre_tokenize_re = if byte_fallback {
+        // fall back to GPT-2/Qwen heuristic if not present or unsupported.
-            // Qwen-style: split on whitespace boundaries, keep Unicode words/numbers
+        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 {
-            // GPT-2 style
             Regex::new(r"'s|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+").unwrap()
         };
 
@@ -125,6 +192,7 @@ impl Tokenizer {
             special_token_ids,
             pre_tokenize_re,
             eos_token_id,
+            eos_token_ids,
             byte_fallback,
         }
     }
@@ -170,10 +238,26 @@ impl Tokenizer {
             }
             // Fall back to per-byte encoding
             let word_bytes: Vec<u8> = word.bytes().collect();
-            let mut token_ids: Vec<u32> = word_bytes.iter().map(|&b| {
+            let mut token_ids: Vec<u32> = word_bytes.iter().filter_map(|&b| {
-                *self.encoder.get(&vec![b]).unwrap_or_else(|| {
+                if let Some(&id) = self.encoder.get(&vec![b]) {
-                    panic!("byte {b} (0x{b:02X}) not in vocab")
+                    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
@@ -216,10 +300,29 @@ impl Tokenizer {
         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()
     }
@@ -236,6 +339,31 @@ 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
@@ -265,3 +393,49 @@ fn unicode_to_byte(c: char) -> u8 {
         panic!("unmapped unicode char U+{:04X} in tokenizer", c as u32)
     })
 }
+
+#[cfg(test)]
+mod tests {
+    use super::{take_valid_utf8, Tokenizer};
+
+    #[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());
+    }
+}

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)))
@@ -45,6 +46,37 @@ __global__ void scale_bf16_kernel(const __nv_bfloat16* x, __nv_bfloat16* out, fl
     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;
@@ -71,6 +103,7 @@ 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) {
@@ -78,12 +111,14 @@ 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) {
@@ -91,6 +126,7 @@ 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) {
@@ -98,6 +134,7 @@ void launch_scale_f32(const void* x, void* out, float scale, int n, void* stream
     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) {
@@ -105,6 +142,7 @@ void launch_scale_bf16(const void* x, void* out, float scale, int n, void* strea
     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) {
@@ -112,24 +150,45 @@ void launch_add_f32(const void* a, const void* b, void* out, int n, void* stream
     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_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

@@ -1,4 +1,5 @@
 #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`,
@@ -27,8 +28,7 @@ __global__ void causal_mask_bf16(
     int col = blockIdx.x * blockDim.x + threadIdx.x;
 
     if (col < cols && col > row + offset) {
-        // BF16 doesn't have proper -inf literal, use a very large negative
+        scores[batch_idx * rows * cols + row * cols + col] = __float2bfloat16(-INFINITY);
-        scores[batch_idx * rows * cols + row * cols + col] = __float2bfloat16(-1e9f);
     }
 }
 
@@ -40,6 +40,7 @@ void launch_causal_mask_f32(void* scores, int batch, int rows, int cols,
     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,
@@ -48,6 +49,7 @@ void launch_causal_mask_bf16(void* scores, int batch, int rows, int cols,
     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 +1,621 @@
+#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[HEAD_DIM_MAX]; // final output accumulator
+
+    // 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 (dimension by dimension using shared mem)
+    float inv_sum = (global_sum > 0.0f) ? (1.0f / global_sum) : 0.0f;
+
+    // Process head_dim in chunks: each iteration reduces one dimension
+    // Use shared memory accumulator: each warp contributes via warp reduction + atomic
+    // Actually simpler: iterate over dimensions, warp reduce each, then lane0 atomicAdd to smem_O
+
+    // Initialize smem_O
+    for (int d = tid; d < head_dim; d += DECODE_THREADS) {
+        smem_O[d] = 0.0f;
+    }
+    __syncthreads();
+
+    // Each thread adds its local_O contributions via warp reduction + atomicAdd
+    for (int d = 0; d < head_dim; d++) {
+        float val = local_O[d];
+        // Warp-level reduction
+        #pragma unroll
+        for (int offset = 16; offset > 0; offset >>= 1)
+            val += __shfl_down_sync(0xffffffff, val, offset);
+        if (lane == 0) {
+            atomicAdd(&smem_O[d], val);
+        }
+    }
+    __syncthreads();
+
+    // Thread 0..head_dim-1 write final output
+    for (int d = tid; d < head_dim; d += DECODE_THREADS) {
+        O_ptr[d] = __float2bfloat16(smem_O[d] * 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 +1,411 @@
+#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[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
+    for (int d = tid; d < head_dim; d += PAGED_THREADS) smem_O[d] = 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) atomicAdd(&smem_O[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) {
+        O_ptr[d] = __float2bfloat16(smem_O[d] * 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[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 d = tid; d < head_dim; d += PAGED_THREADS) smem_O[d] = 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) atomicAdd(&smem_O[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) {
+        O_ptr[d] = __float2bfloat16(smem_O[d] * 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_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 +1,161 @@
+#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();
+}
+
+}

csrc/common.cuh

@@ -48,3 +48,17 @@ __device__ __forceinline__ float block_reduce_max(float val) {
     if (warp_id == 0) val = warp_reduce_max(val);
     return val;
 }
+
+// --- Launch error checking (debug builds only) ---
+#ifdef NDEBUG
+#define CUDA_CHECK_LAST_ERROR() ((void)0)
+#else
+#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)
+#endif

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
+// RoPE: Rotary Position Embedding, using the Qwen/Llama rotate_half layout.
-// For each pair (x[2i], x[2i+1]) at position `pos`:
+// For each dimension i in the first half at position `pos`:
-//   y[2i]   = x[2i] * cos - x[2i+1] * sin
+//   y[i]            = x[i] * cos - x[i + half_dim] * sin
-//   y[2i+1] = x[2i] * sin + x[2i+1] * cos
+//   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 x0 = x[base + pair_idx];
-    float x1 = x[base + 2 * pair_idx + 1];
+    float x1 = x[base + pair_idx + half_dim];
 
-    x[base + 2 * pair_idx]     = x0 * cos_val - x1 * sin_val;
+    x[base + pair_idx]            = x0 * cos_val - x1 * sin_val;
-    x[base + 2 * pair_idx + 1] = x0 * sin_val + x1 * cos_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 x0 = __bfloat162float(x[base + pair_idx]);
-    float x1 = __bfloat162float(x[base + 2 * pair_idx + 1]);
+    float x1 = __bfloat162float(x[base + pair_idx + half_dim]);
 
-    x[base + 2 * pair_idx]     = __float2bfloat16(x0 * cos_val - x1 * sin_val);
+    x[base + 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 + 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 +1,110 @@
+#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).
+// All blocks atomicAdd their partial sums into a pre-zeroed FP32 buffer.
+// A separate conversion kernel writes the final BF16 output.
+// Launch sequence: cudaMemsetAsync(fp32) → accumulation kernel → convert kernel.
+
+#define GEMV_TILE_N 128
+#define GEMV_TILE_K 256
+#define GEMV_BLOCK  128
+
+__global__ void gemv_bf16_fused_kernel(
+    const __nv_bfloat16* __restrict__ x,
+    const __nv_bfloat16* __restrict__ W,
+    __nv_bfloat16* __restrict__ y_bf16,
+    float* __restrict__ y_fp32,
+    int K, int N,
+    int num_k_blocks
+) {
+    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]);
+    }
+
+    atomicAdd(&y_fp32[col], sum);
+}
+
+// Conversion kernel: FP32 accumulator -> BF16 output
+__global__ void gemv_fp32_to_bf16_kernel(
+    const float* __restrict__ src,
+    __nv_bfloat16* __restrict__ dst,
+    int n
+) {
+    int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    if (idx < n) {
+        dst[idx] = __float2bfloat16(src[idx]);
+    }
+}
+
+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;
+
+    // Zero the FP32 accumulator BEFORE the kernel — the kernel uses atomicAdd
+    // across K-blocks with no inter-block ordering, so the buffer must be
+    // pre-zeroed to avoid accumulating on stale data.
+    cudaMemsetAsync(y_fp32_buf, 0, (size_t)N * sizeof(float), s);
+
+    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_fused_kernel<<<grid, GEMV_BLOCK, 0, s>>>(
+        (const __nv_bfloat16*)x,
+        (const __nv_bfloat16*)W,
+        (__nv_bfloat16*)y_bf16,
+        (float*)y_fp32_buf,
+        K, N, num_k_blocks
+    );
+    CUDA_CHECK_LAST_ERROR();
+
+    // FP32 → BF16 conversion (must wait for all K-blocks to finish)
+    int conv_block = 256;
+    int conv_grid = (N + conv_block - 1) / conv_block;
+    gemv_fp32_to_bf16_kernel<<<conv_grid, conv_block, 0, s>>>(
+        (const float*)y_fp32_buf,
+        (__nv_bfloat16*)y_bf16,
+        N
+    );
+    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 +1,247 @@
+#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
+) {
+    int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    int total = local_experts * num_tokens * hidden;
+    if (idx >= total) return;
+
+    int expert = idx / (num_tokens * hidden);
+    int remainder = idx % (num_tokens * hidden);
+    // x_rep[expert, token, dim] = x[token, dim]
+    x_rep[idx] = x[remainder];
+    (void)expert; // suppress unused warning
+}
+
+// ============================================================
+// 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
+) {
+    int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    int total = batch * num_tokens * dim;
+    if (idx >= total) return;
+
+    int b = idx / (num_tokens * dim);
+    int d = idx % dim;
+    float v = __bfloat162float(x[idx]) + __bfloat162float(bias[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
+) {
+    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;
+
+    int expert_stride = 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
+) {
+    int total = local_experts * num_tokens * hidden;
+    int block = 256;
+    int grid = (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
+) {
+    int total = batch * num_tokens * dim;
+    int block = 256;
+    int grid = (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
+) {
+    int total = num_tokens * hidden;
+    int block = 256;
+    int grid = (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/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 += x_row[i];
-        local_sum += v;
-        local_sum_sq += v * v;
     }
     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;
+        s_mean = local_sum / hidden_size;
-        float var = local_sum_sq / hidden_size - mean * mean;
-        s_mean = mean;
-        s_inv_std = rsqrtf(var + eps);
     }
     __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 += __bfloat162float(x_row[i]);
-        local_sum += v;
-        local_sum_sq += v * v;
     }
     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;
+        s_mean = local_sum / hidden_size;
-        float var = local_sum_sq / hidden_size - mean * mean;
-        s_mean = mean;
-        s_inv_std = rsqrtf(var + eps);
     }
     __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 +1,51 @@
+#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
+) {
+    int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    int total = num_experts * rows * cols;
+    if (idx >= total) return;
+
+    int expert_stride = rows * cols;
+    int expert = 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
+) {
+    int total = num_experts * rows * cols;
+    int block = 256;
+    int grid = (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

@@ -94,6 +94,7 @@ void launch_softmax_f32(const void* x, void* out, int rows, int cols, void* stre
     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) {
@@ -101,6 +102,7 @@ void launch_softmax_bf16(const void* x, void* out, int rows, int cols, void* str
     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
+- [ ] 加载 Qwen3-8B 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)
+- [ ] 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) |
+| Hidden Size | 768 | 4,096 (8B) |
-| Layers | 12 | 28 |
+| 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
+模型权重:  8B × 2B = ~16 GB
-KV cache:  28 layers × 2(KV) × 8 heads × 4096 tokens × 128 dim × 2B ≈ 4.5 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)
+   - 5090 SM120: shared memory per SM = 100 KB (需实测确认)
-   - 需要同时存 Q_tile, K_tile, V_tile, S_tile
+   - 需同时存 Q_tile, K_tile, V_tile, S_tile
-   - 典型值: Br=Bc=128 for D=128, BF16
+   - 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**:
+4. **GQA 支持**:
-   - Flash Attention 的 K/V tile 遍历逻辑需要适配间接寻址
+   - K/V heads 数量 < Q heads 时，kernel 中做 `kv_head = q_head / num_groups` 索引
-   - 每个 tile 查 block_table 得到物理地址
+   - 不需要 repeat_kv 操作，直接在 kernel 内部解决
-   - 这是 "Flash-Decoding" / "FlashInfer" 的核心
+
+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
 
 ---
@@ -1727,7 +1753,7 @@ Text  → Tokenizer   → Text Tokens   ────────────→
 | 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 |

docs/10-qwen3.md

@@ -1,12 +1,12 @@
-# Phase 10: Qwen3-7B Support — Design Document (Milestone ②)
+# Phase 10: Qwen3-8B Support — Design Document (Milestone ②)
 
 ## Goal
 
-扩展模型定义支持 Qwen3-7B 架构，验证输出正确性。与 GPT-2 的关键差异：RMSNorm、RoPE、GQA、SwiGLU、不共享 embedding。
+扩展模型定义支持 Qwen3-8B 架构，验证输出正确性。与 GPT-2 的关键差异：RMSNorm、RoPE、GQA、SwiGLU、不共享 embedding。
 
 ## 架构差异 (GPT-2 → Qwen3)
 
-| 特性 | GPT-2 | Qwen3-7B |
+| 特性 | GPT-2 | Qwen3-8B |
 |------|-------|----------|
 | Norm | LayerNorm(gamma, beta) | RMSNorm(gamma only) |
 | Position | Learned absolute (wpe) | RoPE (no params) |
@@ -15,8 +15,8 @@
 | FFN | 2 Linear (fc, proj) + GELU | 3 Linear (gate, up, down) + SwiGLU |
 | Weight layout | [in, out] (Conv1D style) | [out, in] (standard Linear) |
 | Tied embeddings | Yes | No (separate lm_head) |
-| hidden_size | 768 | 3584 |
+| hidden_size | 768 | 4096 |
-| num_layers | 12 | 28 |
+| num_layers | 12 | 36 |
 | head_dim | 64 | 128 |
 
 ## Weight Names (HuggingFace)
@@ -67,17 +67,17 @@ out  = down_proj(out)    # [S, 18944] @ [18944, 3584]^T → [S, 3584]
 ## 显存预算 (BF16, 单卡 5090)
 
 ```
-权重: 7B × 2B = ~14 GB (BF16)
+权重: 8B × 2B = ~16 GB (BF16)
-        7B × 4B = ~28 GB (FP32) — 不够! 必须用 BF16
+        8B × 4B = ~32 GB (FP32) — 不够! 必须用 BF16
 KV cache (S=256, B=1): ~0.1 GB
-总计: ~14 GB (BF16), 单卡可运行
+总计: ~16 GB (BF16), 单卡可运行
 ```
 
-**关键**: Qwen3-7B 必须用 BF16 才能在单张 5090 (32GB) 上运行。当前 GPT-2 用 FP32，需要支持 BF16 forward pass。
+**关键**: Qwen3-8B 必须用 BF16 才能在单张 5090 (32GB) 上运行。当前 GPT-2 用 FP32，需要支持 BF16 forward pass。
 
 ## Implementation Plan
 
-1. 下载 Qwen3-7B 模型 (BF16, ~14GB)
+1. 下载 Qwen3-8B 模型 (BF16, ~14GB)
 2. 实现 Qwen3 模型结构 (qwen3.rs)
 3. 支持 BF16 forward pass (linear_transpose for [out, in] weights)
 4. 实现 GQA (K/V repeat in split)

docs/11-paged-attention.md

@@ -0,0 +1,61 @@
+# Phase 11: GPU-Resident KV Cache — Design Document
+
+> **注意**: 原计划为 "Paged Attention + KV Cache Manager"，实际实现为 GPU 连续预分配 KV cache（非 paged）。Paged allocation 留待后续优化。
+
+## Goal
+
+将 KV cache 从 CPU Vec 迁移到 GPU，消除每步 decode 的 CPU round-trip（当前 KV cache 最大性能瓶颈之一）。
+
+## 当前问题
+
+每步 decode 的 KV cache 路径：
+```
+GPU tensor (K_new) → CPU (per-head Vec append) → reconstruct → CPU tensor → GPU tensor
+```
+这涉及 2 次 GPU↔CPU 拷贝 × 36 层 × 2(K,V) = 144 次 transfer/token。
+
+## 目标设计
+
+KV cache 直接存在 GPU 上，decode 时只做 GPU→GPU append：
+```
+GPU tensor (K_new) → GPU KV cache (in-place append, no CPU)
+```
+
+## 实现方案
+
+### GPU KV Cache（简化版，非 paged）
+
+先实现连续分配的 GPU KV cache（预分配 max_seq_len），消除 CPU round-trip。Paged allocation 留待后续优化。
+
+```rust
+pub struct GpuKVCache {
+    // 预分配: [num_layers, 2, num_kv_heads, max_seq_len, head_dim] on GPU
+    k_caches: Vec<Tensor>,  // per layer: [1, num_kv_heads, max_seq_len, head_dim]
+    v_caches: Vec<Tensor>,
+    seq_len: usize,         // 当前已填充的长度
+    max_seq_len: usize,
+}
+```
+
+### Append 操作
+
+用 cudaMemcpy D2D 将新 K/V 写入 cache 的正确偏移位置：
+```
+k_cache[layer][0, :, seq_len:seq_len+new, :] = k_new[0, :, :, :]
+```
+
+### 读取操作
+
+不需要拷贝——直接用 view/slice 返回 [0, :, 0:seq_len, :] 的 GPU tensor。
+
+## 需要的新功能
+
+1. Tensor slice 支持（view into sub-range of a dimension）
+2. GPU D2D copy at offset（写入 cache 指定位置）
+3. 去掉 Qwen3/GPT-2 forward 中的 CPU round-trip KV cache 路径
+
+## Test Plan
+
+- [ ] GPU KV cache 输出与 CPU KV cache bit-identical
+- [ ] Benchmark: TBT 应显著降低（消除 144 次 CPU round-trip）
+- [ ] 50-prompt correctness re-validation

docs/12-continuous-batching.md

@@ -0,0 +1,157 @@
+# Phase 12: Continuous Batching + Request Scheduler — Design Document
+
+## Goal
+
+实现 iteration-level 请求调度，支持多个请求并发生成 token。核心能力：同时发 N 个请求，N 个请求同时产出 token，新请求可以在 mid-generation 加入 batch。
+
+## 为什么需要 Continuous Batching
+
+**当前问题（串行）**：
+```
+时间 →  [req1 prefill][req1 decode x 100][req2 prefill][req2 decode x 50]...
+GPU利用: ████████████████████████████████████████████████████████████████████
+         req2 等了 100 个 token 的时间才开始
+```
+
+**目标（continuous batching）**：
+```
+时间 →  [req1+req2 prefill][req1+req2 decode][req1 done, req3 加入][req2+req3 decode]...
+GPU利用: ████████████████████████████████████████████████████████████████████
+         req2 和 req1 同时推理，req3 在 req1 完成后立即加入
+```
+
+## 核心设计
+
+### 数据结构
+
+```rust
+pub struct Sequence {
+    pub id: u64,
+    pub prompt_tokens: Vec<u32>,
+    pub generated_tokens: Vec<u32>,
+    pub status: SeqStatus,
+    pub max_tokens: usize,
+    pub kv_cache: GpuKVCache,          // 每个 seq 独立的 KV cache
+    pub output_tx: mpsc::Sender<GenerateEvent>,
+}
+
+pub enum SeqStatus {
+    Waiting,     // 在队列中等待被 admit
+    Running,     // 正在参与 batch forward
+    Finished,    // EOS 或 max_tokens 达到
+}
+
+pub struct Scheduler {
+    waiting: VecDeque<Sequence>,
+    running: Vec<Sequence>,
+    max_batch_size: usize,             // 最大并发请求数
+    next_seq_id: u64,
+}
+```
+
+### 调度循环（Engine 主循环）
+
+```rust
+loop {
+    // Step 1: 回收已完成的 sequence
+    running.retain(|seq| seq.status != Finished);
+
+    // Step 2: Admit 新请求（如果 running < max_batch_size）
+    while running.len() < max_batch_size {
+        if let Some(seq) = waiting.pop_front() {
+            running.push(seq);
+        } else {
+            break;
+        }
+    }
+
+    if running.is_empty() {
+        // 没有任何工作，等待新请求
+        let new_req = request_rx.recv();  // blocking wait
+        waiting.push_back(new_req);
+        continue;
+    }
+
+    // Step 3: 分类 — 哪些需要 prefill，哪些需要 decode
+    let to_prefill: 新加入的 seq（generated_tokens 为空）
+    let to_decode: 已在运行的 seq
+
+    // Step 4: 执行
+    for seq in to_prefill {
+        // Prefill: 完整 prompt 一次 forward
+        model.forward_gpu_cache(&seq.prompt_tokens, &mut seq.kv_cache);
+        seq.status = Running;
+    }
+
+    // Decode: 每个 seq 独立做一步（当前不做 batch forward，留待优化）
+    for seq in to_decode {
+        let last_token = seq.last_generated_token();
+        let logits = model.forward_gpu_cache(&[last_token], &mut seq.kv_cache);
+        let next = sample_greedy(&logits);
+        seq.generated_tokens.push(next);
+        // 发送 token 给客户端
+        seq.output_tx.blocking_send(Token { id: next, text: decode(next) });
+        // 检查完成
+        if next == eos || seq.generated_tokens.len() >= seq.max_tokens {
+            seq.output_tx.blocking_send(Done);
+            seq.status = Finished;
+        }
+    }
+
+    // Step 5: 检查是否有新请求到达（non-blocking）
+    while let Ok(new_req) = request_rx.try_recv() {
+        waiting.push_back(new_req);
+    }
+}
+```
+
+### 关键设计决策
+
+1. **每个 seq 独立 KV cache**：当前不做 batch forward（需要对齐 seq_len），而是每个 seq 独立调用 model.forward_gpu_cache。未来优化为 batched forward。
+
+2. **Prefill 和 Decode 混合**：新加入的 seq 先 prefill（一次 forward），然后下一轮加入 decode batch。
+
+3. **Non-blocking request receive**：decode 循环中用 `try_recv()` 检查新请求，不阻塞推理。
+
+4. **max_batch_size**：受限于 GPU 显存（每个 seq 的 KV cache 占用）。Qwen3-8B 单卡 32GB，每个 seq 的 KV cache 约 256 tokens × 8 heads × 128 dim × 2(KV) × 2B = 1MB。可以并发 ~100 seq。实际受限于推理速度。
+
+## 与 Phase 13 (HTTP API) 的接口
+
+```
+HTTP Handler                              Engine Thread
+     │                                          │
+     │  ──── GenerateRequest ────────►          │
+     │       (prompt_tokens, max_tokens,        │
+     │        output_tx)                        │
+     │                                          │
+     │  ◄──── GenerateEvent (Token/Done) ────   │
+     │       (via tokio::sync::mpsc)            │
+     │                                          │
+```
+
+多个 HTTP handler 可以同时提交请求。Engine 线程内部通过 Scheduler 管理并发。
+
+## 验收测试
+
+必须通过以下测试才算 Phase 12 完成：
+
+1. **并发 3 请求测试**：同时发 3 个请求，验证 3 个请求同时产出 token（不是串行等待）
+2. **吞吐量测试**：并发请求的总 token 吞吐量应接近单请求（因为单个 seq 的 decode 是串行的）
+3. **动态加入测试**：先发 1 个请求开始生成，过 2 秒再发第 2 个，验证第 2 个立即开始（不等第 1 个完成）
+4. **正确性测试**：并发请求的输出内容应与单独跑每个请求一致
+
+## 实现计划
+
+1. 重构 Engine：从 `while recv → generate` 改为 scheduler loop
+2. 每个 Sequence 持有独立的 GpuKVCache
+3. 调度循环实现 admit + prefill + decode + finish
+4. HTTP API 侧改为 unbounded channel（允许多请求同时提交）
+5. 编写并发测试脚本
+
+## 当前状态
+
+**已实现: iteration-level scheduling**。多请求可以并发进入 batch (max_batch_size)，新请求在 mid-generation 动态加入。Prefill 和 decode 阶段在每轮迭代内分离处理。
+
+**未实现: batched GPU forward**。每个 seq 的 model forward 仍是串行调用 (per-seq forward_gpu_cache)。真正的 batched decode (多 seq 的 token 合并为一次 GPU forward) 需要 Flash Attention 的 variable-length attention 支持。Phase 14 实现了 FA2 kernel，为后续 batched forward 提供了基础。
+
+**验证**: 8 个并发请求 (max_batch=4) 总 wall clock 22.5s，各请求延迟之和 135.0s，调度加速 6.0x。Server log 确认 `decode batch_size=4`。

docs/13-http-api.md

@@ -0,0 +1,133 @@
+# Phase 13: HTTP API + Streaming — Design Document (Milestone ③)
+
+## Goal
+
+提供 OpenAI 兼容的 HTTP API，让 xserv 可以作为一个 serving 后端被任何 OpenAI SDK 调用。
+
+## 职责划分
+
+| 组件 | 职责 |
+|------|------|
+| Phase 12 (Scheduler/Engine) | 模型推理 + 请求调度 + token 生成循环 |
+| **Phase 13 (HTTP API)** | HTTP 请求解析 → 内部格式 → 提交给 engine → 从 channel 接收 token → 编码为 HTTP 响应 |
+
+Phase 13 不关心模型如何推理，只负责 HTTP 协议层。
+
+## 技术栈
+
+- **HTTP framework**: axum 0.8
+- **Async runtime**: tokio
+- **Serialization**: serde_json
+- **Channel**: tokio::sync::mpsc (API ↔ Engine)
+
+## API 端点
+
+```
+GET  /health                    → "ok"
+GET  /v1/models                 → {"data": [{"id": "qwen3-8b", ...}]}
+POST /v1/chat/completions       → JSON response (non-streaming)
+POST /v1/chat/completions       → SSE stream (streaming, TODO)
+```
+
+## Architecture
+
+```
+Client
+  │ HTTP POST /v1/chat/completions
+  ▼
+┌──────────────────────────────┐
+│  axum handler                │
+│  1. Deserialize ChatRequest  │
+│  2. Build prompt text        │
+│  3. Tokenize (Mutex<Tokenizer>)│
+│  4. Create mpsc channel      │
+│  5. Submit GenerateRequest   │
+│  6. await tokens from rx     │
+│  7. Build JSON response      │
+└──────────────────────────────┘
+  │ GenerateRequest via SyncSender
+  ▼
+┌──────────────────────────────┐
+│  Engine thread (Phase 12)    │
+│  - recv() request            │
+│  - model.forward_gpu_cache() │
+│  - blocking_send() tokens    │
+└──────────────────────────────┘
+```
+
+## OpenAI 兼容格式
+
+### Request
+```json
+{
+  "model": "qwen3-8b",
+  "messages": [
+    {"role": "system", "content": "You are helpful."},
+    {"role": "user", "content": "Hello"}
+  ],
+  "max_tokens": 256,
+  "stream": false
+}
+```
+
+### Response (non-streaming)
+```json
+{
+  "id": "chatcmpl-xxx",
+  "object": "chat.completion",
+  "created": 1234567890,
+  "model": "qwen3-8b",
+  "choices": [{
+    "index": 0,
+    "message": {"role": "assistant", "content": "Hi there!"},
+    "finish_reason": "stop"
+  }],
+  "usage": {"prompt_tokens": 5, "completion_tokens": 3, "total_tokens": 8}
+}
+```
+
+### SSE Streaming (TODO)
+```
+data: {"choices":[{"delta":{"content":"Hi"}}]}
+
+data: {"choices":[{"delta":{},"finish_reason":"stop"}]}
+
+data: [DONE]
+```
+
+## 当前实现状态
+
+- [x] `/health` — 健康检查
+- [x] `/v1/models` — 模型列表
+- [x] `/v1/chat/completions` (non-streaming) — JSON response
+- [ ] `/v1/chat/completions` (streaming) — SSE
+- [ ] 完整的 `usage` 统计 (token 计数)
+- [ ] 错误处理 (400 for bad request, etc.)
+- [ ] 多轮对话 chat template
+
+## Key Design Decisions
+
+1. **Extension vs State**: 用 `axum::Extension<Arc<AppState>>` 而不是 `Router::with_state`，因为 `SyncSender` 不是 `Sync`（需要 Mutex 包装）。
+
+2. **Engine 在独立 thread**: GPU 同步操作 block 线程，不能放在 tokio runtime 中。
+
+3. **tokio::sync::mpsc 做 token 传输**: Engine (std thread) 用 `blocking_send()`，API (async) 用 `.recv().await`。跨 async/sync 边界通信。
+
+## Test Plan
+
+- [x] curl /health → "ok"
+- [x] curl /v1/models → JSON model list
+- [x] curl /v1/chat/completions → JSON with generated text
+- [ ] Python OpenAI SDK 兼容性测试
+- [ ] SSE streaming 测试
+- [ ] 多轮对话测试
+
+## Takeaways
+
+1. **axum 0.8 的 Handler trait 对 Send 很严格**：async fn 返回的 Future 必须是 Send。`std::sync::MutexGuard` 不是 Send，必须确保它不活过 await point（用 scope 或显式 drop）。
+
+2. **std::sync::mpsc::SyncSender 不是 Sync**：不能直接放在 `Arc<T>` 中被多个 async task 共享。解决方案：`Mutex<SyncSender>` 或换用 `tokio::sync::mpsc::Sender`（是 Sync 的）。
+
+3. **非 streaming 更简单，先跑通再加 SSE**：SSE streaming 涉及 `Stream` trait、lifetime 问题和复杂的类型推导。先用 collect-all-then-respond 跑通 E2E，streaming 作为增量优化。
+
+4. **Engine 加载时间 ~20s（Qwen3-8B）**：需要在 server 启动后等 engine ready 才接受请求，否则请求会 hang 在 channel send 上。当前靠 sync_channel(1) 的背压天然处理。

docs/14-flash-attention.md

@@ -0,0 +1,167 @@
+# Phase 14: Flash Attention 2 for SM120 — Design Document
+
+## Goal
+
+用自写的 Flash Attention 2 CUDA kernel 替换 naive attention (Phase 5)。消除 O(S²) 显存分配，支持 GQA kernel 内部索引（消除 repeat_kv 开销）。
+
+## 硬件约束: FA4 不适用于 RTX 5090
+
+Flash Attention 已发展到第 4 代 (FA4, arxiv 2603.05451)，但各版本有明确硬件依赖：
+
+| 版本 | 目标架构 | 关键硬件特性 | RTX 5090 (SM120) |
+|------|---------|------------|-----------------|
+| FA2 | 通用 CUDA (SM75+) | shared memory + HMMA | **兼容** |
+| FA3 | Hopper SM90 (H100) | TMA + WGMMA + warp specialization | 不兼容 |
+| FA4 | Blackwell SM100 (B200/B300) | TMEM + async MMA + 2-CTA mode | 不兼容 |
+
+RTX 5090 使用消费级 Blackwell (GB202, SM120)，与数据中心 Blackwell (B200, SM100) 是不同硅片。SM120 **没有 TMEM (Tensor Memory)**，这是 FA4 kernel 设计的核心硬件依赖。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 + 标准 HMMA)。
+
+## Naive Attention 的问题
+
+Phase 5 的 naive attention 流程:
+```
+k_t = K.transpose(2,3).contiguous()    ← 分配 K^T 显存
+scores = batched_matmul(Q, k_t)         ← 分配 [B,H,S,S] score 矩阵 (O(S²) 显存)
+scores = scale(scores, 1/sqrt(d))       ← 逐元素 kernel
+causal_mask(scores)                     ← 逐元素 kernel
+weights = softmax(scores)               ← 分配 [B,H,S,S] weight 矩阵
+output = batched_matmul(weights, V)     ← 最终结果
+```
+
+问题:
+1. **显存 O(S²)**: score 和 weight 矩阵各需 `B × H × S × S × dtype_size`。S=2048, H=32, BF16 → 256 MB。S=8192 → 4 GB。
+2. **GQA 预处理**: 在调用 attention 前需要 `repeat_kv_gpu` 将 K/V 从 8 heads 扩展到 32 heads，每层额外分配和拷贝。
+3. **多次 kernel launch**: scale, mask, softmax 各一次 kernel launch + global memory round-trip。
+4. **K^T materialization**: `K.transpose().contiguous()` 需要分配和拷贝。
+
+## FA2 算法
+
+核心思想: **不 materialize S×S 矩阵**。将 Q, K, V 分成 tiles，在 shared memory (SRAM) 中计算，使用 **online softmax trick** 边算边更新 running max 和 sum。
+
+FA2 (Dao 2023) 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，提高并行性。
+
+```
+scale = 1 / sqrt(head_dim)
+
+for each Q tile (q_start..q_start + BR):                   // 外层: Q tiles
+    load Q_tile [BR, D] to shared memory (一次加载，内层复用)
+    init per-row: O[D] = 0, m = -inf, l = 0
+
+    for each K/V tile j (kv_start..kv_start + BC):          // 内层: K/V tiles
+        // Causal tile-skip: 如果整个 K tile 在 Q tile "未来"，跳过
+        if causal && kv_start > max_q_pos + kv_offset: skip
+
+        load K_tile [BC, D] to shared memory
+        S = Q_tile @ K_tile^T * scale                       // [BR, BC], in registers
+        if causal: mask S[r][c] = -inf where kv_pos > q_pos
+
+        // Online softmax update
+        m_new = max(m, rowmax(S))
+        P = exp(S - m_new)
+        l_new = exp(m - m_new) * l + rowsum(P)
+        O = exp(m - m_new) * O                              // rescale accumulator
+
+        load V_tile [BC, D] to shared memory (复用 K 的空间)
+        O += P @ V_tile                                     // accumulate
+
+        m = m_new, l = l_new
+
+    O = O / l                                               // final normalize
+    write O[BR, D] to HBM (convert FP32 → BF16)
+```
+
+## 实现细节
+
+### Kernel 配置
+
+| 参数 | 值 | 说明 |
+|------|---|------|
+| BR (Q tile rows) | 64 | Q tile 大小 |
+| BC (K/V tile rows) | 64 | K/V tile 大小 |
+| head_dim | 运行时参数 (≤128) | 支持 64 (GPT-2) 和 128 (Qwen3) |
+| Block size | 128 threads | 64 线程各 own 一行 Q，其余协助加载 |
+| Grid | (q_tiles, batch × num_q_heads) | 每个 block 处理一个 Q tile + 一个 head |
+
+### Shared Memory (BF16 存储)
+
+```
+smem_q [BR × head_dim] BF16 = 64 × 128 × 2 = 16 KB (加载一次，内层复用)
+smem_kv[BC × head_dim] BF16 = 64 × 128 × 2 = 16 KB (K 和 V 交替使用)
+────────────────────────────────────────────
+Total: 32 KB (SM120 默认 48 KB，余量充足)
+```
+
+### 线程映射
+
+- Thread 0..63: 各 own Q_tile 的一行。负责该行的全部计算：dot products、softmax、PV 累加。
+- Thread 64..127: 协助 shared memory 加载 (K/V tile)，不参与计算。
+- 加载模式: 每个 thread 加载 `(BR × head_dim) / 128 = 64` 个 BF16 元素。
+
+### Per-Thread Register 使用
+
+```
+O_acc[128] FP32     = 512 bytes (128 regs)  — 输出累加器
+P[64]      FP32     = 256 bytes (64 regs)   — 当前 tile 的 softmax 后权重
+m, l       FP32     = 8 bytes (2 regs)      — online softmax running state
+循环变量 + 临时      ≈ 16 regs
+────────────────────────────────────────────
+Total: ~210 regs/thread (max 255，在限制内)
+```
+
+### GQA 支持
+
+每个 thread block 处理一个 Q head，通过 `kv_head = q_head / (num_q_heads / num_kv_heads)` 映射到对应的 KV head。K/V 的数据指针直接指向 KV head 的存储，无需 repeat_kv。
+
+```
+// 32 Q heads, 8 KV heads → heads_per_group = 4
+// Q head 0,1,2,3 → KV head 0
+// Q head 4,5,6,7 → KV head 1
+// ...
+kv_head = q_head / heads_per_group;
+K_ptr = K + (batch * num_kv_heads + kv_head) * kv_len * head_dim;
+```
+
+### Causal Mask
+
+两级优化:
+1. **Tile-level skip**: 如果 `kv_tile_start > max_q_pos + kv_offset`，整个 K/V tile 都在未来，跳过（减少 ~50% 计算）。
+2. **Element-level mask**: 在 tile 内部，`if kv_pos > q_pos + kv_offset: S = -inf`。
+
+`kv_offset = kv_len - q_len` 处理 decode 时 KV cache 长于 Q 的情况。
+
+## 与 Naive Attention 的对比
+
+| 特性 | Naive (Phase 5) | FA2 (Phase 14) |
+|------|----------------|----------------|
+| 显存 | O(B × H × S²) | O(B × H × S × D) |
+| GQA | 需要 repeat_kv (分配+拷贝) | Kernel 内部索引 (零开销) |
+| K^T | 需要 transpose+contiguous | Kernel 内部计算 |
+| Kernel launches | 6 (matmul, scale, mask, softmax, matmul, ...) | 1 (单个 fused kernel) |
+| S=8192 可行性 | OOM (~4 GB score matrix) | 可行 (32 KB shared memory) |
+
+## 源码结构
+
+```
+csrc/attention/flash_attention.cu     — FA2 kernel (BF16 in, FP32 accumulate, BF16 out)
+crates/xserv-kernels/src/attention.rs — flash_attention() Rust wrapper + 原 attention() 保留
+crates/xserv-model/src/qwen3.rs       — forward_gpu_cache 调用 flash_attention
+```
+
+## 已知局限与后续优化方向
+
+1. **Decode (Q_len=1) 效率低**: BR=64 线程中只有 1 个 active（owns_row）。应写专用 decode attention kernel，沿 KV 维度 parallel reduction。
+2. **无向量化加载**: 当前逐元素 bf16→f32 转换，应改用 `float4` 或 `__nv_bfloat162` 批量加载。
+3. **Register tiling**: 每个 thread 目前串行计算 dot product (128 MADs per K column)。可改为多线程协作。
+4. **K/V double buffering**: 可在计算当前 tile 时预加载下一个 tile 到另一半 shared memory。
+5. **Tile size 调优**: 更大的 tile (BR=128) 可能在长 sequence 时更优，需要 opt-in shared memory。
+
+## Test Plan
+
+- [x] 正确性: logits 与 HF transformers 对比 (top-1 match 9/10, top-5 overlap 4.0/5)
+- [x] 生成质量: 52/52 prompt 生成连贯文本，中英文均可
+- [x] SSE streaming 正常工作
+- [x] 性能: 12.9 tok/s (vs naive 10.3 tok/s, +25%)
+- [ ] 长 sequence (S=4096, S=8192): 验证 naive OOM 而 FA2 正常
+- [ ] ncu profile: compute utilization, memory throughput

docs/15-performance.md

@@ -0,0 +1,177 @@
+# Phase 15: Performance Optimization — Design Document (Milestone ④)
+
+## Goal
+
+系统性 profiling + 优化，从 12.9 tok/s (Phase 14 结束) 逼近 RTX 5090 的理论带宽上限 (112 tok/s)。
+
+## 硬件 Roofline
+
+RTX 5090 (SM120, CC 12.0) 的 decode 理论极限：
+
+```
+模型权重:          16 GB (Qwen3-8B BF16)
+内存带宽:          1.79 TB/s (GDDR7)
+理论最优 decode:   16 GB / 1.79 TB/s = 8.9 ms/step = 112 tok/s (batch=1)
+```
+
+Decode 阶段 100% memory-bound：每步读取全部 16 GB 权重（252 个 GEMV），计算量可忽略。
+
+## 瓶颈分析
+
+Phase 14 结束时性能 12.9 tok/s = 77.5 ms/step，roofline 利用率仅 12%。
+
+### 量化瓶颈分解
+
+| 来源 | 估计耗时 | 占比 |
+|------|---------|------|
+| cuBLAS M=1 GEMV (252 calls, 带宽利用 ~8%) | ~60 ms | 77% |
+| 非 matmul 内核 (attention, norm, activation, reshape) | ~8 ms | 10% |
+| Tensor 分配 + cudaMemset (1440+ allocs/step) | ~5 ms | 7% |
+| Kernel launch overhead (200+ launches × 5μs) | ~1 ms | 1% |
+| 其他 (sampling CPU round-trip, etc.) | ~3.5 ms | 5% |
+
+**核心发现: cuBLAS 对 M=1 GEMM (GEMV) 的带宽利用率极低（~8%），是 9x gap 的根本原因。**
+
+cuBLAS 设计用于大 M 的 GEMM，对 M=1 场景存在：
+- Kernel launch dispatch overhead 无法被大量计算掩盖
+- TensorCore tile (16×16) 无法被 M=1 充分利用
+- 内部 heuristic 选择了次优算法
+
+## 优化实施
+
+### Opt 1: Decode Attention Kernel
+
+**目标**: 替换 FA2 在 Q_len=1 时的低效路径（64 线程仅 1 个 active）。
+
+**实现** (`csrc/attention/flash_attention.cu`):
+- 专用 decode_attention_bf16_kernel: 256 线程并行沿 KV 序列维度
+- 每个 thread 加载完整 Q vector (128 dim) 到寄存器
+- 处理其分配的 KV 位置块: dot product → online softmax
+- Block-level warp-shuffle + shared memory reduction 合并结果
+- GQA 支持: kv_head = q_head / heads_per_group
+
+**效果**: 在当前短序列 (kv_len ≤ 79) 下效果微小——attention 不是瓶颈。在长序列时会显著受益。
+
+### Opt 2: Fused SiLU×Mul
+
+**目标**: `silu(gate) * up` 两个 element-wise op 合并为一个 kernel。
+
+**实现** (`csrc/activation/activations.cu`):
+```
+Before: read gate → silu → write temp → read temp + up → mul → write out
+After:  read gate + up → silu(gate) * up → write out
+Saved:  1 HBM read + 1 HBM write per element
+```
+
+**效果**: 每层省 1 次 HBM round-trip，36 层总计可观但在 GEMV 瓶颈下被掩盖。
+
+### Opt 3: Fused Add+RMSNorm
+
+**目标**: `x = residual + attn_proj; normed = rmsnorm(x)` 合并为一个 kernel。
+
+**实现** (`csrc/normalization/rmsnorm.cu`):
+```
+Before: read residual + x → add → write sum → read sum + gamma → norm → write out
+After:  read residual + x + gamma → add + norm → write sum + normed
+Saved:  1 full HBM round-trip per attention block
+```
+
+### Opt 4: Batched Decode Forward ⭐
+
+**目标**: 多序列 decode token 合并为 M=batch_size 的 GEMM，提升 cuBLAS 效率。
+
+**实现** (`crates/xserv-model/src/qwen3.rs` + `crates/xserv-server/src/engine.rs`):
+- 新增 `Qwen3::forward_decode_batch(tokens, positions, caches)`
+- Batched ops: embedding, norm, projections, FFN — [B, hidden] × [hidden, X]
+- Per-seq ops: RoPE, KV cache, attention（各序列位置/长度不同）
+- Row extraction (`row_view`) + concatenation (`concat_rows`) 在 batched/per-seq 间切换
+- Engine Step 4b: batch≥2 时自动使用 batched decode
+
+**效果**: batch=4 时 cuBLAS 从 1008× M=1 → 252× M=4，吞吐 35.1 tok/s (vs serial 13.2)。
+
+### Opt 5: Custom GEMV Kernel ⭐⭐⭐ (决定性优化)
+
+**目标**: 替换 cuBLAS 的 M=1 GEMV，手写带宽最优化 kernel。
+
+**实现** (`csrc/gemm/gemv.cu`):
+```
+设计: K-split tiled GEMV
+- TILE_N = 128 (output columns per block, one thread per column)
+- TILE_K = 256 (K-dimension slice per block)
+- BLOCK_SIZE = 128 threads
+- Grid: (ceil(N/128), ceil(K/256)) — 对 K=N=4096 得到 512 blocks
+  512 blocks / 170 SMs ≈ 3 blocks/SM (良好 occupancy)
+
+内存访问:
+- 相邻线程读 W 矩阵的相邻列 → 完美 coalesced
+- x vector 加载到 shared memory (每 K-chunk 仅加载一次)
+- FP32 accumulation via atomicAdd (K-split partial sums)
+- 独立 kernel 做 FP32→BF16 转换
+
+调度:
+- matmul() 中检测 M==1 && dtype==BF16 → 自动使用 custom GEMV
+- M>1 保持 cuBLAS
+```
+
+**效果**: 13.2 → 46.6 tok/s (+253%)。带宽利用率从 ~8% 提升到 ~42%。
+
+### Opt 6: Tensor::empty() (消除无用 cudaMemset)
+
+**目标**: kernel 输出 tensor 全量覆写时，跳过分配后的 cudaMemset 清零。
+
+**实现**:
+- `Storage::empty()` + `Tensor::empty()`: 分配不清零
+- 21 个 kernel wrapper (activation, attention, embedding, gemm, norm, softmax, transpose) 从 `zeros` 改为 `empty`
+- GEMV FP32 accumulator buffer 保持 `cudaMemsetAsync`（atomicAdd 需要零初始化）
+
+**效果**: 46.6 → 50.3 tok/s (+8%)。消除 ~756 个 cudaMemset/step。
+
+### Infra: CUDA Graph 基础设施
+
+- FFI bindings: `cudaStreamBeginCapture`, `cudaGraphInstantiate`, `cudaGraphLaunch`
+- RAII wrapper: `CudaGraph` (capture/instantiate/launch lifecycle)
+- 当前未在 forward path 使用（variable kv_len 限制），为后续优化预留
+
+## Ablation 结果
+
+dash5, RTX 5090, Qwen3-8B BF16, greedy decode, max_tokens=64:
+
+| 优化叠加 | tok/s | 增量 | vs HF | Roofline |
+|---------|-------|------|-------|----------|
+| Phase 14 baseline (FA2) | 12.9 | — | 36% | 12% |
+| + Decode attention | 12.9 | +0% | 36% | 12% |
+| + Fused SiLU×Mul | 13.0 | +1% | 36% | 12% |
+| + Fused Add+RMSNorm | 13.2 | +2% | 37% | 12% |
+| + Batched decode (batch=4) | 35.1 | — | 97% | — |
+| + Custom GEMV (M=1) | 46.6 | +253% | 130% | 42% |
+| + Tensor::empty | **50.3** | +8% | **140%** | **45%** |
+
+对比:
+
+| 系统 | tok/s | Roofline |
+|------|-------|----------|
+| HF transformers | 36.0 | 32% |
+| **xserv (Phase 15)** | **50.3** | **45%** |
+| 理论极限 (1.79 TB/s) | 112.0 | 100% |
+
+## 剩余 55% Roofline Gap 分析
+
+| 来源 | 估计占比 | 优化方向 |
+|------|---------|---------|
+| GEMV kernel 非满带宽 (atomicAdd contention, K-split overhead) | 25% | 无 K-split GEMV (更大 block), 向量化加载 |
+| Non-matmul kernels (attention, norm, RoPE, reshape) | 15% | Fused layer kernel, 更高效的 decode attention |
+| Kernel launch overhead (200+ launches/step) | 5% | CUDA Graphs (需解决 variable kv_len) |
+| Memory allocator overhead (Arc, SmallVec per tensor) | 5% | Pre-allocated decode workspace |
+| Sampling D2H copy (pipeline stall) | 3% | GPU-side argmax kernel |
+| 其他 (host-side logic, channel overhead) | 2% | — |
+
+## 下一步
+
+Phase 15 的 Milestone ④ 目标 (50% of HF) 已远超 — 达到 140% of HF, 45% of roofline。
+
+后续优化路径（按 ROI 排序）：
+1. **无 K-split GEMV**: 消除 atomicAdd，减少 kernel launches → 预期 +15-20%
+2. **向量化 GEMV loads**: float4 加载 W 矩阵 → 预期 +10%
+3. **Pre-allocated workspace**: 消除 Tensor 对象分配开销 → 预期 +5%
+4. **CUDA Graphs**: 需要 fixed-shape decode path → 预期 +5%
+5. **GPU-side sampling**: 消除 logits D2H pipeline stall → 预期 +3%

docs/16-llama-cpp-comparison.md

@@ -0,0 +1,201 @@
+# Phase 16: llama.cpp Comparison Baseline
+
+> **Goal.** Replace HF transformers with **llama.cpp** as the standing
+> performance baseline, and add a standard quality (response correctness)
+> benchmark suite (AIME 2025, GSM8K). Provide a one-click entrypoint that runs
+> both systems under identical workloads and emits a side-by-side report.
+
+## Motivation
+
+xserv has cleared 140% of HF transformers throughput on Qwen3-8B (Phase 15).
+HF is no longer a useful performance bar — it's a *correctness* baseline.
+
+**llama.cpp** is the right next bar because:
+- It's a serious C++/CUDA inference engine with active optimization
+- Same OpenAI-compatible API → black-box, fair comparison
+- Same GGUF↔safetensors weight source (we convert BF16, no quantization shortcuts)
+- Used widely as a reference point in the community
+
+We also need **quality benchmarks** so that performance improvements don't
+silently regress model quality (numerical precision, sampling, prompt
+formatting). AIME and GSM8K are the cheapest credible signals.
+
+## Architecture
+
+```
+xserv/
+├── third_party/llama.cpp/         # cloned by setup-llama-cpp.sh
+│   └── build/bin/llama-server     # CUDA build (SM120)
+├── tools/
+│   ├── setup-llama-cpp.sh         # clone + cmake build (idempotent)
+│   ├── convert-to-gguf.sh         # safetensors → BF16 GGUF (same weights)
+│   ├── sync-and-build.sh          # extended with `bench` subcommand
+│   └── bench/                     # Python benchmark driver
+│       ├── runner.py              # entrypoint
+│       ├── servers.py             # subprocess lifecycle (start/stop both)
+│       ├── client.py              # OpenAI streaming client + TTFT/TPOT
+│       ├── speed.py               # speed suite
+│       ├── quality.py             # quality suite
+│       ├── tasks/{aime,gsm8k}.py  # dataset loaders + scorers
+│       ├── report.py              # markdown + json output
+│       └── requirements.txt       # httpx, datasets
+└── bench-out/                     # report artifacts (gitignored)
+    ├── comparison-<stamp>.md
+    ├── comparison-<stamp>.json
+    └── logs/{xserv,llama_cpp}.log
+```
+
+Both systems are treated as **black-box HTTP servers** speaking the OpenAI
+streaming chat API. No in-process integration, no shared Python bindings. This
+keeps the comparison fair (same protocol, same prompt-template path) and
+isolates the test harness from internal API churn on either side.
+
+## Workflow
+
+The GPU host (dash5) has **no outbound network and no rsync**, so anything from
+the internet is fetched locally and shipped over via tar-over-ssh.
+
+```
+local repo (has network)              dash5 (GPU host, no network)
+────────────────────────              ────────────────────────────
+# one-time, on a networked machine:
+python3 -m tools.bench.fetch_datasets  →  tools/bench/data/{aime2025,gsm8k}.json
+git submodule update --init …          →  third_party/llama.cpp source
+
+tools/sync-and-build.sh bench   →  tar project   (excl. target, third_party, bench-out)
+                                →  tar llama.cpp source (excl. build, .git)
+                                →  setup-llama-cpp.sh   (build-only; no-op if built)
+                                →  convert-to-gguf.sh   (no-op if .gguf exists)
+                                →  cargo build --release
+                                →  python3 -m tools.bench.runner ...
+                                →  bench-out/comparison-<stamp>.md
+tools/sync-and-build.sh fetch-bench-out  ←  tar bench-out back
+```
+
+Behind a flaky proxy, fetch datasets through the HF mirror:
+`HF_ENDPOINT=https://hf-mirror.com python3 -m tools.bench.fetch_datasets`.
+
+`tools/__init__.py` exists so `python3 -m tools.bench.runner` resolves our
+package: some site-packages (e.g. nvfuser) ship a regular top-level `tools`
+package that would otherwise shadow a namespace `tools`.
+
+## What gets measured
+
+### Speed (TTFT / TPOT / throughput)
+
+- **Single-stream**, three prompt lengths (short / medium / long), `cfg.speed_prompts` repeats each
+  - `TTFT p50/p95`, `TPOT p50/p95`, per-request throughput
+- **Concurrent**, fixed medium prompt, sweep `concurrency ∈ {1, 2, 4, 8}`
+  - Aggregate `tok/s`, `TTFT p95`, error count
+- Both at `temperature=0`, `max_tokens=128` by default.
+
+### Quality (response correctness)
+
+| Task | N | Source | Scoring | Why |
+|---|---|---|---|---|
+| AIME 2025 | 30 | `MathArena/aime_2025`, fallback `yentinglin/aime_2025` (HF) | exact-match boxed integer (0..999) | reasoning + math, hard signal |
+| GSM8K | 1319 | `openai/gsm8k` (HF), `test` split | exact-match `\boxed{n}` or last number | broad sanity, decimals allowed |
+
+Same `temperature=0` sampling across both systems. Max tokens: 16384 for AIME
+(reasoning long), 2048 for GSM8K. Subsample with `--quality-limit N` for smoke.
+
+**Generation mode must match.** xserv's prompt builder hardcodes Qwen3 thinking
+OFF (it appends an empty `<think></think>` block). llama-server applies the
+GGUF's Qwen3 jinja template, which has thinking ON by default. The driver
+therefore sends `chat_template_kwargs={"enable_thinking": false}` to llama.cpp
+so both engines run the model in the same mode. Pass `--enable-thinking` to
+compare in thinking mode instead (xserv would need a matching change first).
+
+### Report
+
+`bench-out/comparison-<stamp>.md` contains:
+- Environment (GPU, driver, xserv commit, python)
+- Speed table per scenario (xserv | llama.cpp | xserv÷llama.cpp speedup)
+- Quality table per task (n, correct, accuracy, mean tokens, TTFT, TPOT, wall)
+
+A sibling `.json` holds all per-request raw rows and per-problem case detail
+(prediction, gold, response preview) so we can diff regressions in CI later.
+
+## Running it
+
+**One-time prerequisites (on a networked machine):**
+```bash
+git submodule update --init third_party/llama.cpp     # pinned to b9371
+HF_ENDPOINT=https://hf-mirror.com python3 -m tools.bench.fetch_datasets
+```
+
+**Full sweep on dash5 (recommended):**
+```bash
+./tools/sync-and-build.sh bench -- --max-seq-len 8192 --quality-limit 50
+./tools/sync-and-build.sh fetch-bench-out
+open bench-out/comparison-*.md
+```
+
+**Speed-only smoke (fast):**
+```bash
+./tools/sync-and-build.sh bench -- --suite speed --speed-prompts 2
+```
+
+**Quality smoke with 5 problems each:**
+```bash
+./tools/sync-and-build.sh bench -- --suite quality --quality-limit 5
+```
+
+**On a host that already has both servers running** (e.g. local dev with two
+shells open):
+```bash
+python3 -m tools.bench.runner \
+    --xserv-base-url http://127.0.0.1:8080 \
+    --llama-base-url http://127.0.0.1:8081 \
+    --suite all
+```
+
+## Design choices
+
+1. **Black-box HTTP, not FFI.** Both engines bind the same OpenAI surface and
+   real serving traffic uses HTTP. Anything that doesn't show up over the wire
+   doesn't matter for serving.
+2. **Same BF16 weights.** We convert the same safetensors with llama.cpp's
+   `convert_hf_to_gguf.py --outtype bf16`. No quantization at this stage; if we
+   want a quant comparison later we'll add a separate column, not replace this
+   one.
+3. **Streaming everywhere.** TTFT and TPOT only make sense with streaming. We
+   ask both servers for `stream=true` with `include_usage` so we can read
+   server-reported token counts when available.
+4. **Idempotent setup.** `setup-llama-cpp.sh` and `convert-to-gguf.sh` are
+   safe to re-run — they no-op when the build / file already exists. The
+   `bench` subcommand wires them so the first run does a full setup and
+   subsequent runs are fast.
+5. **Subprocess lifecycle owned by the driver.** We spawn each server in its
+   own process group and SIGTERM the group on exit so half-dead llama-server
+   children don't survive. If the user is already running a server somewhere,
+   pass `--xserv-base-url` / `--llama-base-url` to skip launch.
+6. **One server at a time.** The driver starts a system, runs every suite
+   against it, stops it, then moves to the next. Two BF16 8B models (~16GB each)
+   do not co-reside on a single 32GB GPU, and a resident idle engine would
+   distort the other's latency/throughput. This serialization is why the report
+   is assembled from per-system passes rather than a single interleaved run.
+
+## Known constraints / findings
+
+- **xserv OOM at `--max-seq-len 8192` — fixed.** xserv used to pre-allocate its
+  paged-KV pool (`total_blocks = blocks_per_seq · max_batch · 2`, ≈9GB at 8192)
+  on top of the 16GB weights, exceeding 32GB at startup (`paged_kv_cache.rs`
+  `alloc paged K pool: OutOfMemory`). Now the pool is sized to *available VRAM*
+  (`cudaMemGetInfo`) and overflow is swapped to pinned host memory (vLLM-style
+  preemption, `--swap-space-gb`). The 8192 comparison runs cleanly with 0 swap
+  events; swap is verified separately under a forced-small pool. The benchmark
+  surfaced the OOM — a good example of the baseline doing its job.
+- When the xserv engine thread dies, the API now returns a clean 503 (the
+  request handler uses a poison-tolerant lock instead of cascading
+  mutex-poison panics). The driver records any failure as a per-request error,
+  so a broken engine shows up as `errs=N` / `accuracy 0%` rather than a hung run.
+
+## Future extensions
+
+- Add quant runs (Q8_0, Q4_K_M) as separate "system" columns
+- Wire to GitHub Actions for nightly regression
+- Track results across commits to flag regressions (per-commit JSON in
+  `docs/benchmarks/history/`)
+- Add MMLU-Pro / HumanEval when budget allows
+- Long-context benchmark (8K, 32K prompts) to compare prefill scaling

docs/17-tensor-parallelism.md

@@ -0,0 +1,122 @@
+# Phase 17: Tensor Parallelism (TP)
+
+> 目标：在单机多卡上做 **张量并行**，把 Qwen3-8B 的权重、计算和 KV cache 按
+> head / 中间维切分到 TP 个 GPU 上，用 AllReduce 聚合，降低单卡显存压力并提升吞吐。
+> 先做 **TP=2 / 4（组内）**，跳过投机解码（原 Phase 16）。
+
+## 1. 硬件约束（dash5）
+
+- 8× RTX 5090（32GB，SM120），**无 NVLink**，纯 PCIe Gen5。
+- 拓扑：GPU 0–3 一组、4–7 一组，组内 `PHB`（同 host bridge，可 P2P），跨组 `NODE`。
+- **TP 必须在组内**（0–3 或 4–7），否则 AllReduce 走跨组 PCIe，延迟更高。
+- AllReduce 带宽受限于 PCIe（~单向 64GB/s），远低于 NVLink；通信会是 decode 的主要开销。
+
+## 2. 切分方案（Megatron-style）
+
+Qwen3-8B：`hidden=4096`、`num_heads=32`、`num_kv_heads=8`、`head_dim=128`、
+`intermediate=12288`、`layers=36`、`vocab=151936`。TP=2/4/8 都能整除 32/8/12288。
+
+每个 transformer block 的切分（设 world size = `T`，本 rank = `r`）：
+
+### Attention（column → row）
+| 权重 | 原 shape (已转置) | 切分 | 每 rank shape |
+|------|-------------------|------|---------------|
+| `q_proj_wt` | `[hidden, num_heads·head_dim]` | column（按 Q head） | `[hidden, (num_heads/T)·head_dim]` |
+| `k_proj_wt` | `[hidden, num_kv_heads·head_dim]` | column（按 KV head） | `[hidden, (num_kv_heads/T)·head_dim]` |
+| `v_proj_wt` | 同上 | column | 同上 |
+| `o_proj_wt` | `[num_heads·head_dim, hidden]` | **row** | `[(num_heads/T)·head_dim, hidden]` |
+
+- 每个 rank 只算自己的 `num_heads/T` 个 Q head 和对应的 `num_kv_heads/T` 个 KV head；
+  GQA 的 `n_rep = num_heads/num_kv_heads = 4` 在每个 rank 内保持不变。
+- `q_norm`/`k_norm`（`[head_dim]`）逐 head 应用，**复制**到每个 rank。
+- RoPE 逐 head、按 position 应用，每个 rank 独立做。
+- **KV cache 也切分**：每个 rank 的 paged KV 只存自己的 `num_kv_heads/T` 个 head
+  → 每卡 KV 显存降为 1/T（TP 的一大收益）。
+- `attn = merge_heads(...) @ o_proj_wt` 得到**部分** `[T_tok, hidden]` → **AllReduce(sum)** → 完整。
+
+### MLP / SwiGLU（column → row）
+| 权重 | 原 shape | 切分 | 每 rank shape |
+|------|----------|------|---------------|
+| `gate_proj_wt` | `[hidden, intermediate]` | column | `[hidden, intermediate/T]` |
+| `up_proj_wt` | 同上 | column | 同上 |
+| `down_proj_wt` | `[intermediate, hidden]` | **row** | `[intermediate/T, hidden]` |
+
+- `silu(gate)*up` 在切分后的 `[T_tok, intermediate/T]` 上逐元素做，无需通信。
+- `down = (...) @ down_proj_wt` 得到部分 `[T_tok, hidden]` → **AllReduce(sum)** → 完整。
+
+### 复制（不切分）
+- 所有 RMSNorm 权重（`input_norm`/`post_norm`/最终 `norm`）：每个 rank 在 AllReduce 后
+  拿到完整 hidden，本地用复制的权重归一化。
+- **第一版**：`embed_tokens` 和 `lm_head` 复制（各 ~1.2GB）。
+  后续优化：vocab-parallel embedding（local lookup + AllReduce）、column-parallel lm_head + AllGather。
+
+### 通信点
+每层 **2 次 AllReduce**（o_proj 后、down_proj 后）→ 每生成 1 token 共 `2·36 = 72` 次。
+decode 时每次 AllReduce 张量是 `[batch, 4096]` BF16（单 token batch=1 时 8KB），**延迟主导**。
+
+## 3. 进程 / 线程模型
+
+**单进程、多线程**：每个 TP rank 一个 OS 线程，线程启动时 `cudaSetDevice(rank_device)` 并绑定。
+
+选择理由：
+- xserv 的 caching allocator 是 `thread_local`，每线程独立池 → 天然契合「一线程一卡一池」。
+- CUDA context 隐式按 device/thread 管理；线程内只 set 一次 device、不再切换即可。
+- HTTP server / 调度器仍在主线程（rank 0 协调），无需多进程 IPC，改动最小。
+- 单机 8 卡足够；多进程（torchrun 式）留待真正跨节点时再说。
+
+执行流：主线程调度器准备一个 step 的输入（tokens/positions/slots），广播给 `T` 个 rank 线程；
+每个 rank 线程跑自己的分片 forward（含层内 AllReduce），rank 0 拿到完整 logits 后采样。
+用 barrier / channel 同步每个 step。
+
+## 4. 通信库：NCCL
+
+用 **NCCL**（dash5 已装：`/usr/lib/x86_64-linux-gnu/libnccl.so.2`，`/usr/include/nccl.h`）。
+
+- 新建 crate `xserv-distributed`：NCCL FFI（`ncclGetUniqueId`、`ncclCommInitRank`、
+  `ncclAllReduce`、`ncclGroupStart/End`）+ `TpContext`（rank/world/comm/stream）+
+  `all_reduce_sum(&mut GpuBuffer)` 原语。
+- NCCL 多线程模式：主线程生成 `ncclUniqueId`，各 rank 线程用 `ncclCommInitRank(comm, world, id, rank)`
+  初始化（需 `ncclGroupStart/End` 包裹并发 init）。
+- AllReduce 用 BF16（`ncclBfloat16`）+ `ncclSum`，在每个 rank 自己的 stream 上。
+
+> **决策点**：collective 用 NCCL 还是自己手写 P2P ring/tree AllReduce？
+> 本项目是「从零构建」，但 collective 属于基础设施（类比我们也用 cuBLAS 作为可切换后端）。
+> 推荐先用 NCCL 把 TP 跑通、拿到正确性与加速比，后续可选做手写 AllReduce 作为学习项。
+
+## 5. 权重分片加载
+
+每个 rank 只加载/保留自己的分片，省显存：
+- column-parallel 权重：按输出维切片取本 rank 的 `[*, dim/T]` 段。
+- row-parallel 权重：按输入维切片取本 rank 的 `[dim/T, *]` 段。
+- 复制权重（norm/embed/lm_head）：每个 rank 各留一份。
+
+实现：`loader` 读 safetensors（mmap）时按 rank 只搬运需要的切片到该 rank 的 GPU；
+`Qwen3::from_weights_tp(config, weights, rank, world)` 在转置/切分时按 rank 取段。
+
+## 6. 实现步骤（逐步可验证）
+
+1. **P17.1 — `xserv-distributed` 基础**：NCCL FFI + `TpContext` + `all_reduce_sum`。
+   验收：2 卡各放一个已知向量，AllReduce 后两卡结果都等于和。
+2. **P17.2 — 分片权重加载**：`from_weights_tp`，每 rank 只持有自己的分片。
+   验收：各 rank 权重 shape 正确、显存占用约为 1/T（+ 复制项）。
+3. **P17.3 — TP forward**：rank 内 attention/MLP + 层内 AllReduce。
+   验收：**TP=2 的 logits 与 TP=1 在 BF16 容差内一致**（top-1 一致，top-5 重合）。
+4. **P17.4 — 接入 engine/server**：`--tp N`，多线程 rank workers + rank0 调度。
+   验收：`--tp 2` 端到端可服务；用现有 llama.cpp bench 跑正确性；
+   测 TP=2 vs TP=1 的吞吐 / TTFT / 每卡显存。
+
+## 7. 预期与风险
+
+- **显存**：每卡权重 + KV 降到约 1/T（embed/lm_head 暂复制）。TP=2 时单卡 ~8GB 权重 + 更大 KV/并发空间。
+- **吞吐**：PCIe AllReduce 延迟会吃掉部分收益；decode 是延迟敏感的，72 次小 AllReduce/token
+  可能让 TP=2 的单流 tok/s **不一定线性提升**，但能跑更大 batch / 更长 context。先测实测数。
+- **风险**：NCCL 多线程初始化的同步、每 rank stream 与现有 kernel stream 的协调、
+  KV cache 按 rank 切 head 后 paged kernel 的 head 维参数要用 per-rank 值。
+- 正确性优先：先 TP=1 等价（logits 对齐），再谈性能。
+
+## 8. 不在本阶段范围
+
+- 跨组 TP=8、Pipeline Parallelism、多节点。
+- vocab-parallel embedding / lm_head（先复制）。
+- 手写 AllReduce（NCCL 跑通后可选）。
+- 与 CUDA Graph decode 的结合（先走非 graph 路径）。

docs/18-pipeline-parallelism.md

@@ -0,0 +1,151 @@
+# Phase 18: Pipeline Parallelism (PP)
+
+> 目标：在单机多卡上做 **流水线并行**，把 Qwen3-8B 的 **层** 切成 `P` 段（stage），
+> 每张卡只持有连续的一段层（+ stage0 的 `embed_tokens`、最后一段的 `norm`/`lm_head`），
+> 激活（hidden state）在相邻 stage 之间用 **NCCL P2P send/recv** 传递。
+> 与 TP（按 head / 中间维切，每层 2 次 AllReduce）互补：PP 通信量小（每 token 仅 `P-1`
+> 次点对点传 `[tokens, hidden]`），KV 与权重按 **层** 降到约 1/P。
+> 先做 **PP=2 / 4（组内）**，正确性优先。
+
+## 1. 硬件约束（dash5）
+
+- 8× RTX 5090（32GB，SM120），**无 NVLink**，纯 PCIe Gen5。
+- 拓扑：GPU 0–3 一组、4–7 一组，组内 `PHB`（同 host bridge，可 P2P），跨组 `NODE`。
+- **PP 同样建议在组内**（0–3 或 4–7）：虽然 PP 的通信量远小于 TP，但 P2P 仍走 PCIe，
+  跨组延迟更高。PP=2/4 用 0–1 / 0–3。
+- 相比 TP：TP 每 token `2·layers = 72` 次 AllReduce（延迟主导）；PP 每 token 仅
+  `P-1` 次 send/recv，每次 `[tokens, hidden]` BF16（decode batch=1 时 8KB）。
+  **PP 对慢互联（PCIe / 无 NVLink）更友好**，这是在 dash5 上做 PP 的主要动机之一。
+
+## 2. 切分方案（layer-wise）
+
+Qwen3-8B：`hidden=4096`、`num_heads=32`、`num_kv_heads=8`、`head_dim=128`、
+`intermediate=12288`、`layers=36`、`vocab=151936`。`36` 能被 `2/4` 整除（PP=3/6 需处理余数，
+本阶段先要求 `layers % P == 0`）。
+
+设 stage 数 `P`，本 stage = `s`，每段 `L = layers / P` 层，本段持有全局层
+`[s·L, (s+1)·L)`：
+
+| 组件 | 持有者 | 说明 |
+|------|--------|------|
+| `embed_tokens` `[vocab, hidden]` | **仅 stage 0** | token → hidden |
+| transformer block `i` 的全部权重 | 持有 `i` 的那个 stage | 不切 head / 中间维（与 TP 正交） |
+| 该层 KV cache | 持有 `i` 的那个 stage | **每卡 KV 降到约 1/P** |
+| 最终 `norm` `[hidden]` | **仅最后一段** | |
+| `lm_head` `[vocab, hidden]` | **仅最后一段** | hidden → logits |
+
+- 注意力 / MLP 的层内计算 **完全不变**（不需要 AllReduce）：每个 stage 用它自己那几层
+  的完整权重、完整 head 做 forward。PP 与 TP 正交，可叠加（本阶段不实现 TP×PP）。
+- **RoPE** 用全局绝对 position，每个 stage 的 `RopeCache` 完全相同（按 position 索引），
+  各 stage 独立做，无需通信。
+- **每个 stage 一个独立的 `PagedKVCache`**，层数 = 本段层数 `L`（不是 36）。forward 时
+  按「本段内的局部层号 `0..L`」索引 cache —— 与单卡代码完全一致，只是 `self.layers`
+  只装了本段的层。实现技巧：给 cache 传一个 `num_hidden_layers` 改写成 `L` 的 config 克隆，
+  **无需改 `PagedKVCache`**。
+
+### 通信点
+- prefill：stage `s` 算完本段层，得到 `[S, hidden]` → **send 给 `s+1`**；`s+1` recv 后接着算。
+- decode：同理传 `[B, hidden]`（batch=1 时 `[1, hidden]`）。
+- 每 token 共 `P-1` 次 send/recv；最后一段算出 logits 并采样。
+- 采样得到的 token id（一个 `u32`）由 **最后一段经线程内 channel 回传给 stage0**
+  （同进程多线程，无需走 NCCL）。
+
+## 3. 进程 / 线程模型
+
+沿用 TP 的 **单进程、多线程**：每个 stage 一个 OS 线程，线程启动时 `cudaSetDevice(stage)`。
+- **stage 0 = 协调者（coordinator）**，跑在调用线程上：持有 scheduler、tokenizer、HTTP
+  response sender、停止判定（eos / max_tokens）与「下一步输入 token」。
+- **stage 1..P-1 = worker 线程**：从控制 channel 收命令（Register/Prefill/Decode/Free/Shutdown），
+  每步 `recv` 上游 hidden → 跑本段层 → `send` 给下游；最后一段 `head`+采样 → 把 token 回传 stage0。
+- 控制信息（命令、采样参数、token id）走 `mpsc`（极小）；**重活（hidden 张量）走 NCCL P2P（GPU↔GPU）**。
+
+> **v1 串行语义**：一次处理一个请求、一次一个 token，流水线每步「灌满又排空」
+> （stage0 decode 第 `t+1` 步依赖最后一段第 `t` 步采出的 token）。这保证 **正确性**，
+> 并拿到 TTFT/TPOT 与每卡显存；**throughput 的真正收益来自 microbatch/请求级流水线
+> 重叠（1F1B）**，列为后续工作（见 §7）。
+
+执行流（每请求）：
+```
+coordinator                     worker s (1..P-1)            last stage (P-1)
+─────────────                   ─────────────────            ────────────────
+broadcast Register(slot)        cache.register(slot)         cache.register(slot)
+broadcast Prefill{n,slot,samp}
+ x=embed(prompt)
+ x=layers_prefill(x,slot)
+ send x → stage1                recv x ← s-1
+                                x=layers_prefill(x,slot)
+                                send x → s+1 ───────────────► recv x ← P-2
+                                                             x=layers_prefill(x,slot)
+                                                             logits=head(x); next=sample
+ next ◄────────────── token channel ◄──────────────────────  token_tx.send(next)
+ stream(next); loop Decode{slot} 直到 eos/length
+broadcast Free(slot)            cache.free(slot)             cache.free(slot)
+```
+
+## 4. 通信库：NCCL P2P
+
+复用 `xserv-distributed`（已有 NCCL FFI + `TpContext`/AllReduce），新增：
+- FFI：`ncclSend(sendbuff, count, dtype, peer, comm, stream)`、
+  `ncclRecv(recvbuff, count, dtype, peer, comm, stream)`。
+- `PpContext`：与 `TpContext` 同样的 `ncclCommInitRank`（一个 comm 跨 `P` 个 stage），
+  外加 `send_bf16_ptr(ptr, count, peer)` / `recv_bf16_ptr(ptr, count, peer)`，在 **null
+  stream** 上发起（与模型 kernel 同流，天然有序）。
+- 线性流水线无死锁：stage0 只 send、最后一段只 recv、中间段「先 recv 上游、再 send 下游」，
+  依赖链无环，从头解锁。每个 stage 在 send/recv + 本段计算后 `synchronize()`，
+  确保 NCCL 读完发送缓冲再复用/释放（v1 串行下成本可接受）。
+
+> **决策点**：和 TP 一样，collective/P2P 先用 NCCL 把 PP 跑通拿正确性与基线；
+> 手写 P2P（PCIe 上的 cudaMemcpyPeer）作为后续学习项。
+
+## 5. 权重分片加载
+
+`Qwen3::from_weights_pp(config, weights, stage, num_stages, device)`：
+- 只把全局层 `[s·L, (s+1)·L)` 搬到本 stage 的 GPU（其余层的权重直接 drop，不占显存）。
+- `embed_tokens`：仅 stage 0 加载；其余 stage 放一个 1×1 占位张量（forward 用 `is_first_stage`
+  守卫，永不触碰）。
+- `norm`/`lm_head`：仅最后一段加载；其余放占位。
+- head 不切（不做 TP），所以 `local_num_heads = num_heads`、`local_num_kv_heads = num_kv_heads`。
+
+每卡显存 ≈ `权重(transformer 1/P) + KV(1/P) + (stage0: embed) + (last: norm+lm_head)`。
+对 Qwen3-8B：transformer 层约 14GB，PP=2 每卡约 7GB 层权重 + embed 或 lm_head（各 ~1.2GB）。
+
+## 6. 实现步骤（逐步可验证）
+
+1. **P18.1 — `xserv-distributed` P2P**：`ncclSend/Recv` FFI + `PpContext`。
+   验收：2 卡，rank0 send 已知向量、rank1 recv，校验一致（`tests/sendrecv.rs`）。
+2. **P18.2 — 分段权重加载**：`from_weights_pp`，每 stage 只持有本段层 + 该有的 embed/head。
+   验收：各 stage 层数 = `L`、显存约 1/P（+ embed/head）。
+3. **P18.3 — stage forward**：`embed` / `forward_layers_prefill` / `forward_layers_decode` /
+   `head`，每段独立 KV cache。
+   验收：**PP=1 与单卡 `forward_*_paged` 逐 token 一致**（同一条代码路径退化）。
+4. **P18.4 — PP engine + `--pp N`**：多线程 stage workers + NCCL 传递 + stage0 协调。
+   验收：`--pp 2/4` 端到端可服务；**greedy 输出与单卡（PP=1）逐 token 一致**；
+   用现有 llama.cpp bench 跑正确性（GSM8K/AIME）；测 PP=1/2/4 的 TTFT/TPOT/每卡显存。
+
+## 7. 预期与风险
+
+- **显存**：每卡 transformer 权重 + KV ≈ 1/P，这是 PP 的主要收益（可上更大模型 / 更长 context）。
+- **单流吞吐**：v1 串行无 stage 重叠 → 单流 tok/s **不会超过单卡**（多一份 P2P + sync 开销，
+  可能略低）。这是 PP 的本质：**没有 microbatch 重叠就没有加速**。诚实记录实测，并与
+  llama.cpp 的 `--split-mode layer`（同样是层切流水线、单序列也串行跨卡）对比 —— 两者单流
+  都应≈单卡。
+- **真正的 throughput 收益**（后续）：请求级 / microbatch 流水线（1F1B），让 stage 间重叠：
+  stage1 算 microbatch A 时 stage0 算 B。需要把 scheduler 改成跨 stage 连续批处理。
+- **风险**：NCCL 多线程 init 同步；send 缓冲生命周期（必须 sync 后再复用）；
+  `layers % P != 0` 的余数分配（本阶段先约束整除）；与 CUDA Graph decode 的结合（先走非 graph 路径）。
+- 正确性优先：先 PP=1 等价（逐 token 对齐），PP=2/4 与单卡对齐，再谈性能。
+
+## 8. 与 llama.cpp 的对比口径
+
+- **xserv**：`--pp N`，`CUDA_VISIBLE_DEVICES=0..N-1`。
+- **llama.cpp**：`-sm layer`（默认即层切流水线）+ `--tensor-split` 均分层，`CUDA_VISIBLE_DEVICES=0..N-1`。
+  （对照 TP 用的是 `-sm row`。）
+- 指标：正确性（GSM8K / AIME exact-match）、单流 TTFT/TPOT、并发吞吐、每卡 VRAM。
+- 复用 `tools/bench/runner.py` 与 `run_pp_parallel.sh`（仿 `run_tp_parallel.sh`）。
+
+## 9. 不在本阶段范围
+
+- TP×PP 混合（2D 并行）、跨组 / 多节点。
+- microbatch / 1F1B 流水线重叠（throughput 收益，后续）。
+- vocab-parallel embedding / lm_head。
+- `layers % P != 0` 的非均匀切分；与 CUDA Graph decode 结合。

docs/benchmarks/fp8-quantization.md

@@ -0,0 +1,99 @@
+# FP8 W8A8 quantization — gpt-oss-20b (dash5, 8× RTX 5090)
+
+Operator-level FP8 E4M3 quantization of the MoE expert weights, with real
+cuBLASLt FP8 tensor-core GEMM (W8A8: FP8 weights × dynamically-quantized FP8
+activations). All other tensors (attention, router, embeddings, norms, biases)
+stay BF16.
+
+## Scheme
+
+- **Weights** (`tools/quantize_fp8.py`): expert `gate_up_proj` / `down_proj`
+  quantized BF16 → FP8 E4M3 with a **per-expert scalar** scale (`absmax/448`).
+  Stored transposed `[E, N, K]` because cuBLASLt FP8 on Blackwell (sm120)
+  requires `transA=T`.
+- **Activations**: quantized dynamically at runtime, **per-token** (per-row
+  absmax), recovered by a post-GEMM row scale.
+- **Compute**: `batched_gemm_fp8` (`crates/xserv-kernels/src/quantization.rs`)
+  runs **one strided-batched cuBLASLt FP8 matmul for all experts** (`alpha=1`,
+  in-GEMM scales `1.0`); a fused kernel then applies `a_scale[token]·b_scale[expert]`
+  in a single pass. BF16's relative error is scale-invariant, so applying both
+  scales post-GEMM is precision-equivalent to folding them into the epilogue.
+- Model size: **22 GB** (FP8) vs **39 GB** (BF16). The FP8 model fits on a
+  single 32 GB 5090; BF16 needs ≥ 2.
+
+## The performance bug that was fixed
+
+`batched_gemm_fp8` originally rebuilt the entire cuBLASLt plan **per expert,
+per GEMM, per layer, on every forward pass** — running the algo heuristic
+search, creating/destroying the descriptor + 4 layouts + preference, and
+`cudaMalloc`-ing a 4-byte scale buffer — roughly 1500 heuristic searches per
+decoded token. This made FP8 **slower than BF16**:
+
+| | FP8 (buggy) | FP8 (fixed) | BF16 |
+|---|---|---|---|
+| Decode TPOT | 27.0 ms | **17.9 ms** | 18.8 ms |
+| Throughput | 37 tok/s | **55.8 tok/s** | 53.2 tok/s |
+
+Fix: cache the cuBLASLt plan (descriptor + layouts + heuristically-chosen algo)
+in a thread-local map keyed by `(M, N, K, batch)` so the heuristic runs once per
+shape, and allocate the scale buffer once.
+
+## Reducing launches: one strided-batched matmul
+
+The per-expert loop still issued one `cublasLtMatmul` per expert — ~768 tiny
+launches per decoded token (16 local experts × 2 GEMMs × 24 layers). Collapsing
+each MoE GEMM into a single **strided-batched** cuBLASLt FP8 matmul (BATCH_COUNT
++ strided-batch offsets) drops that to ~48, with a fused post-scale kernel
+applying both scales. This required moving the per-expert weight scale out of the
+GEMM epilogue (a single strided call can't carry a per-batch scalar) into the
+post-scale kernel — precision-equivalent, as noted above.
+
+| (gpt-oss-20b, TP=2) | per-expert FP8 | batched FP8 | BF16 |
+|---|---|---|---|
+| Decode TPOT | 17.9 ms | **13.8 ms** | 18.8 ms |
+| Throughput | 55.8 tok/s | **72.3 tok/s** | 53.2 tok/s |
+
+## Results — GSM8K (greedy, TP=2 on the same 2 GPUs)
+
+200-problem run is the per-expert plan-cache fix; 100-problem run is the
+strided-batched version. BF16 is the unchanged baseline in both.
+
+Harness: `tools/fp8_compare.py` — a warm `xserv-server` per model, GSM8K streamed
+through `/v1/chat/completions`; TTFT = time to first token, TPOT = mean
+inter-token latency, per request.
+
+| metric | FP8 per-expert (n=200) | FP8 batched (n=100) | BF16 |
+|---|---|---|---|
+| GSM8K accuracy | 93.0 % | 91.0 % | 90.5 / 90.0 % |
+| TTFT median | 67.4 ms | 65.0 ms | 68.8 / 69.5 ms |
+| TPOT median | 17.45 ms | **13.08 ms** | 18.26 / 18.39 ms |
+| TPOT p90 | 17.65 ms | **13.28 ms** | 18.38 / 18.52 ms |
+| Throughput | 57.3 tok/s | **76.4 tok/s** | 54.8 / 54.4 tok/s |
+| Decode speedup vs BF16 | 1.05× | **1.41×** | 1.00× |
+
+- **Accuracy: unchanged.** FP8 is nominally +0.5 … +2.5 pts above BF16, but at
+  n=100–200 the standard error is ~2–3 pts, so they are statistically
+  indistinguishable. The takeaway is that neither FP8 quantization nor the
+  strided-batched rounding degrades accuracy.
+- **Decode: FP8 1.41× faster** once batched (TPOT 13.08 vs 18.39 ms), with a
+  tight p90. The per-expert version was only ~1.05× — the ~768 tiny M=1 launches
+  per token dominated; batching them into ~48 unlocked most of the FP8
+  expert-weight-bandwidth saving.
+- **Prefill (TTFT): comparable.** A multi-length sweep (113 / 561 / 1681 tokens)
+  gave FP8 480 / 362 / 2451 ms vs BF16 558 / 282 / 2287 ms — non-monotonic, i.e.
+  dominated by fixed overhead (cuBLAS lazy init + FP8's one-time per-shape
+  heuristic), not prefill compute, at these lengths.
+
+## Single-GPU (TP=1)
+
+FP8 runs gpt-oss-20b on **one** 5090 (`bench-gpt-oss --tp 1`, GPU6): TTFT 538 ms,
+TPOT 29.0 ms, 34.5 tok/s. BF16 cannot (39 GB > 32 GB). This — fitting a model
+that otherwise needs two GPUs onto one — is the largest practical win.
+
+## Follow-ups (not done)
+
+- Per-channel (per-output-row) weight scales for better accuracy headroom than
+  per-tensor.
+- Warm common prefill shapes at load to hide the first-request heuristic stall.
+- Sparse (top-k only) MoE compute instead of dense — currently every token runs
+  all experts, so only ~top_k/num_experts of the FP8 GEMM work is used.

docs/benchmarks/llama-cpp-comparison.md

@@ -0,0 +1,102 @@
+# Benchmark: xserv vs llama.cpp (Qwen3-8B)
+
+**What this adds.** A standing baseline that compares xserv against **llama.cpp**
+on both **response quality (correctness)** and **performance (TTFT / TPOT /
+throughput)**, using the same model weights and standard public datasets. This
+replaces HF transformers as our reference point — xserv already beat HF, so it
+is no longer a useful performance bar.
+
+- **Baseline engine**: llama.cpp, vendored as a submodule pinned to `b9371`,
+  built with CUDA for SM120 (RTX 5090).
+- **Same weights**: the Qwen3-8B safetensors are converted to a **BF16 GGUF**
+  (`convert_hf_to_gguf.py --outtype bf16`) — no quantization, so the comparison
+  is apples-to-apples.
+- **Standard quality datasets**: **AIME 2025** (30 competition-math problems,
+  exact-match boxed integer) and **GSM8K** (grade-school math, exact-match).
+- **Black-box HTTP**: both engines are driven through the OpenAI-compatible
+  streaming API; the driver measures TTFT/TPOT/throughput and scores answers.
+
+See `docs/16-llama-cpp-comparison.md` for the design and `tools/bench/` for the
+driver. One-click: `tools/sync-and-build.sh bench`.
+
+## How it runs
+
+The GPU host (dash5) has no outbound network, so datasets are fetched locally
+(`tools/bench/fetch_datasets.py`) into JSON and the llama.cpp source is shipped
+over with the project; everything builds and runs on the GPU host. The driver
+runs **one engine at a time** (two BF16 8B models do not co-reside on a 32GB
+GPU, and a resident idle engine would distort the other's numbers).
+
+Generation mode is matched: xserv hardcodes Qwen3 **thinking off**, so the
+driver sends `chat_template_kwargs={enable_thinking:false}` to llama.cpp.
+
+## Results (RTX 5090, BF16, greedy, 8192 ctx, max_batch 4)
+
+### Performance — llama.cpp is the stronger baseline
+
+| scenario | metric | xserv | llama.cpp | xserv ÷ llama.cpp |
+|---|---|---|---|---|
+| single / medium | TTFT p50 (ms) | 28.0 | 17.7 | 0.63× |
+| single / medium | TPOT p50 (ms/tok) | 17.5 | 10.4 | 0.60× |
+| single / medium | throughput (tok/s) | 56.6 | 95.1 | 0.60× |
+| concurrent-4 | throughput (tok/s) | 135.2 | 317.1 | 0.43× |
+| concurrent-8 | throughput (tok/s) | 135.5 | 322.5 | 0.42× |
+
+xserv runs at **~0.42–0.60×** llama.cpp. It saturates at `max_batch` (~135 tok/s)
+while llama.cpp keeps scaling under load (~322 tok/s). This is the honest new bar.
+The ratio is the same at 4096 and 8192 — TPOT is bandwidth-bound, not
+context-bound at these sizes.
+
+### Quality — parity, confirming xserv's numerical fidelity
+
+| task | n | xserv | llama.cpp |
+|---|---|---|---|
+| GSM8K | 50 | 100.0% (50/50) | 96.0% (48/50) |
+| AIME 2025 | 30 | 16.7% (5/30) | 23.3% (7/30) |
+
+With equal context the two engines land at comparable AIME accuracy (within
+the ±2-problem greedy-decode wobble band) and xserv edges ahead on GSM8K. At
+8192 both generate full-length solutions (mean ~4.2k tokens), so neither is
+truncated. The AIME difference (2 problems) is entirely within the run-to-run
+non-determinism documented below. Per-problem analysis shows the disagreements
+are due to different greedy-decode paths (different token at position ~500+
+cascades into a different solution), not systematic precision errors.
+
+On GSM8K, xserv strictly dominates: it gets 2 problems right that llama.cpp
+misses, and never misses one that llama.cpp gets.
+
+## Findings the benchmark surfaced
+
+1. **Context must be provisioned per-request, not total.** A first run showed
+   xserv 20.0% vs llama.cpp 3.3% on AIME — an artifact: llama.cpp divides total
+   `-c` across `--parallel` slots, so `-c 4096 --parallel 4` gave each request
+   only **1024 tokens**, truncating long AIME solutions before the boxed answer
+   (capped at ~940 generated tokens). GSM8K (~280 tokens) was unaffected, which
+   is how we caught it. Fixed: per-slot context = `max_seq_len` (total
+   `-c = max_seq_len × parallel`). After the fix, AIME is at parity (above).
+2. **xserv OOM'd at `--max-seq-len 8192` — now fixed.** xserv used to eagerly
+   pre-allocate its paged-KV pool (`blocks_per_seq × max_batch × 2`, ~9GB at
+   8192) on top of the 16GB weights, exceeding 32GB at startup. Fixed by sizing
+   the pool to *available VRAM* (`cudaMemGetInfo`) instead of worst-case demand,
+   plus vLLM-style **swap to pinned host memory**: when running sequences grow
+   past the GPU pool, the newest are evicted to host and swapped back when blocks
+   free up (`--swap-space-gb`, default 8). The results above run at 8192 with **0
+   swap events** — the VRAM-sized pool alone covers this load; swap is the
+   overload safety net (verified lossless under a forced-small pool).
+3. **xserv decode is not run-to-run deterministic.** The same greedy (temp 0)
+   AIME config produced 6/30 / 7/30 / 6/30 across runs — non-deterministic CUDA
+   reductions flip an argmax over long (~3k-token) generations. Harmless for
+   serving, but it explains why long-sequence accuracy wobbles by a problem.
+4. **GEMV race condition corrupted decode outputs — now fixed.** The custom
+   K-split GEMV kernel (used for all M=1 decode-step projections with N≥256)
+   had a race condition: block k=0 zeroed the FP32 accumulator (`y_fp32[col] =
+   0.0`) while other K-blocks were already atomicAdding to it. Since CUDA
+   provides no inter-block ordering within a single kernel launch, the zero
+   could land before, during, or after other blocks' writes. Fix:
+   `cudaMemsetAsync` on the stream before the kernel launch, which guarantees
+   the buffer is zeroed before any block executes. This bug was introduced
+   after the initial benchmark and caused systematic decode-time precision
+   errors that degraded GSM8K accuracy from 98→80% range.
+
+Raw artifacts (per-request timings, per-problem prediction/gold) are written to
+`bench-out/` as `comparison-<stamp>.{md,json}` (gitignored).

docs/benchmarks/mxfp4-and-llama-decode.md

@@ -0,0 +1,71 @@
+# MXFP4 W4A16 + decode-speed vs llama.cpp (gpt-oss-20b, 2×RTX 5090)
+
+## xserv vs llama.cpp — single-stream decode (TP=2, same GPUs)
+
+`tools/xserv_vs_llama.py` streams identical prompts through each server's
+OpenAI endpoint (counting llama's `reasoning_content` as real decode tokens).
+
+| metric | xserv FP8 | llama MXFP4 |
+|---|---|---|
+| Decode TPOT (medium) | 13.1 ms | **6.6 ms** (2.0× faster) |
+| Throughput | 76 tok/s | **151 tok/s** |
+| TTFT (short/medium) | 35–50 ms | 60–63 ms |
+| TTFT (long, 1.6k tok) | 94 ms | **35 ms** |
+
+llama.cpp decodes ~2× faster; prefill is comparable-to-better.
+
+## Why — decode is memory/comm-bound, not launch-bound
+
+Traced + measured (not assumed):
+
+- The 24-layer decode loop is already fully async (no per-layer syncs), so kernel
+  launches hide behind GPU work — a CUDA graph would buy ~0.5–1.5 ms, not 2×.
+- **TP=2→TP=4 probe**: TPOT 13.5→10.2 ms (FP8) with the *same* launch count and
+  *more* NCCL — confirms the bottleneck is **expert HBM traffic + all-reduce**,
+  not launch overhead.
+- Even FP8 TP=4 (10.2 ms) can't catch llama TP=2 (6.6 ms): the gap is
+  *algorithmic*. llama is **sparse (top-4 of 32 experts) + 4-bit (MXFP4)**;
+  xserv is **dense (all 16 local experts) + 8-bit (FP8)** → ~8× the expert bytes
+  per token. Dense also makes xserv's long-prefill TTFT worse.
+
+The two levers that close it: **sparse top-k MoE** (≈4×, the bigger structural
+change) and **4-bit weights** (≈2×).
+
+## MXFP4 W4A16 (this change) — correct, smallest, not yet faster than FP8
+
+Weight-only 4-bit: expert weights are MXFP4 (E2M1 + per-32 UE8M0 scale,
+`tools/quantize_mxfp4.py`); a fused kernel reads the 4-bit weights and
+dequantizes on-chip to BF16. Decode uses `batched_gemv_mxfp4`; prefill (M>1)
+dequantizes to BF16 then reuses the BF16 batched GEMM.
+
+| | MXFP4 W4A16 | FP8 W8A8 | BF16 |
+|---|---|---|---|
+| Model size | **13 GB** | 22 GB | 39 GB |
+| Greedy tokens | identical | identical | baseline |
+| Decode TPOT (TP=2) | 17.0 ms | **13.5 ms** | 18.8 ms |
+| Decode TPOT (TP=4) | 11.8 ms | **10.2 ms** | — |
+| Prefill TTFT | 350 ms | **134 ms** | 135 ms |
+
+- **Correct** (byte-identical greedy tokens to FP8/BF16) and **smallest
+  footprint** — fits one 32 GB 5090 with ample room for KV cache.
+- **Not faster than FP8**: the hand-written W4A16 dequant-GEMV (no tensor cores)
+  is less efficient than cuBLASLt's FP8 tensor-core GEMM, so even reading half
+  the bytes it stays ~2–3.5 ms behind FP8 at every TP. The TP=4 scaling
+  (17→11.8) shows it *is* partly memory-bound; a fixed per-GEMM inefficiency
+  dominates. Vectorized loads, hoisted scale, warp reduction, and shared-memory
+  activation tiling did not change it.
+- **Prefill regresses** (350 vs 134 ms) — the dequant-to-BF16 fallback.
+
+Committed as a **memory-optimization foundation**, not a decode speedup.
+
+## To make 4-bit actually win
+
+- **FP4 tensor cores (W4A4)** — cuBLASLt block-scaled MXFP4 GEMM
+  (`CUDA_R_4F_E2M1` + `CUBLASLT_MATMUL_MATRIX_SCALE_VEC32_UE8M0`, available on
+  sm_120). Tensor-core throughput *at* 4-bit would beat FP8. Risk: the scale
+  swizzle layout.
+- A **Marlin-class W4A16 kernel** (register-blocked, async-copy pipelined).
+- **Sparse top-k MoE** for the larger, llama-matching win.
+
+FP8 (the plan-cache fix + strided-batched optimization, 1.41× over BF16) remains
+xserv's best-performing quantization today.

docs/benchmarks/phase14-flash-attention.md

@@ -0,0 +1,109 @@
+# Phase 14 Benchmark: Flash Attention 2
+
+**Date**: 2026-05-22
+**Hardware**: RTX 5090 (32GB GDDR7, SM120 CC 12.0, 170 SMs)
+**Model**: Qwen3-8B (BF16, 36 layers, 4096 hidden, 32 Q / 8 KV GQA heads, head_dim=128)
+**Config**: greedy decoding (temperature=0), max_tokens=64, single-request serial
+
+## Correctness
+
+Logits comparison with HuggingFace transformers (10 prompts, raw text without ChatML):
+
+| Metric | Result |
+|--------|--------|
+| Prefill Top-1 match vs HF | **9/10 (90%)** |
+| Avg Top-5 overlap vs HF | **4.0/5** |
+| Result vs pre-FA2 naive attention | **Identical** (same 9/10 top-1, same 4.0/5 overlap) |
+
+The single top-1 mismatch ("Explain quantum computing.") has logits differing by 0.125
+(22.000 vs 21.875) — within BF16 precision. The top-5 sets are identical (5/5 overlap).
+
+FA2 introduces no precision degradation compared to the naive attention path.
+
+## API Generation
+
+52 diverse prompts (English, Chinese, code) via `/v1/chat/completions`:
+
+| Metric | Result |
+|--------|--------|
+| Success rate | **52/52 (100%)** |
+| SSE streaming | **Working** (role chunk, content chunks, finish_reason, [DONE]) |
+| Usage stats | Correct (prompt_tokens + completion_tokens = total_tokens) |
+
+## Performance
+
+### xserv vs HuggingFace transformers
+
+8 prompts (short/medium/long) × max_tokens=64, greedy:
+
+| Category | Prompt Tokens | xserv (tok/s) | HF (tok/s) | Ratio |
+|----------|--------------|---------------|------------|-------|
+| Short (~12 tok) | 12-14 | 12.5 | 38.5 | 0.32x |
+| Medium (~28 tok) | 27-28 | 13.6 | 44.1 | 0.31x |
+| Long (~60 tok) | 58-64 | 13.0 | 36.0 | 0.36x |
+| **Overall** | — | **12.9** | **36.6** | **0.35x** |
+
+### Phase-over-Phase Improvement
+
+| Phase | Attention | repeat_kv | tok/s | vs HF |
+|-------|-----------|-----------|-------|-------|
+| 10 | Naive (O(S²), cuBLAS batched) | CPU round-trip | 6.9 | 15% |
+| 11 | Naive + GPU KV cache | GPU repeat_kv | 10.3 | 30% |
+| **14** | **FA2 (O(1), fused kernel)** | **None (GQA in kernel)** | **12.9** | **35%** |
+
+Phase 14 vs Phase 11: **+25% throughput** (10.3 → 12.9 tok/s).
+
+### Improvement Breakdown (estimated)
+
+| Factor | Contribution |
+|--------|-------------|
+| Eliminating repeat_kv GPU alloc + copy (per layer) | ~10% |
+| Eliminating K^T transpose + contiguous | ~5% |
+| Eliminating S×S score matrix alloc | ~5% |
+| Fused kernel (1 launch vs 6) | ~5% |
+
+### Concurrent Requests
+
+8 concurrent requests, max_batch=4:
+
+| Metric | Result |
+|--------|--------|
+| Wall clock | 22.5s |
+| Sum of individual latencies | 135.0s |
+| Scheduling speedup | **6.0x** |
+| Throughput | 11.4 tok/s |
+
+Continuous batching scheduling confirmed working (decode batch_size=4 in logs).
+
+## Remaining Performance Gap
+
+35% of HF throughput. Main bottlenecks:
+
+| Bottleneck | Impact | Fix |
+|-----------|--------|-----|
+| **Decode Q_len=1 inefficiency** | FA2 kernel: 64 threads, only 1 active (owns_row=true for single query) | Specialized decode attention kernel (vector-dot against KV, parallel reduction along S) |
+| **No kernel fusion** | RMSNorm+residual, SiLU*up: separate kernels, redundant HBM reads/writes | Fused kernels (Phase 15) |
+| **No CUDA Graphs** | ~100+ kernel launches per decode step, each has host-side overhead | Capture decode iteration as CUDA Graph (Phase 15) |
+| **Per-seq forward (no batched decode)** | With batch=4, 4 serial forward passes per iteration | Batched projections + per-seq attention (Phase 15, depends on FA2 decode kernel) |
+| **No vectorized loads in FA2** | Scalar bf16→f32 conversion in dot product loop | float4 / bfloat162 vectorized loads |
+
+## Memory Usage
+
+| Component | Naive (Phase 11) | FA2 (Phase 14) |
+|-----------|-----------------|----------------|
+| Score matrix [1, 32, S, S] | S² × 32 × 2B | **0** |
+| repeat_kv K/V [1, 32, S, 128] | 2 × S × 32 × 128 × 2B per layer | **0** |
+| K^T contiguous copy | S × 32 × 128 × 2B per layer | **0** |
+
+For S=256 (current max): savings ~6 MB per layer × 36 layers ≈ 216 MB.
+For S=2048: savings ~384 MB per layer × 36 layers ≈ 13.5 GB (naive would OOM).
+
+## Tracking
+
+| Phase | Attention | tok/s | vs HF | Correctness |
+|-------|-----------|-------|-------|-------------|
+| 8 | Naive (no cache) | 2.5 | 5% | 50/50 vs HF |
+| 9 | Naive + CPU KV cache | 44.3 (GPT-2) | — | 50/50 self |
+| 10 | Naive + CPU KV cache | 6.9 (Qwen3-8B) | 15% | 100% top-5 |
+| 11 | Naive + GPU KV cache | 10.3 | 30% | 9/10 top-1 |
+| **14** | **FA2 + GQA in kernel** | **12.9** | **35%** | **9/10 top-1** |

docs/benchmarks/phase15-performance.md

@@ -0,0 +1,85 @@
+# Phase 15 Benchmark: Performance Optimization
+
+**Date**: 2026-05-23
+**Hardware**: RTX 5090 (32GB GDDR7, SM120 CC 12.0, 170 SMs, 1.79 TB/s)
+**Model**: Qwen3-8B (BF16, 36 layers, 4096 hidden, 32 Q / 8 KV GQA heads, head_dim=128)
+**Config**: greedy decoding (temperature=0), max_tokens=64, serial (batch=1)
+
+## Ablation: Each Optimization Measured Independently
+
+| # | Optimization | tok/s | Delta | ms/token | Roofline |
+|---|-------------|-------|-------|----------|----------|
+| 0 | Phase 14 baseline (FA2 + naive cuBLAS GEMV) | 12.9 | — | 77.5 | 12% |
+| 1 | + Decode attention kernel (256 threads) | 12.9 | +0% | 77.5 | 12% |
+| 2 | + Fused SiLU×Mul | 13.0 | +1% | 76.9 | 12% |
+| 3 | + Fused Add+RMSNorm | 13.2 | +2% | 75.8 | 12% |
+| 4 | + Custom GEMV (M=1, K-split tiled) | 46.6 | +253% | 21.5 | 42% |
+| 5 | + Tensor::empty (skip cudaMemset) | **50.3** | **+8%** | **19.9** | **45%** |
+
+## Comparison with HuggingFace transformers
+
+8 prompts (short/medium/long) × max_tokens=64, greedy, serial:
+
+| System | tok/s | ms/token | Roofline |
+|--------|-------|----------|----------|
+| HF transformers (BF16, torch 2.8, SDPA) | 36.0 | 27.8 | 32% |
+| **xserv Phase 15** | **50.3** | **19.9** | **45%** |
+| Roofline (1.79 TB/s, 16GB model) | 112.0 | 8.9 | 100% |
+
+**xserv is 140% of HF transformers throughput.**
+
+## Per-Prompt Detail (Phase 15 Final)
+
+| # | Prompt | pt | ct | Time | tok/s |
+|---|--------|----|----|------|-------|
+| 1 | What is gravity? | 12 | 64 | 1.39s | 46.0 |
+| 2 | Hello, how are you? | 14 | 64 | 1.27s | 50.5 |
+| 3 | Explain DNA briefly. | 13 | 64 | 1.25s | 51.2 |
+| 4 | Write a detailed explanation of photosynthesis... | 27 | 64 | 1.26s | 50.7 |
+| 5 | Describe machine learning. | 13 | 64 | 1.25s | 51.2 |
+| 6 | What causes earthquakes? | 12 | 64 | 1.25s | 51.1 |
+| 7 | How does the internet work? | 14 | 64 | 1.25s | 51.1 |
+| 8 | What is the speed of light? | 15 | 64 | 1.25s | 51.0 |
+
+Prompt 1 is slower (46.0 vs 51.x) due to first-request warmup (caching allocator cold start).
+
+## Concurrent Throughput
+
+8 requests concurrent, max_batch=4:
+
+| Config | tok/s | Wall clock | Speedup |
+|--------|-------|-----------|---------|
+| Serial (batch=1, custom GEMV) | 50.3 | — | — |
+| Concurrent (batch=4, cuBLAS M=4) | 28.2 | 9.09s | 6.47x scheduling |
+| Concurrent (batch=4, custom GEMV) | 35.1* | ~7.3s | ~6x scheduling |
+
+*Note: batch=4 with custom GEMV is slower than serial because:
+1. Batched decode path uses cuBLAS for M>1 matmuls, losing the GEMV advantage
+2. Per-seq attention/reshape overhead in the batched path adds ~2ms/step
+3. Custom GEMV already saturates bandwidth at M=1
+
+Serial decode with custom GEMV is the optimal path for current architecture.
+
+## Correctness Verification
+
+| Test | Result |
+|------|--------|
+| Top-1 logits match vs HF (10 prompts) | 9/10 (90%) |
+| Top-5 overlap vs HF (10 prompts) | 4.0/5 avg |
+| vs pre-optimization baseline | Identical (same 9/10) |
+| API generation (52 prompts) | 52/52 pass |
+| SSE streaming | Working |
+| Chinese prompts | Working |
+
+## Phase-over-Phase Performance Tracking
+
+| Phase | Key Change | tok/s | vs HF | Roofline |
+|-------|-----------|-------|-------|----------|
+| 8 | GPT-2 inference (no cache) | 2.5 | 7% | — |
+| 9 | + KV cache (CPU) | 44.3 (GPT-2) | — | — |
+| 10 | Qwen3-8B (CPU KV cache) | 6.9 | 19% | 6% |
+| 11 | + GPU KV cache | 10.3 | 29% | 9% |
+| 14 | + Flash Attention 2 | 12.9 | 36% | 12% |
+| **15** | **+ Custom GEMV + fused + empty** | **50.3** | **140%** | **45%** |
+
+Total speedup from Phase 10 to Phase 15: **7.3x** (6.9 → 50.3 tok/s).

docs/benchmarks/pp-sweep.md

@@ -0,0 +1,118 @@
+# PP sweep — xserv vs llama.cpp (Qwen3-8B BF16, 8×RTX 5090)
+
+Pipeline parallelism (layer split), verified end-to-end on dash5. Qwen3-8B BF16,
+greedy, single stream, no NVLink (hand-off / split traffic over PCIe Gen5).
+xserv `--pp N` puts stage `s` on GPU `s` and hands the hidden state stage→stage
+over NCCL P2P; llama.cpp uses `-sm layer` (its default pipeline split) over N GPUs.
+
+## Single-stream latency + per-GPU VRAM (measured, `--max-seq-len 2048`)
+
+Measured strictly sequentially, one server at a time, each config gated on a real
+successful generation (so VRAM snapshots are post-load). Driver:
+`tools/pp_final.sh`.
+
+| engine | PP | TTFT_ms | TPOT_ms | tok/s | per-GPU VRAM (MiB) |
+|--------|----|---------|---------|-------|--------------------|
+| xserv  | 1  | 33.2    | 17.39   | 57.5  | 24010 |
+| xserv  | 2  | 35.9    | 18.07   | 55.3  | 11580, 13632 |
+| xserv  | 4  | 36.1    | 17.91   | 55.8  | 7298, 5250, 5250, 9350 |
+| llama  | 1  | 133.3   | 9.38    | 106.7 | 15604 |
+| llama  | 2  | 131.4   | 9.10    | 109.9 | 7862, 8494 |
+| llama  | 4  | 161.2   | 8.88    | 112.6 | 4476, 4090, 4090, 5108 |
+
+(xserv VRAM with `XSERV_MAX_KV_BLOCKS=160` so the number is weights + a minimal
+KV pool. `tok/s = 1000 / TPOT`. This latency probe's TTFT differs from the
+quality-suite TTFT below because the suite includes scheduler/HTTP overhead.)
+
+## Correctness — PP is numerically exact
+
+The hidden-state hand-off between stages is a bit-exact BF16 P2P copy and each
+stage runs the same kernels over its layers, so PP must reproduce the single-GPU
+result. Verified by byte-comparing generated text (greedy, temp 0), running each
+config **twice** to separate PP effects from run-to-run GEMM noise:
+
+| comparison | result |
+|------------|--------|
+| single run A == single run B | **DIFFER** (cuBLAS GEMM is not bit-reproducible run-to-run) |
+| pp4 run A == pp4 run B | **IDENTICAL** |
+| single run A == pp4 run A | **IDENTICAL** |
+| single == pp2 (single run each) | **IDENTICAL** |
+
+Takeaway: **single-GPU itself is non-deterministic** under greedy (a 1-ULP logit
+difference flips a late argmax and the suffix changes), so a one-shot single-vs-PP
+byte compare can spuriously "DIFFER". The 2×2 control shows PP=4 is *more*
+reproducible than re-running single-GPU, and it lands exactly on a single-GPU
+trajectory. NCCL P2P (`tests/sendrecv.rs`) and AllReduce (`tests/allreduce.rs`)
+unit tests pass.
+
+## Quality matrix — AIME 2025 (30) + GSM8K (30), greedy, both engines × PP=1/2/4
+
+Full measured matrix (`tools/bench/summarize_fullq.py`; raw in
+`bench-out/FULLQ_SUMMARY.txt`). Qwen3-8B BF16, thinking OFF, `max_seq_len 4096`.
+xserv on GPUs 0-3, llama.cpp on GPUs 4-7 (disjoint groups, run in parallel).
+
+| engine | PP | AIME 2025 | GSM8K | AIME mean_tok | TTFT_ms | TPOT_ms |
+|--------|----|-----------|-------|---------------|---------|---------|
+| xserv  | 1  | 8/30 (26.7%) | 29/30 (96.7%) | 2383 | 485 | 22.42 |
+| xserv  | 2  | 7/30 (23.3%) | 29/30 (96.7%) | 2367 | 457 | 22.55 |
+| xserv  | 4  | 7/30 (23.3%) | 29/30 (96.7%) | 2652 | 494 | 23.31 |
+| llama  | 1  | 7/30 (23.3%) | 29/30 (96.7%) | 2651 | 119 | 10.37 |
+| llama  | 2  | 7/30 (23.3%) | 29/30 (96.7%) | 2651 | 118 | 10.41 |
+| llama  | 4  | 7/30 (23.3%) | 29/30 (96.7%) | 2651 | 119 | 10.39 |
+
+Reading the matrix:
+
+- **GSM8K = 29/30 (96.7%) in every cell** — identical across both engines and all
+  PP levels. xserv's accuracy matches llama.cpp exactly on the same weights.
+- **AIME = 7/30 (23.3%) everywhere except xserv PP=1 (8/30)**. That single +1 is
+  the run-to-run greedy nondeterminism documented above (an AIME solution is
+  ~2400 tokens; one late argmax flip changes one problem's outcome) — not a PP or
+  engine effect. AIME accuracy is low because this is an 8B model with thinking
+  disabled; the point here is the *cross-engine / cross-PP agreement*, which holds.
+- **TPOT is flat across PP** for both engines (xserv 22.4→23.3 ms, llama
+  10.3→10.4 ms), reconfirming PP doesn't slow single-stream decode. The ~2.2×
+  TPOT gap to llama.cpp is the single-GPU gap (`llama-cpp-comparison.md`),
+  orthogonal to PP.
+
+## Takeaways
+
+- **Memory is the win.** Per-GPU weights+KV scale ~1/P: xserv 24.0 GB (1 GPU) →
+  ~11–14 GB (PP=2) → ~5–9 GB (PP=4); llama 15.6 → ~8 → ~4–5 GB. The two end
+  stages sit higher (stage 0 holds `embed_tokens`, the last stage `norm`+`lm_head`,
+  ~1.1 GB each). This is what PP buys: a model / context that does not fit on one
+  card fits across P.
+- **Single-stream latency is flat, not faster.** v1 PP is serial across stages
+  (no microbatch overlap): per-token latency = sum of all stages' compute +
+  (P-1) P2P hops + a blocking sync per stage. The `[1, hidden]` BF16 hop (8 KB)
+  over PCIe is cheap relative to per-token compute, so TPOT is ~constant across P.
+  PP does **not** speed up single-stream decode; it trades (almost no) latency for
+  large memory headroom.
+- **Quality is preserved and matches llama.cpp.** GSM8K 96.7% in all 12 cells;
+  AIME within the greedy noise band. PP=1/2/4 agree, and xserv tracks llama.cpp.
+
+## Reproduce
+
+```bash
+./tools/sync-and-build.sh build
+# latency + VRAM + byte-exact correctness (writes bench-out/PP_FINAL.md):
+ssh <host> 'cd <repo> && bash tools/pp_final.sh'
+# determinism control (single×2 vs pp4×2):
+ssh <host> 'cd <repo> && bash tools/pp_diag.sh'
+# NCCL P2P + AllReduce unit tests:
+ssh <host> 'cd <repo> && cargo test -p xserv-distributed --release'
+# full quality matrix AIME-30 + GSM8K-30 (xserv 0-3 serial; or parallel w/ llama 4-7):
+ssh <host> 'cd <repo> && bash tools/pp_quality_full.sh'   # xserv+llama serial, GPU 0-3
+ssh <host> 'cd <repo> && bash tools/pp_llama_47.sh'        # llama on GPU 4-7 (parallel)
+python3 tools/bench/summarize_fullq.py bench-out
+```
+
+## Next (where PP actually raises throughput)
+
+- **Microbatch / 1F1B overlap**: while stage 1 runs microbatch A, stage 0 runs B.
+  This is the only thing that turns PP into a *throughput* win; v1 is serial, so
+  P GPUs give 1 GPU's single-stream rate (but P× the memory headroom / batch room).
+- Persistent per-stage recv buffers (drop the per-token CPU alloc + H2D) and
+  event-based ordering instead of a full device sync per hop.
+- 2D TP×PP, and `layers % P != 0` non-uniform splits.
+
+🤖 Generated with [Claude Code](https://claude.com/claude-code)

docs/benchmarks/tensor-parallelism.md

@@ -0,0 +1,73 @@
+# Benchmark: Tensor Parallelism (TP=1/2/4) — xserv vs llama.cpp
+
+**Setup.** Qwen3-8B BF16 on 8× RTX 5090 (PCIe Gen5, **no NVLink**; GPUs grouped
+0-3 / 4-7 by PHB). Both engines driven over the same OpenAI HTTP harness, same
+scorers, thinking-off, greedy (temp 0), `max_tokens` 2048. Datasets: **AIME
+2025** (30) + **GSM8K** (30). The two engines run **concurrently on disjoint
+groups** — xserv on GPU 0..N-1, llama.cpp (`--split-mode row`) on GPU 4..4+N-1
+(`tools/bench/run_tp_parallel.sh`).
+
+## Correctness — on par across engines and TP
+
+| TP | task | xserv | llama.cpp |
+|----|------|-------|-----------|
+| 1 | AIME 2025 | 16.7% (5/30) | 13.3% (4/30) |
+| 1 | GSM8K | 96.7% (29/30) | 96.7% (29/30) |
+| 2 | AIME 2025 | 13.3% (4/30) | 13.3% (4/30) |
+| 2 | GSM8K | 93.3% (28/30) | 96.7% (29/30) |
+| 4 | AIME 2025 | 16.7% (5/30) | 13.3% (4/30) |
+| 4 | GSM8K | 96.7% (29/30) | 96.7% (29/30) |
+
+Within ±1 problem everywhere — TP changes nothing about quality on either
+engine, and the two engines agree. (AIME is low for both: Qwen3-8B thinking-off,
+capped at 2048 tokens.)
+
+## Performance — TPOT (ms/token, lower is better)
+
+| TP | xserv AIME / GSM8K | llama.cpp AIME / GSM8K |
+|----|--------------------|------------------------|
+| 1 | 21.0 / 17.8 | **10.4 / 10.3** |
+| 2 | 17.2 / 13.9 | 19.0 / 18.9 |
+| 4 | **15.2 / 12.1** | 20.2 / 20.2 |
+
+**Opposite TP scaling, with a crossover:**
+
+- **xserv TP scales positively**: TPOT 21.0 → 17.2 → 15.2 ms (AIME),
+  17.8 → 13.9 → 12.1 ms (GSM8K) — TP=4 is ~1.4–1.5× faster than TP=1. GPU 0-3
+  all ~82% utilized. (Sublinear because of the 72 PCIe AllReduces/token.)
+- **llama.cpp row-split regresses**: TPOT 10.4 → 19.0 → 20.2 ms — TP=1 is its
+  best; TP=2/4 nearly double the latency. GPU 4-7 only ~24% utilized
+  (communication-bound). Row-split's per-layer cross-GPU traffic over PCIe
+  without NVLink dominates.
+- **Crossover**: llama.cpp is ~2× faster at TP=1, still ahead at TP=2, and xserv
+  is ~1.3× faster at TP=4 (15.2 vs 20.2 ms AIME). AIME-30 wall clock: xserv
+  1046 → 846 → 730 s (falling), llama.cpp 520 → 952 → 1012 s (rising).
+
+On a NVLink-less PCIe box, **xserv's TP is a genuine win and llama.cpp's
+tensor-split is counterproductive** — exactly what the topology predicts.
+
+### Clean same-path xserv scaling (bench-tp)
+
+The HTTP numbers above mix engines (xserv TP=1 uses the production continuous-
+batching engine; TP≥2 uses the serial TP coordinator). The single-stream,
+same-code-path scaling from `bench-tp` (greedy, 8 prompts × 64 tokens):
+
+| TP | xserv decode tok/s | speedup | TTFT |
+|----|--------------------|---------|------|
+| 1 | 58.5 | 1.00× | 18.0 ms |
+| 2 | 75.7 | 1.29× | 13.4 ms |
+| 4 | 86.1 | 1.47× | 11.5 ms |
+
+## Caveats
+
+- xserv **TP=1 uses the production `Engine`**, TP≥2 the serial `tp_engine`
+  coordinator — different per-token paths, so the HTTP TP=1→2 step has an engine
+  confound. The clean same-path scaling (bench-tp, above) confirms the trend.
+- xserv **TTFT is weaker** on long AIME prompts (~460–500 ms vs llama ~100–190 ms)
+  — prefill is a known optimization target.
+- llama.cpp uses `--split-mode row` (its tensor-parallel mode); the default
+  `layer` split only memory-splits, without parallel compute.
+- The TP HTTP server processes requests **serially** (sufficient for this serial
+  quality benchmark); continuous-batching TP is future work.
+
+Raw artifacts: `bench-out/tp{1,2,4}-{xserv,llama}/comparison-*.{md,json}`.