moe(wip): gpt-oss-20b groundwork — config fields, arch doc, MXFP4 tools

Phase 19 start. config.rs: explicit head_dim (gpt-oss=64) + MoE fields
(num_local_experts, num_experts_per_tok, swiglu_limit, sliding_window,
layer_types) with accessors; Qwen3/GPT-2 paths unchanged (fall back to
hidden/num_heads when head_dim absent).

docs/19-moe-gpt-oss.md: architecture + exact HF reference math (router
softmax-after-topk, interleaved clamped (up+1)*glu experts, attention
sinks, alternating sliding window, rotate_half RoPE theta=150000,
head_dim 64), verified tensor layout, MXFP4 dequant plan.
docs/MOE_PROGRESS.md: resume/handoff snapshot.

tools/mxfp4_probe.py: inspect safetensors + validate MXFP4 decode (done).
tools/gptoss_dequant.py: MXFP4 experts -> plain BF16 safetensors dir so
the existing loader reads it (no MXFP4 in Rust for the first pass).

Verified: llama.cpp (dash5, LLM_ARCH_OPENAI_MOE) runs the gpt-oss-20b
MXFP4 GGUF correctly (17*24 -> 408) = the correctness oracle. MXFP4 decode
validated in numpy. Model + GGUF staged on dash5.

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

crates/xserv-model/src/config.rs

@@ -46,6 +46,28 @@ pub struct ModelConfig {
     pub rope_theta: Option<f64>,
     #[serde(default)]
     pub tie_word_embeddings: Option<bool>,
+
+    // Explicit head_dim (gpt-oss: 64, which is NOT hidden/num_heads). When
+    // absent, head_dim() falls back to hidden/num_heads (Qwen3, GPT-2).
+    #[serde(default)]
+    pub head_dim: Option<usize>,
+
+    // MoE (gpt-oss). Absent for dense models.
+    #[serde(default)]
+    pub num_local_experts: Option<usize>,
+    #[serde(default)]
+    pub num_experts_per_tok: Option<usize>,
+    // gpt-oss clamped-SwiGLU limit (config: swiglu_limit, default 7.0).
+    #[serde(default)]
+    pub swiglu_limit: Option<f64>,
+
+    // Sliding-window attention (gpt-oss: 128 on alternating layers). The
+    // pattern is given by `layer_types` (e.g. "sliding_attention" /
+    // "full_attention" per layer); absent for dense models.
+    #[serde(default)]
+    pub sliding_window: Option<usize>,
+    #[serde(default)]
+    pub layer_types: Option<Vec<String>>,
 }
 
 impl ModelConfig {
@@ -81,7 +103,48 @@ impl ModelConfig {
     }
 
     pub fn head_dim(&self) -> usize {
-        self.hidden() / self.num_heads()
+        // gpt-oss sets head_dim explicitly (64 != 2880/64). Dense models omit it.
+        self.head_dim.unwrap_or_else(|| self.hidden() / self.num_heads())
+    }
+
+    // ----- MoE (gpt-oss) -----
+
+    /// True for MoE models (have an expert count in config).
+    pub fn is_moe(&self) -> bool {
+        self.num_local_experts.is_some()
+    }
+
+    pub fn num_experts(&self) -> usize {
+        self.num_local_experts.unwrap_or(0)
+    }
+
+    pub fn experts_per_tok(&self) -> usize {
+        self.num_experts_per_tok.unwrap_or(0)
+    }
+
+    /// Clamp bound for gpt-oss SwiGLU (config `swiglu_limit`, default 7.0).
+    pub fn swiglu_limit(&self) -> f32 {
+        self.swiglu_limit.unwrap_or(7.0) as f32
+    }
+
+    /// Whether layer `i` uses sliding-window attention. gpt-oss alternates per
+    /// `layer_types`; if that's absent but `sliding_window` is set, fall back to
+    /// the common "every other layer" pattern (even = sliding). Dense → false.
+    pub fn layer_uses_sliding_window(&self, layer: usize) -> bool {
+        if self.sliding_window.is_none() {
+            return false;
+        }
+        match &self.layer_types {
+            Some(types) => types
+                .get(layer)
+                .map(|t| t.contains("sliding"))
+                .unwrap_or(false),
+            None => layer % 2 == 0,
+        }
+    }
+
+    pub fn sliding_window(&self) -> Option<usize> {
+        self.sliding_window
     }
 
     pub fn ln_eps(&self) -> f32 {

docs/19-moe-gpt-oss.md

@@ -59,6 +59,41 @@ o = (scores @ v) -> merge heads -> @Wo + bo
 与 Qwen3 的新增点：MoE FFN、MXFP4 反量化、attention sinks（softmax 多一列再丢）、
 交替 sliding window、q/k/v/o bias、head_dim=64、clamped `(up+1)*glu`、rope_theta=150000。
 
+### 实测张量布局（layer 0，已用 `tools/mxfp4_probe.py` 核对）
+```
+self_attn.q_proj.weight [4096,2880] +bias[4096]   # 64 heads*64
+self_attn.k_proj.weight [512,2880]  +bias[512]    # 8 kv*64
+self_attn.v_proj.weight [512,2880]  +bias[512]
+self_attn.o_proj.weight [2880,4096] +bias[2880]
+self_attn.sinks [64]                               # 每 q-head 一个标量（BF16）
+input_layernorm.weight [2880]; post_attention_layernorm.weight [2880]
+mlp.router.weight [32,2880] +bias[32]
+mlp.experts.gate_up_proj_blocks [32,5760,90,16] U8 + _scales [32,5760,90] U8 + _bias[32,5760] BF16
+mlp.experts.down_proj_blocks    [32,2880,90,16] U8 + _scales [32,2880,90] U8 + _bias[32,2880] BF16
+# 全局: model.embed_tokens.weight, model.norm.weight, lm_head.weight (BF16)
+```
+MXFP4 打包：`[..., nblk=90, 16]` U8，每 16 字节 = 32 个 FP4 码（低 nibble=偶 idx，高 nibble=奇 idx），
+每 block 一个 E8M0 scale；`90*32 = 2880 = 输入(hidden)维`。即 gate_up 每 expert 权重逻辑 shape
+`[5760 out, 2880 in]`（**已转置存储**：行=out，列=in，与 HF `nn.Linear` 一致 `y=x·Wᵀ`）。
+
+### RoPE（**rotate_half，非 interleave**）
+```
+dim = head_dim = 64; base = rope_theta = 150000
+inv_freq = 1 / base^(arange(0,64,2)/64)        # 32 项
+freqs = pos ⊗ inv_freq                          # [S, 32]；cos/sin = cos(freqs)/sin(freqs)  (不 doubling)
+# 应用: x=[.., 64], first=x[:32], second=x[32:]
+#   out_first  = first*cos - second*sin
+#   out_second = second*cos + first*sin
+```
+> ⚠️ 与 Qwen3 的 RoPE kernel（interleave）不同 —— gptoss 走 rotate_half。需单独处理。
+
+### Decoder layer（pre-norm 残差，结构同 Qwen3）
+```
+h = x + attn(input_norm(x))            # attn 含 sinks/bias/滑窗
+out = h + moe(post_norm(h))            # moe = router + top4 experts 加权和
+```
+最终：`logits = lm_head(norm(h_last))`。无 q_norm/k_norm（与 Qwen3 不同，gptoss 没有）。
+
 ## 3. MXFP4 反量化（expert 权重）
 
 expert 张量名：`model.layers.{i}.mlp.experts.gate_up_proj_blocks/_scales`、

tools/gptoss_dequant.py

@@ -0,0 +1,79 @@
+#!/usr/bin/env python3
+"""Dequantize gpt-oss-20b MXFP4 expert weights -> a plain BF16 safetensors dir.
+
+Only the expert MLPs are MXFP4 (`*_blocks` uint8 packed 4-bit + `*_scales` uint8
+E8M0, block=32); everything else is already BF16. We decode experts to BF16 and
+re-emit a standard HF-format dir so xserv's normal safetensors loader reads it
+(keeps the first MoE pass free of any MXFP4 code in Rust).
+
+Fused expert outputs (per layer i), matching HF `GptOssExperts` param shapes:
+  model.layers.{i}.mlp.experts.gate_up_proj   [E, hidden, 2*inter]  bf16
+  model.layers.{i}.mlp.experts.down_proj      [E, inter, hidden]    bf16
+  (*_bias tensors pass through unchanged)
+
+NOTE on transpose: the MXFP4 `_blocks` decode to [E, OUT, IN] (out-major, the
+contraction dim = nblk*32 last). HF's nn.Parameter for these is [E, IN, OUT]
+(it does `x @ gate_up_proj`). We emit [E, IN, OUT] (transpose last two dims) so
+the names/shapes match HF exactly and xserv can treat them uniformly.
+
+Run on the GPU host (torch + the model + disk):
+  python3 tools/gptoss_dequant.py /opt/wjh/models/gpt-oss-20b /opt/wjh/models/gpt-oss-20b-bf16
+"""
+import sys, os, json, glob
+import torch
+from safetensors import safe_open
+from safetensors.torch import save_file
+
+# FP4 E2M1 code -> value (OCP MX). 16 entries.
+FP4 = torch.tensor(
+    [0.0, 0.5, 1.0, 1.5, 2.0, 3.0, 4.0, 6.0,
+     -0.0, -0.5, -1.0, -1.5, -2.0, -3.0, -4.0, -6.0], dtype=torch.float32)
+
+def dequant(blocks: torch.Tensor, scales: torch.Tensor) -> torch.Tensor:
+    """blocks uint8 [..., nblk, 16], scales uint8 [..., nblk] -> bf16 [..., nblk*32]."""
+    blocks = blocks.to(torch.int64)
+    lo = blocks & 0xF
+    hi = (blocks >> 4) & 0xF
+    codes = torch.stack([lo, hi], dim=-1).reshape(*blocks.shape[:-1], blocks.shape[-1] * 2)
+    vals = FP4[codes]                                   # [..., nblk, 32] f32
+    scale = torch.exp2(scales.to(torch.float32) - 127.0)  # [..., nblk]
+    out = vals * scale[..., None]
+    out = out.reshape(*out.shape[:-2], out.shape[-2] * out.shape[-1])
+    return out.to(torch.bfloat16)
+
+def main():
+    src, dst = sys.argv[1], sys.argv[2]
+    os.makedirs(dst, exist_ok=True)
+    wm = json.load(open(os.path.join(src, "model.safetensors.index.json")))["weight_map"]
+
+    shards = {}
+    for name, shard in wm.items():
+        shards.setdefault(shard, []).append(name)
+
+    out = {}
+    for shard in sorted(shards):
+        h = safe_open(os.path.join(src, shard), framework="pt")
+        keys = set(h.keys())
+        for name in shards[shard]:
+            if name.endswith("_scales"):
+                continue
+            if name.endswith("_blocks"):
+                base = name[:-len("_blocks")]                 # ...gate_up_proj / ...down_proj
+                sc = base + "_scales"
+                sc_h = h if sc in keys else safe_open(os.path.join(src, wm[sc]), framework="pt")
+                deq = dequant(h.get_tensor(name), sc_h.get_tensor(sc))  # [E, OUT, IN]
+                out[base] = deq.transpose(1, 2).contiguous()           # [E, IN, OUT] (HF param layout)
+                print("dequant", base, tuple(out[base].shape), flush=True)
+            else:
+                out[name] = h.get_tensor(name)                # already bf16/other; pass through
+
+    save_file(out, os.path.join(dst, "model.safetensors"), metadata={"format": "pt"})
+    for f in glob.glob(os.path.join(src, "*.json")) + glob.glob(os.path.join(src, "*.jinja")):
+        b = os.path.basename(f)
+        if b == "model.safetensors.index.json":
+            continue
+        open(os.path.join(dst, b), "wb").write(open(f, "rb").read())
+    print("DEQUANT_DONE ->", dst, flush=True)
+
+if __name__ == "__main__":
+    main()

tools/mxfp4_probe.py

@@ -0,0 +1,82 @@
+#!/usr/bin/env python3
+"""P19.1 — inspect gpt-oss-20b safetensors + validate MXFP4 dequant (CPU only).
+
+Run: python3 tools/mxfp4_probe.py /path/to/gpt-oss-20b
+
+Does three things:
+  1. List the layer-0 tensor names, shapes, dtypes (esp. expert _blocks/_scales).
+  2. Dequantize one expert's gate_up_proj from MXFP4 -> fp32 with our own LUT/decode.
+  3. Print stats so we can eyeball sanity (range, mean, a few values).
+
+MXFP4 (OCP microscaling) as used by gpt-oss:
+  - elements are FP4 E2M1 (4-bit), packed 2-per-byte.
+  - every block of 32 consecutive elements shares one E8M0 scale (8-bit exponent).
+  - value = fp4_e2m1_lut[code] * 2**(scale_e8m0 - 127)
+The `_blocks` tensor holds the packed 4-bit codes; `_scales` holds the per-block
+E8M0 exponents (uint8). We confirm shapes line up (last dim of blocks * 2 / ... )
+and that decoded values are in a sane range.
+"""
+import sys, json, os
+import numpy as np
+
+# FP4 E2M1 code -> value lookup (OCP MX spec). 16 codes: sign(1) exp(2) mant(1).
+# Values: 0, 0.5, 1, 1.5, 2, 3, 4, 6 and their negatives.
+FP4_E2M1 = np.array(
+    [0.0, 0.5, 1.0, 1.5, 2.0, 3.0, 4.0, 6.0,
+     -0.0, -0.5, -1.0, -1.5, -2.0, -3.0, -4.0, -6.0],
+    dtype=np.float32,
+)
+
+def dequant_mxfp4(blocks: np.ndarray, scales: np.ndarray) -> np.ndarray:
+    """gpt-oss layout:
+      blocks: uint8 [..., nblk, 16]  -> each 16-byte row = 32 FP4 codes (2/byte)
+      scales: uint8 [..., nblk]      -> one E8M0 exponent per block
+    Returns fp32 [..., nblk*32] (the input/contraction dim)."""
+    lo = blocks & 0x0F
+    hi = (blocks >> 4) & 0x0F
+    # within a byte, low nibble is element 2i, high nibble is element 2i+1
+    codes = np.empty(blocks.shape[:-1] + (blocks.shape[-1] * 2,), dtype=np.uint8)
+    codes[..., 0::2] = lo
+    codes[..., 1::2] = hi                          # [..., nblk, 32]
+    vals = FP4_E2M1[codes]                         # [..., nblk, 32]
+    scale_f = np.power(2.0, scales.astype(np.float32) - 127.0)  # [..., nblk]
+    out = vals * scale_f[..., None]                # [..., nblk, 32]
+    return out.reshape(out.shape[:-2] + (out.shape[-2] * out.shape[-1],))  # [..., nblk*32]
+
+def main():
+    d = sys.argv[1] if len(sys.argv) > 1 else "/home/gahow/models/gpt-oss-20b"
+    from safetensors import safe_open
+    idx = json.load(open(os.path.join(d, "model.safetensors.index.json")))
+    wm = idx["weight_map"]
+    l0 = {k: v for k, v in wm.items() if "layers.0." in k}
+    print("=== layer 0 tensors ===")
+    # open each shard lazily
+    handles = {}
+    def get(name):
+        shard = wm[name]
+        if shard not in handles:
+            handles[shard] = safe_open(os.path.join(d, shard), framework="numpy")
+        return handles[shard]
+    for k in sorted(l0):
+        h = get(k)
+        t = h.get_slice(k)
+        print(f"  {k.replace('model.layers.0.','L0.')}  shape={t.get_shape()} dtype={t.get_dtype()}")
+
+    # find expert gate_up blocks/scales
+    gu_b = next((k for k in l0 if "gate_up_proj_blocks" in k), None)
+    gu_s = next((k for k in l0 if "gate_up_proj_scales" in k), None)
+    if gu_b and gu_s:
+        print(f"\n=== dequant {gu_b.split('.')[-1]} (expert 0) ===")
+        blocks = get(gu_b).get_tensor(gu_b)
+        scales = get(gu_s).get_tensor(gu_s)
+        print("  blocks", blocks.shape, blocks.dtype, " scales", scales.shape, scales.dtype)
+        b0 = blocks[0]; s0 = scales[0]   # expert 0: blocks [5760,90,16], scales [5760,90]
+        deq = dequant_mxfp4(b0.astype(np.uint8), s0.astype(np.uint8))  # -> [5760, 2880]
+        print("  expert0 dequant shape", deq.shape, "(expect [5760, 2880] = [2*inter, hidden])",
+              "\n  min %.4f max %.4f mean %.4f std %.4f" % (deq.min(), deq.max(), deq.mean(), deq.std()))
+        print("  row0 first 8 vals:", np.round(deq[0, :8], 4))
+    else:
+        print("\n(no gate_up_proj_blocks/_scales found — check tensor names above)")
+
+if __name__ == "__main__":
+    main()