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>

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()