Gahow Wang
c7d0750c32
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>
80 lines
3.6 KiB
Python
80 lines
3.6 KiB
Python
#!/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()
|