mirror of
https://github.com/deepseek-ai/DeepInfer.git
synced 2026-11-01 01:06:23 +08:00
28fb291db1
No behaviour change. `CONTRIBUTING.md` requires `ruff check --no-cache` and `ruff format --check --no-cache` to pass, and the SM90 branch predates that requirement: 72 findings, all in the files added here. Most of it is mechanical (`ruff format`), plus: - annotations for the helper signatures (`torch.Tensor` / `int` / `Generator` / return types); hooks are typed as `Callable[[nn.Module, tuple, object], None]` rather than left bare; - two `if`/`else` assignments become ternaries (SIM108), and the index read in `h20_real_requant_check.py` uses a context manager (SIM115). Also corrects the `sparse_mla_fwd` docstring, which still described the old `0 <= idx <= q_position` masking contract. Passing `is_causal=True` there is what produced the all-NaN layer 0: indices address the assembled `[swa cache | current swa | compressed]` buffer, so the row number is not a valid causal bound and the engine must pass False.
33 lines
1.2 KiB
Python
33 lines
1.2 KiB
Python
"""Validate the SM90 dense requant on a real checkpoint tensor (wq_a + shared expert)."""
|
|
|
|
import json
|
|
|
|
import torch
|
|
from safetensors import safe_open
|
|
|
|
from deepinfer.h20_ops.quant import requant_dense_fp8, ue8m0_to_fp32
|
|
|
|
root = "/data/models/DeepSeek-V4.1-Flash-1017"
|
|
with open(f"{root}/model.safetensors.index.json") as f:
|
|
idx = json.load(f)["weight_map"]
|
|
|
|
targets = [
|
|
"layers.0.attn.wq_a.weight",
|
|
"layers.0.ffn.shared_experts.w1.weight",
|
|
]
|
|
for name in targets:
|
|
with safe_open(f"{root}/{idx[name]}", framework="pt", device="cpu") as f:
|
|
w = f.get_tensor(name)
|
|
s = f.get_tensor(name.replace(".weight", ".scale"))
|
|
dev = "cuda"
|
|
w, s = w.to(dev), s.to(dev)
|
|
ref = w.to(torch.float32) * ue8m0_to_fp32(s).repeat_interleave(32, 0).repeat_interleave(32, 1)
|
|
w8, sf = requant_dense_fp8(w, ue8m0_to_fp32(s), dst_block=128)
|
|
rec = w8.to(torch.float32) * sf.repeat_interleave(128, 0).repeat_interleave(128, 1)
|
|
err = (rec - ref).abs()
|
|
print(
|
|
f"{name}: {tuple(w.shape)} sf {tuple(s.shape)} -> {tuple(sf.shape)} | "
|
|
f"max_abs {err.max().item():.5g} rel {(err.max() / ref.abs().max()).item():.3e}"
|
|
)
|
|
print("REAL_WEIGHT_REQUANT_OK")
|