Skip to content

fix(models): preserve glm4 router correction bias in fp32 #469

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
jason-fxz merged 1 commit into from
Sep 22, 2026
Merged

fix(models): preserve glm4 router correction bias in fp32 #469

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 2 additions & 4 deletions python/freetoken/models/glm4_moe/moe.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -32,10 +32,8 @@ def __init__(self, config: ModelConfig, layer_id: int, *, prefix: str = ""):
self.topk_group = config.topk_group

self.gate = LinearReplicated(config.hidden_size, config.num_experts, has_bias=False)
# DeepSeek-style selection bias; a registered buffer in HF (kept fp32 there). We
# store it in the model's bf16 dtype and upcast at use, which is exact enough for
# the argmax-style top-k selection.
self.e_score_correction_bias = torch.empty(config.num_experts)
# Keep selection bias in fp32: rounding can change the selected experts.
self.e_score_correction_bias = torch.empty(config.num_experts, dtype=torch.float32)

# The offload cache indexes experts by *MoE* layer (global layer minus
# first_k_dense_replace), matching how the loader packs the expert banks. The
Expand Down
4 changes: 2 additions & 2 deletions python/freetoken/models/glm4_moe/weight.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -168,11 +168,11 @@ def _iter_resident_weights(reader, config, primary) -> Iterator[tuple[str, torch
for proj in ("gate_proj", "up_proj", "down_proj"):
yield from _iter_nvfp4_resident(reader, f"{m}.{proj}", f"{m}.{proj}")
else:
# router (bf16 gate + fp32 selection bias -> bf16) and shared expert.
# router (bf16 gate + fp32 selection bias) and shared expert.
yield f"{m}.gate.weight", reader.get(f"{m}.gate.weight")
yield (
f"{m}.e_score_correction_bias",
reader.get(f"{m}.gate.e_score_correction_bias").to(torch.bfloat16),
reader.get(f"{m}.gate.e_score_correction_bias").to(torch.float32),
)
s = f"{m}.shared_experts"
for proj in ("gate_proj", "up_proj", "down_proj"):
Expand Down