Skip to content

qwen3_5_moe: run lm_head on sampled rows only (fixes 32 GB first-prefill OOM) #342

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

Open
chrisqianz wants to merge 1 commit into
base: main
Choose a base branch
from
+9 −2
Open

qwen3_5_moe: run lm_head on sampled rows only (fixes 32 GB first-prefill OOM) #342

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
11 changes: 9 additions & 2 deletions python/freetoken/models/qwen3_5_moe/model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -110,8 +110,15 @@ def __init__(self, config: ModelConfig):
super().__init__()

def forward(self) -> torch.Tensor:
output = self.model.forward(get_global_ctx().batch.input_ids)
return self.lm_head.forward(output)
ctx = get_global_ctx()
output = self.model.forward(ctx.batch.input_ids)
# Project only the rows the sampler reads. engine.forward_batch slices
# logits[:batch.size] (one row per request, the first rows of the forward
# window); the remaining rows are overlap context. Running the vocab GEMM on
# the whole window allocates M x vocab bf16 (a default 8192-token chunk at
# Qwen3.8's 248k vocab is 3.79 GiB -- a guaranteed OOM on 32 GB cards)
# and spends ~M x vocab x hidden FLOPs whose results nobody consumes.
return self.lm_head.forward(output[: ctx.batch.size])


__all__ = ["Qwen3_5MoEForCausalLM"]