Skip to content

Tiny fix trtllm_fp8_per_tensor_scale_moe_wrapper router_logits dtype #22006

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
Qiaolin-Yu merged 2 commits into from
Apr 6, 2026
Merged

Tiny fix trtllm_fp8_per_tensor_scale_moe_wrapper router_logits dtype #22006

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
abf659e
upd
Qiaolin-Yu Apr 3, 2026
7dd86f6
Merge branch 'main' into fix_per_tensor
Qiaolin-Yu Apr 3, 2026
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
9 changes: 8 additions & 1 deletion python/sglang/srt/layers/moe/moe_runner/flashinfer_trtllm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -470,8 +470,15 @@ def fused_experts_none_to_flashinfer_trtllm_fp8(
# Move kernel call outside context manager to avoid graph breaks
# during torch.compile for piecewise cuda graph.
# Use custom op wrapper for torch.compile compatibility.

# The DeepSeekV3 routing method requires float32 router logits.
Copy link
Copy Markdown
Collaborator

@nvpohanh nvpohanh Apr 6, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@leejnau @trevor-m is this true? If so, why didn't we run into issues before?

Copy link
Copy Markdown
Collaborator

@nvpohanh nvpohanh Apr 6, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Maybe will be fixed by flashinfer-ai/flashinfer#2993 ?

Copy link
Copy Markdown
Collaborator

@trevor-m trevor-m Apr 6, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The path for block scale had this fix already, I think we didn't use per tensor scaling before?

Copy link
Copy Markdown
Collaborator

@nvpohanh nvpohanh Apr 7, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

got it. we have never run DSV3/R1 with per-tensor FP8 before.

if routing_method_type == RoutingMethodType.DeepSeekV3:
router_logits = router_logits.to(torch.float32)
else:
router_logits = router_logits.to(torch.bfloat16)

output = trtllm_fp8_per_tensor_scale_moe_wrapper(
routing_logits=router_logits.to(torch.bfloat16),
routing_logits=router_logits,
routing_bias=routing_bias_cast,
hidden_states=a_q,
gemm1_weights=quant_info.w13_weight,
Expand Down
Loading