Skip to content

fix: restore FP4 deepgemm path for Blackwell broken by #23686 #23948

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
Fridge003 merged 2 commits into from
Apr 28, 2026
Merged

fix: restore FP4 deepgemm path for Blackwell broken by #23686 #23948

Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
018b03b
fix: restore FP4 deepgemm path for Blackwell broken by #23686
Apr 28, 2026
d3e63aa
refactor: address review feedback - simplify marlin branch and add ba…
Apr 28, 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
16 changes: 4 additions & 12 deletions python/sglang/srt/layers/quantization/fp8.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -932,20 +932,12 @@ def process_weights_after_loading_block_quant(self, layer: Module) -> None:
will_use_deepgemm = self.is_deepgemm_moe_runner_backend_enabled()

if self.is_fp4_expert:
if get_moe_runner_backend().is_marlin():
layer.w13_weight.data = layer.w13_weight.data.view(torch.int8)
layer.w2_weight.data = layer.w2_weight.data.view(torch.int8)
elif not get_moe_runner_backend().is_flashinfer_mxfp4():
raise NotImplementedError(
"DeepSeekV4 FP4 experts now require a native FP4 MoE backend. "
"Use `--moe-runner-backend marlin` on Hopper or "
"`--moe-runner-backend flashinfer_mxfp4` when available."
)
layer.w13_weight.data = layer.w13_weight.data.view(torch.int8)
layer.w2_weight.data = layer.w2_weight.data.view(torch.int8)

if get_moe_runner_backend().is_marlin():

@ispobock ispobock Apr 28, 2026

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

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

we don't need the pass branch here. just comment here if you want to notice marlin.

@ispobock ispobock Apr 28, 2026

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

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

I think we also need to add comments to clarify the three moe backends we supported and which branch they should go.

pass # Hopper w4a16: int8 view is sufficient
else:
layer.w13_weight.data = layer.w13_weight.data.view(torch.int8)
layer.w2_weight.data = layer.w2_weight.data.view(torch.int8)

if envs.SGLANG_OPT_USE_DEEPGEMM_MEGA_MOE.get():
from sglang.srt.models.deepseek_v4 import (
build_mega_moe_experts_weights,
Expand Down
Loading