Skip to content

Fix DEEPEP BF16 compatibility for Deepseek Style model like GLM 4.5 #8469

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
merged 1 commit into from
Jul 28, 2025
Merged

Fix DEEPEP BF16 compatibility for Deepseek Style model like GLM 4.5 #8469

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
7 changes: 1 addition & 6 deletions python/sglang/srt/layers/moe/ep_moe/layer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -800,11 +800,6 @@ def __init__(
routed_scaling_factor=routed_scaling_factor,
)
self.deepep_mode = deepep_mode
if deep_gemm_wrapper.ENABLE_JIT_DEEPGEMM:
assert self.use_fp8_w8a8, (
"DeepGEMM requires an fp8_w8a8 model; "
"alternatively, you can disable DeepGEMM by turning off the ENABLE_JIT_DEEPGEMM environment variable."
)

# TODO: move to the beginning of the file
from sglang.srt.distributed.parallel_state import get_tp_group
Expand Down Expand Up @@ -897,7 +892,7 @@ def moe_impl(self, dispatch_output: DispatchOutput):
# in forward_aiter, we skip token permutation and unpermutation, which have been fused inside aiter kernel
return self.forward_aiter(dispatch_output)
if dispatch_output.format.is_deepep_normal():
if deep_gemm_wrapper.ENABLE_JIT_DEEPGEMM:
if deep_gemm_wrapper.ENABLE_JIT_DEEPGEMM and self.use_fp8_w8a8:
return self.forward_deepgemm_contiguous(dispatch_output)
else:
return self.forward_normal(dispatch_output)
Expand Down
Loading