Skip to content

Update batch size limitation of dsv3_router_gemm kernel to 16 #8051

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
ispobock merged 2 commits into from
Aug 1, 2025
Merged

Update batch size limitation of dsv3_router_gemm kernel to 16 #8051

Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
f45e066
Apply updated dsv3_router_gemm kernel
Fridge003 Jul 15, 2025
724a816
Merge branch 'main' into apply_router
Fridge003 Jul 26, 2025
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: 3 additions & 4 deletions python/sglang/srt/models/deepseek_v2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -243,14 +243,13 @@ def forward(self, hidden_states):
# NOTE: For some unknown reason, router_gemm seems degrade accept length.
if (
_is_cuda
and not self.is_nextn
and hidden_states.shape[0] < 4
and hidden_states.shape[0] <= 16
and hidden_states.shape[1] == 7168
and self.weight.shape[0] == 256
and _device_sm >= 90
):
logits = dsv3_router_gemm(hidden_states, self.weight).to(
hidden_states.dtype
logits = dsv3_router_gemm(
hidden_states, self.weight, out_dtype=torch.bfloat16
)
else:
logits = F.linear(hidden_states, self.weight, None)
Expand Down