Skip to content

[DRAFT][DeepSeek-V4] Emit fp32 directly in linear_bf16_fp32 (drop .float() copy) #3

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

Closed
AMD-yanfeiwang wants to merge 1 commit into from
Closed

[DRAFT][DeepSeek-V4] Emit fp32 directly in linear_bf16_fp32 (drop .float() copy) #3

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
9 changes: 8 additions & 1 deletion python/sglang/jit_kernel/dsv4/gemm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15,7 +15,14 @@

def linear_bf16_fp32(x: torch.Tensor, y: torch.Tensor) -> torch.Tensor:
if _use_aiter:
return tgemm.mm(x, y, otype=x.dtype).float()
# Request fp32 output directly instead of computing a bf16 GEMM and then
# casting with `.float()`. The explicit cast shows up as a separate
# `bfloat16tofloat32_copy` kernel per call in the DSV4 indexer
# (compute_kv_score) path; emitting fp32 from the GEMM epilogue removes
# that copy and also avoids rounding the result to bf16 first.
# Requires AITER to expose a native fp32-output tuned kernel for these
# shapes (companion AITER change); otherwise tgemm falls back internally.
return tgemm.mm(x, y, otype=torch.float32)
elif _linear_bf16_fp32_algo == "deep_gemm":
z = torch.empty(x.size(0), y.size(0), dtype=torch.float32, device=x.device)
deep_gemm_wrapper.gemm_nt_bf16bf16f32(x, y, z)
Expand Down
Loading