Skip to content

Fix SGL_CUDA_ARCH mismatch in NVFP4 JIT compilation #20965

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
csahithi wants to merge 1 commit into from
Closed

Fix SGL_CUDA_ARCH mismatch in NVFP4 JIT compilation #20965

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
15 changes: 1 addition & 14 deletions python/sglang/jit_kernel/nvfp4.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -93,20 +93,7 @@ def _get_nvfp4_cuda_arch_list() -> str:
)
# NVFP4 kernels use architecture-family-specific instructions and must be
# compiled for `sm_*a` targets (e.g. sm_100a), not plain sm_100.
archs = [f"{major}.{minor}a"]
cuda_major, _cuda_minor = _parse_cuda_version()
if cuda_major >= 13 and "10.3a" not in archs:
# Match sgl-kernel AOT fatbin behavior on CUDA 13+ for Blackwell.
archs.append("10.3a")
# Preserve order while de-duplicating.
seen = set()
ordered_archs: list[str] = []
for arch in archs:
if arch in seen:
continue
seen.add(arch)
ordered_archs.append(arch)
return " ".join(ordered_archs)
return f"{major}.{minor}a"


@contextmanager
Expand Down
Loading