Skip to content

Fix SM120 triton_kernels MXFP4 block_k for GPT-OSS #20040

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
Kangyan-Zhou merged 1 commit into from
Mar 6, 2026
Merged

Fix SM120 triton_kernels MXFP4 block_k for GPT-OSS #20040

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
6 changes: 4 additions & 2 deletions python/sglang/srt/layers/quantization/mxfp4.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -139,8 +139,9 @@ def _swizzle_mxfp4(quant_tensor, scale, num_warps):
from triton_kernels.tensor_details import layout

if is_sm120_supported():
# SM120 (Blackwell desktop) doesn't support persistent kernels / TMA block layout
# Use StridedLayout and disable persistent kernels to avoid assertion errors
# SM120 desktop Blackwell does not support the persistent/TMA MXFP4 path.
# This MXFP4 path uses StridedLayout and the non-persistent kernel with
# block_k=128 so the selected tile stays within the per-block shared-memory budget.
from triton_kernels.tensor_details.layout import StridedLayout

value_layout = StridedLayout
Expand All @@ -149,6 +150,7 @@ def _swizzle_mxfp4(quant_tensor, scale, num_warps):
scale_layout_opts = {}
constraints = {
"is_persistent": False,
"block_k": 128,
"num_stages": 1,
}
opt_flags.update_opt_flags_constraints(constraints)
Expand Down
Loading