Skip to content

[Kernel] Hotfix: update sgl-kernel imports of relocated fp8_kernel (RFC #29630 #30784) - #31089

Merged
BBuf merged 1 commit into
sgl-project:mainfrom
BBuf:bbuf/hotfix-sgl-kernel-fp8-import
Jul 14, 2026
Merged

BBuf merged 1 commit into
sgl-project:mainfrom
BBuf:bbuf/hotfix-sgl-kernel-fp8-import

Conversation

@BBuf

@BBuf BBuf commented Jul 14, 2026

Copy link
Copy Markdown
Collaborator

Problem

RFC #29630 Phase 2.5 PR #30784 relocated fp8_kernel from
sglang.srt.layers.quantization.fp8_kernelsglang.kernels.ops.quantization.fp8_kernel
(move + rewrite all imports, no compatibility shim).

The import rewrite covered python/sglang/, test/ and benchmark/, but missed the
sgl-kernel/ subtree
. Three sgl-kernel test/benchmark modules still import the old
path, so after #30784 merged they fail at collection with:

ModuleNotFoundError: No module named 'sglang.srt.layers.quantization.fp8_kernel'

This breaks the Run sgl-kernel unit tests on B200 job on main for every PR (e.g. it
surfaced on an unrelated PR's CI), not just the migration PRs.

Fix

Rewrite the three stale imports to the new path:

File lines
sgl-kernel/tests/test_per_token_group_quant_8bit.py 14, 17
sgl-kernel/benchmark/bench_per_token_group_quant_8bit.py 11, 14, 17
sgl-kernel/benchmark/bench_fp8_blockwise_gemm.py 23

sglang.srt.layers.quantization.fp8_kernelsglang.kernels.ops.quantization.fp8_kernel

All imported symbols — per_token_group_quant_8bit, sglang_per_token_group_quant_8bit,
w8a8_block_fp8_matmul_triton, create_per_token_group_quant_fp8_output_scale — exist at
the new path (the relocation was byte-identical). Pure import-path change, no behavior change.

Verified: no srt.layers.quantization.fp8_kernel references remain anywhere in sgl-kernel/,
and a repo-wide sweep confirms these three were the only stale old-path Python imports left
(the other relocated quant modules — int8_kernel / mxfp8 / awq / nvfp4 — have no old-path
consumers in the skipped dirs).

🤖 Generated with Claude Code


CI States

Latest PR Test (Base): ❌ Run #29296300849
Latest PR Test (Extra): ❌ Run #29296408618

RFC sgl-project#29630 Phase 2.5 PR sgl-project#30784 relocated fp8_kernel from
sglang.srt.layers.quantization.fp8_kernel to
sglang.kernels.ops.quantization.fp8_kernel (no shim), but the import
rewrite only covered python/sglang, test/ and benchmark/ — it missed the
sgl-kernel/ subtree, so three sgl-kernel test/benchmark modules still
imported the old path and now fail collection (ModuleNotFoundError),
breaking the 'sgl-kernel unit tests on B200' CI job on main.

Rewrite the three stale imports to the new path; all imported symbols
(per_token_group_quant_8bit, sglang_per_token_group_quant_8bit,
w8a8_block_fp8_matmul_triton, create_per_token_group_quant_fp8_output_scale)
exist there.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
@gemini-code-assist

Copy link
Copy Markdown
Contributor

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

@github-actions github-actions Bot added quant LLM Quantization sgl-kernel labels Jul 14, 2026
@BBuf BBuf added the run-ci label Jul 14, 2026
@BBuf
BBuf merged commit 4c99731 into sgl-project:main Jul 14, 2026
123 of 140 checks passed
Chronostasys pushed a commit to MindLab-Research/sglang that referenced this pull request Aug 24, 2026
jakki-amd pushed a commit to jakki-amd/sglang that referenced this pull request Sep 9, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant