Repository navigation
[Fix] docker: Pin sgl-nvidia-mega to 3333e4c so the megakernel starts on main - #46030
Merged
Oasis-Git merged 1 commit intoOct 24, 2026
Merged
Conversation
… on main Root cause: both CUDA Dockerfiles pin sgl-nvidia-mega d47deb33. Since sgl-project#44912/sgl-project#44913 (TopKConfig) and sgl-project#45112 (kernel regrouping), that commit fails while building its MoE slots, so images built from main ship a megakernel that cannot start. Fix: pin 3333e4cf in docker/Dockerfile and docker/Dockerfile.cu134. It is the plugin main commit that follows both changes (sgl-nvidia-mega sgl-project#15, sgl-project#16) and includes the Engram prefetch fix (sgl-project#7). Validation: DeepSeek-V4.1-Flash on 4x GB300, TP4, megakernel enabled: every rank reports the megakernel active; GSM8K-50 49/50; an AgentX c1 trace replay runs with 0 request errors.
weireweire
requested review from
Fridge003,
HaiShaw,
ishandhanani,
ispobock,
sogalin and
yctseng0211
as code owners
October 23, 2026 09:46
Oasis-Git
approved these changes
Oct 24, 2026
4 of 5 tasks
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Motivation
docker/Dockerfileanddocker/Dockerfile.cu134pin sgl-nvidia-megad47deb33. That commit predates two SGLang changes it depends on: the TopKConfig shared-expert field rename (#44912, #44913) and the kernel regrouping (#45112). As a result, it fails while building its MoE slots on current main, and every image built from main ships a megakernel that cannot start under--enable-nvidia-megakernel. This also blocks the DSpark megakernel path from #45450 in public images.#45595 made the same bump and was closed in favor of #45389. #45389 is a larger refactor and currently changes only
docker/rocm.Dockerfile. This PR is the standalone pin bump for the two CUDA Dockerfiles.Modifications
SGL_NVIDIA_MEGA_COMMITto3333e4cfac0f2b15603565d702b4c67538f1628findocker/Dockerfileanddocker/Dockerfile.cu134.3333e4cfis the plugin main commit that follows both SGLang changes (sgl-nvidia-mega Rename image_url to image_file #15, Fix for T4 GPUs #16) and includes the Engram prefetch fix (fix radix cache match #7). It keeps thenvidia_megakernelentry point, so the existing install check is unchanged.Validation
DeepSeek-V4.1-Flash on 4x GB300, TP4/EP1, DSpark block size 5,
--enable-nvidia-megakernel, plugin at this commit:NVIDIA megakernel: activefor the target and the draft.pre-commit run --all-filespasses.CI States
Latest PR Test (Base): ✅ Run #39934258095
Latest PR Test (Extra): ❌ Run #39934258045
Latest PR Test (AMD ROCm 10): ➖ No AMD PR run found for this commit.