docker/scripts: remove dead GB10 / DGX Spark support - #35
Merged
Conversation
The GB10 flow is the last remaining SGLang user in the repo after PR #18 ripped the SGLang runtime out of vime. It was a one-host port (NVIDIA DGX Spark, aarch64 + sm_121a) that never made it past "validates against a smoke test"; not in any active CI matrix and not referenced from any production training script. Per user direction ("都不用了"), removing it as a unit so the repo stops carrying a parallel SGLang-on-aarch64 build path. Files removed: - docker/Dockerfile.gb10 NGC vllm:26.03-py3 base + source-built sgl-kernel for sm_121a + sglang==0.5.9. The whole point of this image was getting SGLang onto an arch where no prebuilt sgl-kernel wheel existed. - docker/NOTES_GB10.md 99-line debug log: 15 documented blockers from porting SGLang to GB10 (cicc OOM, missing cuda_profiler_api.h shim, cuDNN engines_precompiled, etc.). Useful only as long as the Dockerfile.gb10 above exists. - docker/patch/gb10/cuda_profiler_api.h 20-line shim header for CUDA 13's removal of the public profiler header. TE 2.10 #includes it without calling the APIs. - docker/patch/gb10/patch_sgl_kernel.py Inserts SGL_KERNEL_GB10_ONLY CMake option into sgl-kernel's build. - docker/patch/gb10/sgl-kernel-arch.patch Same patch in textual diff form. - scripts/run-qwen2.5-0.5B-gb10-smoke.sh The smoke test header explicitly says "Used only to validate the GB10 port; not a training recipe." With the GB10 image gone this script has no environment to run in. No other file in the tree references gb10/GB10 after this. The mainline CUDA path (docker/Dockerfile + build_conda.sh) and the AMD path (docker/Dockerfile.rocm + docker/amd_patch/) are untouched. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Code Review
This pull request removes several files and scripts that were previously introduced to support porting and running the "slime" framework on NVIDIA DGX Spark (GB10) hardware. The deleted files include a specialized Dockerfile, porting notes, a CUDA 13 header shim, a CMake patching script, a patch file for sgl-kernel, and a GRPO smoke test script. I have no feedback to provide as there are no active review comments.
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
1 participant
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.
Summary
Removes the NVIDIA DGX Spark (GB10 / sm_121a, aarch64) port as a unit. The port has been static for ~5 weeks since landing in #1835, has no follow-up activity from any contributor, no CI coverage, and exists primarily because GB10 required a from-source sgl-kernel build for sm_121a that no published wheel covered.
With the SGLang runtime removal underway (
gcl/clean-sglang/ PR #18 follow-on), the GB10 Dockerfile's central motivation — getting sgl-kernel onto sm_121a aarch64 — disappears. Rather than carry a half-supported port that nobody is exercising, this PR removes it cleanly. Git history (f44af13/ #1835) preserves the full port for anyone wishing to revive it as a separate effort.Files removed (538 lines):
docker/Dockerfile.gb10— NGCvllm:26.03-py3base + source-built sgl-kernel + sglang 0.5.9docker/NOTES_GB10.md— 99-line porting journal (15 documented blockers)docker/patch/gb10/cuda_profiler_api.h— CUDA 13 profiler header shimdocker/patch/gb10/patch_sgl_kernel.py— sgl-kernel CMake injectiondocker/patch/gb10/sgl-kernel-arch.patch— same in textual diff formscripts/run-qwen2.5-0.5B-gb10-smoke.sh— GB10-only smoke test (the script's own header notes it is "not a training recipe")No other file in the tree references
gb10orGB10after this change. Mainline CUDA (docker/Dockerfile) and ROCm (docker/Dockerfile.rocm) paths are untouched.Test plan
git grep -i gb10returns no matchesdocker/Dockerfilebuilds unchanged (no references to deleted files)🤖 Generated with Claude Code