Rebases Relax's sglang patch onto a new sglang release and updates the training Docker image.
npx skills add https://github.com/redai-infra/Relax --skill sglang-upgrade
Please help me install this Agent Skill. Skill: sglang-upgrade Source: https://github.com/redai-infra/Relax/tree/main/skills/sglang-upgrade Command: npx skills add https://github.com/redai-infra/Relax --skill sglang-upgrade
Upgrading the sglang version behind Relax's training Docker image is dominated by one painful task: rebasing the large private patch docker/patch/latest/sglang.patch onto a new upstream release without losing Relax-specific changes or re-porting code that upstream has already absorbed.
BASE_IMAGE ARG in docker/Dockerfile, not by requirements files, and how source overlays affect it.scripts/classify_patch.sh.git apply --check on a clean worktree.sglang.patch, update the Dockerfile and requirements, and record keep/drop decisions in an upgrade plan document.Upgrade Relax's sglang base image to the new tag vX.Y.Z and rebase the sglang patch accordingly.
High-intent search queries and answers about installing and using this skill.
Change the BASE_IMAGE ARG tag in docker/Dockerfile, since the sglang version is pinned there rather than in requirements files. Then rebase the private sglang.patch onto the new release and validate it with git apply --check on a clean worktree of the target tag.
Use an already-rebased reference patch (THUDM/slime maintains per-version sglang patches) as a skeleton, then classify your private hunks into identical, shared, and unique files. Port only the deltas that upstream has not already absorbed, and verify with git apply --check.
Check out the target release source tree and grep for the key symbols or function names your patch adds. If the functionality exists upstream, drop your hunk; model-layer features like NSA or BF16 MoE kernels are frequently absorbed by upstream sglang.
An rsync source overlay replaces only the source files, not the .dist-info metadata that setuptools_scm froze at image build time. The installed package metadata therefore disagrees with the actual source, so trust the Dockerfile base image tag instead.
Patch @@ line-number offsets pollute change comparisons, so compare only real +/- lines. Also avoid copying slime-private changes meant for their custom base image, and confirm Relax actually references a private feature before deleting it.
GregsGreyCodeCoordinate end-to-end RL training for LLMs with Megatron and SGLang.
NVIDIADerives per-rank GEMM and fused-MoE shapes from Hugging Face models and runs FlashInfer microbenchmarks.
hey-kongQuantize diffusion transformers with NVIDIA ModelOpt to generate FP8 or NVFP4 checkpoints for SGLang Diffusion.
majunbaoInspect, patch, validate, publish, and confirm GHSA advisory private-fork state.