Skip to content

[AMD][CI] Retire the ROCm 7.0 kernel wheel - #38767

Merged
bingxche merged 1 commit into
mainfrom
cursor/drop-rocm700-image-build-and-ci-d06d
Sep 11, 2026
Merged

bingxche merged 1 commit into
mainfrom
cursor/drop-rocm700-image-build-and-ci-d06d

Conversation

@michaelzhang-ai

@michaelzhang-ai michaelzhang-ai commented Sep 10, 2026

Copy link
Copy Markdown
Collaborator

Motivation

#38763 added a ROCm 10 kernel wheel, making the ROCm wheel matrix ["700", "720", "1000"]. rocm700 is now the only leg with no CI behind it, and it is the last ROCm 7.0 artifact SGLang still publishes on every kernel version bump.

Scope note: this PR was originally a broader ROCm 7.0 retirement (workflows, nightly image build, release image matrix). #38632 supersedes all of that with a better end state — it consolidates the AMD workflows back onto a single pr-test-amd.yml / nightly-test-amd.yml pair and retires ROCm 7.0 CI as a consequence. Rather than conflict with it on the same four workflow files, this PR is now reduced to the one ROCm 7.0 surface #38632 does not touch: the kernel wheel.

Modifications

Retire the ROCm 7.0 kernel wheel end-to-end:

  • .github/workflows/release-whl-kernel.yml: drop the rocm700 dispatch option, the rocm700 clause in the build-rocm-matrix guard, the "700" matrix entry, and the release-rocm700 job (which published the artifact and ran update_kernel_whl_index.py --rocm 700). The ROCm wheel matrix becomes ["720", "1000"].
  • 3rdparty/amd/wheel/sgl-kernel/build_rocm.sh: drop the 700 build image (lmsysorg/sglang:v0.5.8.post1-rocm700-mi35x) and its pinned rocm-rel-7.0.2 torch/triton/torchaudio/torchvision install branch. 720 and 1000 are unchanged, and an explicit 700 argument now fails with a clear message.
  • 3rdparty/amd/wheel/sglang/pyproject.toml: remove the rocm700 optional-dependency extra (which pinned sglang_kernel-0.4.0+rocm700) and the stale install comment. rocm720 is unchanged.
  • 3rdparty/amd/wheel/README.md: remove the ROCm 7.0.0 pip install "amd-sglang[all-hip,rocm700]" recipe and the ROCm 7.0.0 Triton upgrade note, and correct the extras description. Also fixes the section heading typo Revolving tritonResolving triton.

#38763's ROCm 10 wheel path is untouched.

What this does not touch

The remaining ROCm 7.0 references are deliberately out of scope:

  • The ROCm 7.0 workflows, nightly image build, and release-docker-amd.yml matrix leg — all handled by [AMD][CI] Consolidate AMD workflows and retire ROCm 7.0 CI #38632.
  • scripts/ci/amd/amd_ci_start_container*.sh still default to ROCM_VERSION=rocm700. This cannot change independently: all 58 container-launch sites in the current pr-test-amd.yml / nightly-test-amd.yml omit --rocm-version and rely on that default, so flipping it to rocm10 before those workflows are deleted would silently repoint the ROCm 7.0 runs at ROCm 10 images. It belongs with [AMD][CI] Consolidate AMD workflows and retire ROCm 7.0 CI #38632 or a follow-up.
  • docker/rocm.Dockerfile retains the unsuffixed gfx942 / gfx950 ROCm 7.0 stages for local builds.
  • SGLANG_USE_ROCM700A is misleadingly named but gates a workaround current ROCm 7.2 images still set.
  • Cookbook pages pin historical rocm700 image tags for reproducibility.

Accuracy Tests

Not applicable — release-artifact configuration only; no runtime or kernel code changed.

Speed Tests and Profiling

Not applicable. Removes one wheel build leg on the shared amd-docker-scale node plus its release job from every kernel version bump.

Verification

  • No rocm700, ROCm 7.0 dependency URL, "700" build input, or --rocm 700 remains in release-whl-kernel.yml or 3rdparty/amd/wheel/.
  • release-whl-kernel.yml parses as YAML; the surviving legs are rocm720 / rocm1000 in the dispatch menu, the ["720", "1000"] matrix, and the release-rocm720 / release-rocm1000 jobs.
  • build_rocm.sh passes bash -n; 720 and 1000 still resolve their images, and 700 exits 1 with Only '720' and '1000' are supported.
  • pyproject.toml parses with tomllib; the only remaining rocm* extra is rocm720.
  • python3 scripts/lint/check_workflow_job_names.py passes.

Checklist


CI States

Latest PR Test (Base): ✅ Run #34489631801
Latest PR Test (Extra): ❌ Run #34489631335
Latest PR Test (AMD ROCm 10): ➖ No AMD PR run found for this commit.

@github-actions github-actions Bot added documentation Improvements or additions to documentation amd labels Sep 10, 2026
@michaelzhang-ai michaelzhang-ai changed the title [AMD][CI] Drop the ROCm 7.0 nightly image build, nightly run, and PR shadow run [AMD][CI] Drop the ROCm 7.0 image builds, nightly run, and PR shadow run Sep 10, 2026
@cursor
cursor Bot force-pushed the cursor/drop-rocm700-image-build-and-ci-d06d branch from 569e64d to 933d989 Compare September 10, 2026 07:04
@michaelzhang-ai
michaelzhang-ai marked this pull request as ready for review September 10, 2026 07:07
@github-actions github-actions Bot added the dependencies Pull requests that update a dependency file label Sep 10, 2026
@cursor cursor Bot changed the title [AMD][CI] Drop the ROCm 7.0 image builds, nightly run, and PR shadow run [AMD][CI] Retire ROCm 7.0 release artifacts and CI Sep 10, 2026
@cursor
cursor Bot force-pushed the cursor/drop-rocm700-image-build-and-ci-d06d branch from 2692672 to e4aae9b Compare September 10, 2026 07:51
The rocm700 leg is the only ROCm kernel-wheel flavor with no CI behind it:
it builds in a frozen v0.5.8.post1-rocm700-mi35x image against pinned
rocm-rel-7.0.2 wheels. Drop the workflow target, the build-script branch,
the amd-sglang extra it fed, and the ROCm 7.0.0 install docs.

Co-authored-by: quitenode <quitenode@users.noreply.github.com>
@cursor
cursor Bot force-pushed the cursor/drop-rocm700-image-build-and-ci-d06d branch from e4aae9b to 81733a6 Compare September 10, 2026 14:31
@michaelzhang-ai michaelzhang-ai changed the title [AMD][CI] Retire ROCm 7.0 release artifacts and CI [AMD][CI] Retire the ROCm 7.0 kernel wheel Sep 10, 2026
@bingxche
bingxche merged commit d006f40 into main Sep 11, 2026
96 of 100 checks passed
@bingxche
bingxche deleted the cursor/drop-rocm700-image-build-and-ci-d06d branch September 11, 2026 00:25
mqhc2020 pushed a commit to mqhc2020/sglang that referenced this pull request Sep 15, 2026
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: quitenode <quitenode@users.noreply.github.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

amd deepseek dependencies Pull requests that update a dependency file documentation Improvements or additions to documentation run-ci

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants