Skip to content

[AMD][CI] Drop the dead miles ROCm 7.0 nightly image build - #38694

Merged
HaiShaw merged 1 commit into
mainfrom
cursor/retire-dead-miles-rocm700-build-f847
Sep 10, 2026
Merged

HaiShaw merged 1 commit into
mainfrom
cursor/retire-dead-miles-rocm700-build-f847

Conversation

@michaelzhang-ai

@michaelzhang-ai michaelzhang-ai commented Sep 9, 2026

Copy link
Copy Markdown
Collaborator

Motivation

release-docker-amd-miles-rocm700-nightly.yml builds the rocm700-mi35x and rocm700-mi30x variants through miles' docker/build.py. radixark/miles#2854 ("retire the rocm700 variants, add a rocm10-mi35x variant") merged 2026-09-09 05:37 UTC and deleted both, leaving rocm720-mi35x and rocm10-mi35x as the only ROCm variants.

The workflow still asks for the retired ones, so the very next scheduled run — run 34348795681, 2026-09-09 12:02 UTC — failed immediately:

python3 docker/build.py --variant rocm700-mi35x --image-tag custom --custom-tag miles
Invalid value for '--variant': 'rocm700-mi35x' is not one of 'cu13', ...

It cannot succeed again. Left in place it burns two amd-docker-scale jobs and posts a red daily build on the MILES dashboard every day at 12:00 UTC, which is noise on top of the real nightly failures.

Modifications

Delete .github/workflows/release-docker-amd-miles-rocm700-nightly.yml.

The ROCm 10 replacements (release-docker-amd-miles-rocm10-nightly.yml and nightly-test-amd-miles-rocm10.yml) are already proposed in #37495, so this only removes the dead half; the two can land in either order.

Verification

  • rg 'miles-rocm700' over the repo returns nothing, so no workflow, script or doc consumes the rocm/sgl-dev:miles-rocm700-* tags this workflow published. The remaining rocm700 hits in scripts/ci/amd/amd_ci_start_container*.sh and docker/rocm.Dockerfile refer to plain SGLang ROCm 7.0 images (v0.5.8.post1-rocm700-*), which are unrelated and untouched.
  • rg 'release-docker-amd-miles-rocm700' returns nothing, so nothing references the workflow by name.
  • python3 scripts/lint/check_workflow_job_names.py passes.
  • Miles' current variant list is ['cu13', 'cu13-x86', 'cu13-aarch64', 'cu12-x86', 'rocm720-mi35x', 'rocm10-mi35x']; rocm700 no longer appears anywhere in that repo outside docs.

release-docker-amd-miles-rocm720-nightly.yml is unaffected — rocm720-mi35x is still a supported variant and that workflow still builds.

Accuracy Tests

Not applicable — removes a CI workflow, no model or kernel code is touched.

Speed Tests and Profiling

Not applicable.

Checklist

Open in Web Open in Cursor 

CI States

Latest PR Test (Base): ✅ Run #34360497621
Latest PR Test (Extra): ❌ Run #34360497168
Latest PR Test (AMD ROCm 10): ➖ No AMD PR run found for this commit.

radixark/miles#2854 removed the rocm700-mi35x and rocm700-mi30x variants
from docker/build.py, leaving rocm720-mi35x and rocm10-mi35x. This
workflow still asks for the retired variants, so since that merge it can
only fail:

  Invalid value for '--variant': 'rocm700-mi35x' is not one of 'cu13', ...

Nothing consumes the miles-rocm700-* tags. The ROCm 10 replacement
workflows come from #37495.

Co-authored-by: Michael <michaelzhang-ai@users.noreply.github.com>
@HaiShaw
HaiShaw merged commit 13fb796 into main Sep 10, 2026
95 of 99 checks passed
@HaiShaw
HaiShaw deleted the cursor/retire-dead-miles-rocm700-build-f847 branch September 10, 2026 07:00
mqhc2020 pushed a commit to mqhc2020/sglang that referenced this pull request Sep 15, 2026
…ct#38694)

Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: Michael <michaelzhang-ai@users.noreply.github.com>
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.

3 participants