Skip to content

fix(deps): restore ROCm runtime dependencies - #1650

Merged
TATP-233 merged 2 commits into
mainfrom
fix/rocm-runtime-dependencies
Sep 26, 2026
Merged

TATP-233 merged 2 commits into
mainfrom
fix/rocm-runtime-dependencies

Conversation

@TATP-233

Copy link
Copy Markdown
Collaborator

Summary

  • Restored numba>=0.67 and prettytable>=3.10 to pyproject.rocm.toml.
  • Regenerated uv.rocm.lock, adding numba, prettytable, and llvmlite.
  • Added a regression test that checks non-substituted runtime dependencies remain mirrored in the ROCm profile and lockfile.

Linked Work

Validation

  • make test-all passed on the final local head before this PR was created or updated
  • Additional task-specific validation listed below

Commands actually run:

uv run ruff format tests/scripts/test_rocm_dependency_profile.py
uv run ruff check tests/scripts/test_rocm_dependency_profile.py
uv run pytest -q tests/scripts/test_rocm_dependency_profile.py

# ROCm profile lock consistency (active manifests temporarily swapped, then restored):
cp pyproject.toml /tmp/unilab-default-pyproject.toml
cp uv.lock /tmp/unilab-default-uv.lock
cp pyproject.rocm.toml pyproject.toml
cp uv.rocm.lock uv.lock
uv lock --check --no-progress
cp /tmp/unilab-default-pyproject.toml pyproject.toml
cp /tmp/unilab-default-uv.lock uv.lock

make check
make test-all

make test-all result: 1549 passed, 27 skipped, 604 deselected; benchmark smoke passed 35/35 module-mode and 36/36 script-mode imports. The first local attempt encountered errors from an ignored, pre-existing scripts/benchmark/outputs/ directory; that unrelated local data was temporarily moved aside and restored before the successful final-head run.

Remote Linux ROCm validation in an isolated worktree:

make PATH=/home/unilab/.local/bin:$PATH sync-rocm
uv run --no-sync python -c 'import numba, prettytable; from unilab.base.registry import ensure_registries; ensure_registries(); print("numba", numba.__version__); print("prettytable", prettytable.__version__); print("registries ok")'
uv lock --check --no-progress
HIP_VISIBLE_DEVICES=7 uv run --no-sync train --algo sac --task g1_walk_flat --sim mujoco algo.max_iterations=2 training.no_play=true training.export_onnx=false training.log_dir=/tmp/unilab-rocm-pr-sac-smoke

The ROCm smoke completed 2/2 SAC iterations on GPU 7. GPU 0-6 jobs were not modified, and no validation process remained afterward.

Remote CI route:

  • Base main: current-head CI recorded below after the run starts.

Impact

  • Backend impact: none / dependency-profile only
  • Platform impact: Linux ROCm
  • Training effect expected: no algorithmic change; restores required runtime imports after make sync-rocm

Artifacts

  • W&B: none
  • benchmark result: none
  • video / screenshot: none
  • ONNX / checkpoint: none (remote smoke checkpoint was temporary)

Checklist

  • Added or updated tests where needed
  • Updated docs if behavior or workflow changed
  • Linked the driving issue
  • Noted any follow-up work explicitly

Follow-up noted in #1649: cold ROCm first-update compile can slightly exceed the current 30-second inference coordination timeout. That behavior is outside this dependency-profile PR.

@TATP-233
TATP-233 force-pushed the fix/rocm-runtime-dependencies branch from f2f2574 to ca11abe Compare September 26, 2026 18:19
@TATP-233
TATP-233 merged commit 1d21f27 into main Sep 26, 2026
7 of 8 checks passed
@TATP-233
TATP-233 deleted the fix/rocm-runtime-dependencies branch September 26, 2026 18:25
@TATP-233 TATP-233 mentioned this pull request Sep 26, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant