Skip to content

feat(genesis): support ROCm collector runtime - #2519

Merged
TATP-233 merged 8 commits into
dev/issue-2520-genesis-dr-lifecyclefrom
feat/issue-471-genesis-rocm
Oct 26, 2026
Merged

TATP-233 merged 8 commits into
dev/issue-2520-genesis-dr-lifecyclefrom
feat/issue-471-genesis-rocm

Conversation

@TATP-233

@TATP-233 TATP-233 commented Oct 26, 2026 •

Copy link
Copy Markdown
Collaborator

Summary

  • relax the UniLab HIP preflight only for UniSim's declared Genesis ROCm profile; MJWarp, Newton, IsaacGym, and IsaacSim remain CUDA-only
  • accept exactly one Genesis CUDA or AMDGPU diagnostic in the production collector benchmark and record torch_runtime_family plus genesis_backend
  • collect ROCm driver provenance through rocm-smi while preserving the NVIDIA nvidia-smi path
  • update the bilingual support matrix and Genesis/tensor-runtime documentation

Refs unilabsim/unisim#471

This PR also carries the six #457 collector-benchmark commits because that integration is not yet in main; the issue #471 commit is 677b709a.

Benchmark evidence

  • AMD MI300X / ROCm 7.2 / Genesis 1.3.3, 2048-env FlashSAC G1 production collector: 33,986 steps/s, tail 33,736 steps/s.
  • Local NVIDIA RTX 4090 / CUDA 13.0 / Genesis 1.3.3, same owner and parameters: 88,644 steps/s, tail 95,611 steps/s.
  • Both records contain the direct in-process device-resident profile. The ROCm artifact records torch_runtime_family=rocm, genesis_backend=amdgpu, and driver 6.19.14.31400000; the CUDA artifact records torch_runtime_family=cuda, genesis_backend=cuda, and driver 580.178.04.

Validation

  • make check — passed.
  • make test-all — 2811 passed, 61 skipped, 373 deselected; benchmark smoke 35/35 module-mode and 36/36 script-mode passed.
  • AMD focused factory/benchmark/support-matrix tests — 63 passed.
  • AMD Genesis fake runtime — 26 passed.
  • Collector benchmark tests — 45 passed.
  • AMD full suite also retains 17 NVIDIA-only-on-ROCm failures and the previously reported Motrix angular-velocity conformance failure; these are not part of this change.

Integration note

Current CI resolves the published unisim-core==2.1.1, which does not yet contain the Genesis ROCm profile. The ROCm preflight and generated-matrix checks therefore require the UniSim change to land in a released package. Final-head local validation used the matching UniSim source through PYTHONPATH; no local-path dependency or unpublished dependency pin is committed. UniSim PR: unilabsim/unisim#473.

@TATP-233
TATP-233 changed the base branch from main to dev/issue-2520-genesis-dr-lifecycle October 26, 2026 10:23
@TATP-233
TATP-233 merged commit 26b91cf into dev/issue-2520-genesis-dr-lifecycle Oct 26, 2026
@TATP-233
TATP-233 deleted the feat/issue-471-genesis-rocm branch October 26, 2026 10:35
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