Fix CI

[pelumi1163]

No description provided.

[github-actions]

Review Summary

Reviewed the diff against HEAD^1 (base dev). Scope is a 9-line addition to .github/workflows/rocm-ci.yml inside the shared &run_container anchor (applied to both sGPU and MGPU jobs).

What it does: parses --device=/dev/dri/renderD* entries from DEVICE_FLAG, counts them, and — when non-zero — injects -e ROCR_VISIBLE_DEVICES=0,…,N-1 -e HIP_VISIBLE_DEVICES=0,…,N-1 into docker run. This restricts ROCm/HIP enumeration to the assigned GPUs (necessary because /dev/kfd is shared and otherwise exposes host-wide topology).

Verdict: LGTM. The --device=/dev/dri fallback path is handled correctly (grep matches nothing → empty array → guard skips → "${GPU_ENV[@]}" expands to nothing). Change is minimal, contained, and ROCm-only (no upstream compatibility concern).

Minor nit (non-blocking): RENDER_DEVICES is only used for its length — GPU_COUNT=$(grep -cE '/dev/dri/renderD[0-9]+' <<< "$DEVICE_FLAG") would be equivalent and drop the array. Not worth changing.

Copyright headers: OK (rocm-ci.yml already carries 2024-2026 AMD line; no NVIDIA line expected on ROCm-only workflow).