[sglang-miles] Cherry-pick #24766: Fix NUMA NVML handle resolution under shuffled CUDA_VISIBLE_DEVICES

[ByronHsu]

Summary

Cherry-pick of #24766 (merge commit 7edb4c3) onto sglang-miles.

• Resolves CUDA logical id via torch.cuda._get_nvml_device_index before calling nvmlDeviceGetHandleByIndex, so NUMA binding follows reordered CUDA_VISIBLE_DEVICES
• Falls back to original index with a warning on torch builds lacking the helper
• Enriches numactl debug string with logical/physical GPU ids and CUDA_VISIBLE_DEVICES

Conflict resolution: Minor conflict in configure_subprocess control flow in numa_utils.py (sglang-miles had a slightly different structure); resolved by taking the upstream version.

Test plan

• Existing NUMA tests should pass
• Manual verification on multi-GPU nodes with shuffled CVD

Made with Cursor

[gemini-code-assist]

Code Review

This pull request improves NUMA node affinity detection by correctly mapping logical CUDA device indices to physical NVML device indices, which is essential when CUDA_VISIBLE_DEVICES reorders or subsets GPUs. It introduces a helper function _get_nvml_device_index that leverages PyTorch's internal mapping and enhances debug logging in configure_subprocess to include logical and physical GPU IDs. I have no feedback to provide as there were no review comments to evaluate.