Use local rank for Responses Triton CUDA binding

[sylvesterkaczmarek]

Summary

Bind the Responses API Triton backend to its node-local CUDA device instead of using global distributed rank as a device index.

This backend selects its device independently and does not call gpt_oss.torch.utils.init_distributed(), so the local-rank correction in #267 does not cover it.

Fixes #273.

Fix

• keep global RANK for distributed identity and log prefixes;
• use LOCAL_RANK for torch.cuda.set_device() and the CUDA device object;
• fall back to global RANK when LOCAL_RANK is absent, preserving existing single-node/custom-launch behavior.

Regression coverage

Adds a CPU-only mocked regression with RANK=5 and LOCAL_RANK=1 verifying that:

• global rank remains 5;
• local rank is 1;
• model loading receives cuda:1;
• CUDA binding uses device index 1.

No model loading, checkpoint format, inference math, or logging identity is otherwise changed.

[chatgpt-codex-connector]

Note

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.