[mercant1982]

What this changes

Restore mixed_gpus.AUTO together with the other patched globals in test_auto_where_found.

On v1.1.62, the released default is AUTO=True, but the test's finally block restores the saved value and then unconditionally sets it to False. A successful test therefore leaks a different default into later tests. Saving AUTO before entering the try also makes cleanup independent of how far the test gets.

This changes one test file only; no engine or frontend runtime source changes.

Answers

• The default answers are unchanged (same tokens as before), or
• It changes answers, and it is opt-in (off by default). Setting:

Test-only change: no runtime setting changes. No inference or token comparison was run.

Tested on

Linux x86_64, Lenovo P620, Threadripper PRO 5975WX, 512 GiB RAM, Python 3.12.15. The host has two CMP170HX 64 GiB cards, but neither GPU was used.

Base: v1.1.62, 30145b8b97b2414e2ea91d9f53acbec58a82caf3.

Before/after:

• Run the existing test_auto_where_found, then assert that mixed_gpus.AUTO equals its entry value: before fails (True -> False); after passes (True -> True). The test itself passes in both revisions.
• python -m unittest serve.test_mixed_gpus serve.test_glm_link -v: 12 passed before and after, no skips.
• Ten synthetic CUDA-only config cases retain their original configuration before and after, with engine/subprocess calls prohibited.
• python tools/check_hip_compat.py: passes on the candidate.

The CUDA-only checks cover frontend configuration gating; they do not qualify a CUDA build, inference, performance or AMD hardware.

Speed (if this is a speed change)

Not a speed change. No GPU benchmark or model load.

Checks

• Builds on the platform I tested
• The tests I touched pass
• Docs updated where a setting or behaviour changed

No CUDA/HIP build was run; the change is Python test cleanup. No user setting or runtime behaviour changes, so no documentation update is needed.

Before/after evidence, probes and reproduction instructions.