[tests] Some fixes for test suite failures on TPU (#49425)
* test: use the CUDA SDPA tolerances on TPU
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* test: request full-precision float32 matmuls in the test suite
enable_tf32(False) only sets the new fp32_precision API. torch_tpu follows
the legacy float32_matmul_precision, which it sets to "medium".
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* test: run subprocess tests in-process on single-process devices
TPU locks the device to the process that opens it, so the pytest process
holds it and the subprocess can never open it. Device spec files can add
their device with SINGLE_PROCESS_DEVICE = True.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* test: restore the default device when the global device test fails
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* test: add TPU expectations to the gemma3 integration tests
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* test: add TPU expectations to the gemma3n integration tests
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* test: add TPU expectations to the gpt2 integration tests
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UiAi1LaVWKhaLmJs7s7KQ7
* test: add TPU expectations to the qwen2 integration tests
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UiAi1LaVWKhaLmJs7s7KQ7
* test: add TPU expectations to the llama integration tests
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UiAi1LaVWKhaLmJs7s7KQ7
* test: add TPU expectations to the internvl integration tests
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UiAi1LaVWKhaLmJs7s7KQ7
* test: add TPU expectations to the mistral3 integration tests
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UiAi1LaVWKhaLmJs7s7KQ7
* test: add TPU expectations to the whisper integration tests
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UiAi1LaVWKhaLmJs7s7KQ7
* review(test): relax whisper expectations on TPU
We just used two digites, and the list matched ROCm, so we only keep
this one.
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
