pytorch - 💡(How to fix) Fix [vllm] [2.12 regression][Inductor] warm_artifacts_saved == 0 and KeyError(None) in standalone_compile startup tests

Summary

Under torch 2.12.0, vLLM's H100 compile-startup tests fail because the inductor standalone-compile warm cache no longer saves artifacts:

• compile/h100/test_startup.py::test_model_startup[deepseek_v3.2] → AssertionError: warm_artifacts_saved: got 0
• compile/h100/test_startup.py::test_model_startup[kimi_k2.5] → AssertionError: warm_artifacts_saved: got 0
• compile/h100/test_startup.py::test_moe_startup[0] → KeyError: None

Same tests pass on torch 2.11.0. Blocking the torch 2.12 upgrade for vLLM (vllm-project/vllm#40077).

Environment

• torch: 2.12.0+cu130
• triton: 3.7.0
• CUDA: 13.0
• Python: 3.12
• GPU: H100

Question / diagnosis

In 2.11, vLLM's compiler_interface.standalone_compile(...) produces warm artifacts that get written to disk; in 2.12 the write count is 0, and at least one test_moe_startup test gets KeyError: None from somewhere in the cache lookup (probably a missing key returning None).

• Is there a known change to torch._inductor.standalone_compile in 2.12 around how warm/aot-autograd artifacts get saved (naming, cache directory layout, or bypass conditions)?
• Was the None key an intentional sentinel for "no cache entry", and something is now blindly indexing it?

vLLM call site: vllm/compilation/compiler_interface.py:351 → torch._inductor.standalone_compile(...).

Links

• vLLM PR: https://github.com/vllm-project/vllm/pull/40077
• Failing build: https://buildkite.com/vllm/ci/builds/62138
• Failed job (PyTorch Compilation Unit Tests H100): https://buildkite.com/vllm/ci/builds/62138#019dab36-87f7-449c-b761-dcc9c2cac612
• Umbrella: #180899

cc @chauhang @penguinwu @voznesenskym @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @aakhundov @coconutruben @jataylo