pytorch - 💡(How to fix) Fix [vllm] [2.12 regression][AOTAutograd] "Cannot access data pointer of Tensor (FakeTensor/FunctionalTensor)" during cache save

Summary

Under torch 2.12.0, vLLM's Quantization suite fails to compile with torch.compile because AOTAutogradCache.save trips:

RuntimeError: Cannot access data pointer of Tensor (e.g. FakeTensor, FunctionalTensor). If you're using torch.compile/export/fx, it is likely that we are erroneously tracing into a custom kernel. To fix this, please wrap the custom kernel into an opaque custom op.

Preceded by:

W torch/_functorch/_aot_autograd/autograd_cache.py:1224] AOTAutograd cache unable to serialize compiled graph: Cannot access data pointer of Tensor (...)

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

• Did 2.12 start routing more graphs through AOTAutogradCache.save, or tighten what's pickleable?
• The warning says the cache write fails, but the resulting RuntimeError propagates out and breaks EngineCore init — is that intentional? In 2.11 a cache-save failure was a silent warning.
• Is there a supported opt-out (env var) to treat cache-save failures as bypass rather than hard fail?

Affected tests

Quantization suite (1 failed out of 261 passed, 27 skipped — 2017s). Top-of-trace is engine core init on the suite's compile warm-up.

(A secondary failure in the same log — AssertionError: Hidden size mismatch 2048 != 1024 — appears later and may be unrelated; will split if so.)

Links

• vLLM PR: https://github.com/vllm-project/vllm/pull/40077
• Failing build: https://buildkite.com/vllm/ci/builds/62138
• Failed job (Quantization): https://buildkite.com/vllm/ci/builds/62138#019dab36-87fd-4fa0-88ee-3878996ca49c
• Umbrella: #180899

cc @chauhang @penguinwu @bdhirsh @bobrenjc93 @aorenste