pytorch - 💡(How to fix) Fix [vllm] [2.12 regression][Inductor] prims.convert_element_type receives MetaProxy instead of Tensor in FP8 rms_norm fusion

Summary

Under torch 2.12.0 with backend='inductor', a graph produced by vLLM's FP8 rms_norm fusion pass fails during compilation with:

RuntimeError: prims::convert_element_type() Expected a value of type 'Tensor' for argument 'a' but instead found type 'MetaProxy'. Value: MetaProxy(arg4_1). Cast error details: Unable to cast MetaProxy(arg4_1) to Tensor

Same graph compiles on torch 2.11.0. This is one of the regressions blocking the torch 2.12 upgrade for vLLM (vllm-project/vllm#40077).

Environment

• torch: 2.12.0+cu130 (test channel)
• triton: 3.7.0
• CUDA: 13.0
• Python: 3.12
• GPU: reproduces on both H100 and B200
• Affected attention backends: FLASHINFER, TRITON_ATTN

Repro

• Model: nvidia/Llama-4-Scout-17B-16E-Instruct-FP8
• Fusion pass: quant_fp8 + rms_norm
• Compile mode: inductor_partition
• Reproduces with and without TP.

Failing vLLM tests (all same root error):

Error + abbreviated traceback

RuntimeError: Worker failed with error 'backend='inductor' raised:
RuntimeError: prims::convert_element_type() Expected a value of type 'Tensor' for argument 'a' but instead found type 'MetaProxy'.
Value: MetaProxy(arg4_1)
Cast error details: Unable to cast MetaProxy(arg4_1) to Tensor

The error originates during inductor lowering when calling prims.convert_element_type on a fx node whose value is wrapped in a MetaProxy rather than unwrapped to the underlying Tensor.

Question / diagnosis

Did torch 2.12 change how MetaProxy is unwrapped (or when it's preserved) during inductor graph lowering? A MetaProxy leaking into a prims.* call that expects a Tensor suggests either:

• a missing unwrap in the fusion pass path, or
• stricter type-checking in prims.convert_element_type on 2.12 that rejects MetaProxy where 2.11 quietly accepted it.

Happy to provide a minimal inductor-only repro extracted from vLLM's fusion graph if helpful.

Links

• vLLM PR: https://github.com/vllm-project/vllm/pull/40077
• Failing Buildkite build: https://buildkite.com/vllm/ci/builds/62138
• Example failed job (Fusion E2E TP2 B200): https://buildkite.com/vllm/ci/builds/62138#019dab36-870f-4089-9996-6c430835bdf8
• Umbrella: #180899

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