🐛 Describe the bug

torch.compile with the Inductor backend produces a different result from eager execution for a cast-to-float16 multiplication pattern.

In eager mode, the input is first cast to float16, then multiplied in float16, which overflows to inf. In the compiled version, the result is 25000000.0, suggesting the multiplication is effectively performed without preserving the intermediate fp16 overflow behavior.

import torch

def fn(x):
    y = x.to(torch.float16)
    return (y * y).to(torch.float32)

x = torch.tensor([5000.0])

eager = fn(x).item()

torch._dynamo.reset()
compiled = torch.compile(fn, backend="inductor")(x).item()

print("eager:", eager, "compile:", compiled)

Error logs

eager: inf compile: 25000000.0

Versions

PyTorch version: 2.11.0+cu130 Is debug build: False CUDA used to build PyTorch: 13.0 ROCM used to build PyTorch: N/A

OS: Ubuntu 24.04.2 LTS (x86_64) GCC version: (Ubuntu 13.3.0-6ubuntu2~24.04.1) 13.3.0 Clang version: Could not collect CMake version: Could not collect Libc version: glibc-2.39

Python version: 3.10.20 (main, Mar 11 2026, 17:46:40) [GCC 14.3.0] (64-bit runtime) Python platform: Linux-5.15.167.4-microsoft-standard-WSL2-x86_64-with-glibc2.39 Is CUDA available: False CUDA runtime version: Could not collect CUDA_MODULE_LOADING set to: N/A GPU models and configuration: GPU 0: NVIDIA GeForce RTX 3080 Laptop GPU Nvidia driver version: 545.92 cuDNN version: Could not collect Is XPU available: False HIP runtime version: N/A MIOpen runtime version: N/A Is XNNPACK available: True Caching allocator config: N/A

Versions of relevant libraries: [pip3] numpy==2.2.6 [pip3] onnx==1.21.0 [pip3] onnx2torch==1.5.15 [pip3] onnxruntime==1.23.2 [pip3] torch==2.11.0 [pip3] torchvision==0.26.0 [pip3] triton==3.6.0

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