How to check PyTorch2 model is compiled

torch.compile
1

Hi, how do I check my model has been successfully compiled in Pytorch 2, after calling torch.compile(). Basically, is there any properties in the model that I can check that it’s a compiled model?

Thanks

2

There’s a test utility that’s very helpful for this called CompileCounter pytorch/test_compile.py at main · pytorch/pytorch · GitHub

But otherwise the compilation happens after the first inference and another easy way to sanity check is to make sure kernels were generated if you add TORCH_COMPILE_DEBUG to True as an environemnt variable

3

Thanks for your reply! For the second solution, how can I check the kernels are generated?

4

The way I usually do it if I set TORCH_COMPILE_DEBUG then I’ll see the Triton or C++ kernels generated in a tmp folder that’ll be printed to your console if the debug flag is set

5

Hi, got it, many thanks!