[inductor] cpp codegen using already deleted header

[xmfan]

🐛 Describe the bug

Inductor seems to be using a header filename generated from a previous test run, but temporary directories are cleared between tests.

Issue boils down to the fresh_inductor_cache wrapper not cleaning up properly once exited

import torch
from torch._inductor.utils import fresh_inductor_cache

def fn():
    x = torch.randn(1, 10)
    y = torch.randn(10, 1)
    return torch.mm(x, y).sum()

def fn2():
    x = torch.randn(10, 100)
    y = torch.randn(100, 10)
    return torch.mm(x, y).sum()

with fresh_inductor_cache():
    torch.compile(fn)()

torch.compile(fn2)()

makes compiled autograd flaky tests: TORCHINDUCTOR_COMPILE_THREADS=1 pytest test/inductor/test_compiled_autograd.py -k 'test_torch_compile or test_access_saved_tensor_twice_without_recomputation_works'

Versions

main

cc @ezyang @msaroufim @bdhirsh @anijain2305 @chauhang @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @peterbell10 @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @ColinPeppler @amjames @desertfire

[jgong5]

The header file about cpp_prefix.h and is named after its hash. It should be generated before compiling the generated cpp code. Things are being tested with other UTs in test_torchinductor.py and test_cpu_repro.py. Is it test_compile_autograd.py specific?

[xmfan]

I only have a repro with test_compiled_autograd.py, do you have a code pointer to where we generate the header name?

[xmfan]

@jgong5 found a smaller repro, and have a fix

import torch
from torch._inductor.utils import fresh_inductor_cache

def fn():
    x = torch.randn(1, 10)
    y = torch.randn(10, 1)
    return torch.mm(x, y).sum()

def fn2():
    x = torch.randn(10, 100)
    y = torch.randn(100, 10)
    return torch.mm(x, y).sum()

with fresh_inductor_cache():
    torch.compile(fn)()

torch.compile(fn2)()