🐛 Describe the bug

I am trying to collect log files generated during torch.compile execution for debugging purposes, but the files do not always appear.

I created the following simple test script:

import torch
import torch.nn as nn
import torch.nn.functional as F

device = torch.device("cuda")

class SimpleModel(nn.Module):
    def __init__(self):
        super(SimpleModel, self).__init__()
        self.proj1 = nn.Linear(10, 10)
        self.proj2 = nn.Linear(10, 10)

    def forward(self, x):
        x = F.relu(self.proj1(x))
        x = self.proj2(x)
        return x

torch.compiler.reset()
model = SimpleModel()
model.to(device)
compiled_model = torch.compile(model)

x = torch.randn(1, 10).to(device)
y = compiled_model(x)

Observed behavior:

After a reboot, the debug log files are correctly generated under torch_compile_debug/<log_dir>/torchinductor/model__0_forward_1.0/. However, no new log files are generated when running the script again without rebooting. This might be related to a cache issue.

Expected behavior:

TORCH_COMPILE_DEBUG=1 would consistently generate debug logs, without requiring a system reboot.

Error logs

No debug files generated.

Versions

env.txt

cc @chauhang @penguinwu @voznesenskym @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @amjames