Prevent saving empty profiler files

[DarkEden-coding]

Hello, current functionality does some weird stuff with threads sometimes, and ends up saving a ton of empty profiling files. Given that these are useless, perhaps a simple warning print statement and no saved files would be sufficient to prevent repo cluttering. The correct report is generated at the end of the code run, so nothing is lost with these empty profiling files.

Example output (in all generated text files):

Timer unit: 1e-07 s

[Theelx]

Huh, that's odd. I can work on a fix, but can you let us know what OS this is running on and give a minimal reproducible example script along with the command you use to run it?

[DarkEden-coding]

Thanks for the quick response! I am using Windows 11, Version 5.0.0 of line_profiler. Here is the minimal test code I can get; it seems to be directly related to the way Torch handles multiprocessing/threading.

"""
Minimal test case demonstrating line_profiler empty file bug.

PROBLEM:
When using line_profiler with multiprocessing (DataLoader workers) and LINE_PROFILE=1,
multiple processes try to write to the same output files simultaneously, possibly causing race conditions.
This results in some profile files being created with only headers and no actual profile data.

HOW TO REPRODUCE:
1. Set environment variable: $env:LINE_PROFILE=1
2. Run: python line_profiler_test.py
3. Check directory for profile_output* files

EXPECTED RESULTS:
- Multiple profile files are created with timestamps
- Some files contain only "Timer unit: 1e-07 s" (empty files)
- One file contains full profiling data
- Typically 5+ (one file per iteration) empty files per run (possibly) due to race conditions

TECHNICAL DETAILS:
- Single @profile decorated function with multiprocessing DataLoader
- 5 iterations within one function call
- num_workers=8 creates multiple processes competing for file writes
- Race conditions (possibly) cause some processes to write incomplete/empty files
"""

import torch
from line_profiler import profile
from torch.utils.data import DataLoader, Dataset


class D(Dataset):
    def __len__(self):
        return 10

    def __getitem__(self, i):
        return torch.randn(2, 2), torch.randn(2, 2)


@profile
def test():
    dataset = D()
    loader = DataLoader(dataset, batch_size=1, num_workers=8)

    # Create multiple iterations within one function (one empty file per iteration)
    for i in range(5):
        print(f"Iter {i}")

        # Access dataloader (error does not trigger if removed)
        for _, _ in loader:
            pass


if __name__ == "__main__":
    test()

This code was run with $env:LINE_PROFILE=1; python line_profiler_test.py

More complicated code using the torch dataloader also results in these outputs; it's not just this very simple script.

Output of the script (and the script) is attached:

line_profiler_test.py
profile_output.txt
profile_output_2025-09-16T210602.txt
profile_output_2025-09-16T210606.txt
profile_output_2025-09-16T210610.txt
profile_output_2025-09-16T210615.txt
profile_output_2025-09-16T210619.txt
profile_output_2025-09-16T210621.txt