Have you ever felt overwhelmed by the complexities of torch.compile? Diving into its workings can feel like black magic, with bytecode and Python internal details that many users fail to understand, hindering them from understanding and adapting to torch.compile.

If you also face the problem, then you might be interested in depyf. As the logo suggests, depyf is a software tool to leverage advanced Python features (the Python snake symbol) to open up internal details (the internal gears symbol) of PyTorch's compiler torch.compile (the PyTorch logo), so that users can understand it, adapt to it, and tune their code (the debugger symbol) to get maximum performance benefit out of it.

⚠️ This project is developed under close collaborations with the PyTorch team. We frequently require new features from PyTorch to support better understanding of torch.compile. Therefore, please use this project along with PyTorch nightly. Visit the PyTorch website for how to install nightly version of PyTorch. We recommend updating your PyTorch nightly installation every week or so.

⚠️ During development, we seek suggestions from the community quite a lot. You may find some early usage examples from some discussion forums or social media platforms. Please follow the latest documentation for how to use this tool.

If you want to understand bytecode generated by torch.compile, then depyf might be the only choice for you. Below we tested several existing decompilers, they struggle to decompile simple Python bytecode across versions, and have poor support for PyTorch.

Stable release: pip install depyf

Nightly version (recommended): pip install git+https://github.com/thuml/depyf.git

The main usage is quite simple: just wrap your code within a context manager:

import torch
from torch import _dynamo as torchdynamo
from typing import List

@torch.compile
def toy_example(a, b):
    x = a / (torch.abs(a) + 1)
    if b.sum() < 0:
        b = b * -1
    return x * b

def main():
    for _ in range(100):
        toy_example(torch.randn(10), torch.randn(10))

if __name__ == "__main__":
-     main()
+     import depyf
+     with depyf.prepare_debug("./dump_src_dir"):
+         main()

Then you can see all the details of torch.compile inside the directory ./dump_src_dir. The details are organized into the following:

• full_code_for_xxx.py for each function using torch.compile
• __transformed_code_for_xxx.py for Python code associated with each graph.
• __transformed_code_for_xxx.py.xxx_bytecode for Python bytecode, dumped code object, can be loaded via dill.load(open("/path/to/file", "wb")). Note that the load function might import some modules like transformers. Make sure you have these modules installed.
• __compiled_fn_xxx.py for each computation graph and its optimization:
  • Captured Graph: a plain forward computation graph
  • Joint Graph: joint forward-backward graph from AOTAutograd
  • Forward Graph: forward graph from AOTAutograd
  • Backward Graph: backward graph from AOTAutograd
  • kernel xxx: compiled CPU/GPU kernel wrapper from Inductor.

We collect all the compilation artifacts when testing over 100 deep learning models. You can take a look to learn how the PyTorch compiler works.

If you want to use debugger to step through the above code, just add another context manager (and launch the script through debuggers):

import torch
from torch import _dynamo as torchdynamo
from typing import List

@torch.compile
def toy_example(a, b):
    x = a / (torch.abs(a) + 1)
    if b.sum() < 0:
        b = b * -1
    return x * b

def main():
    for _ in range(100):
        toy_example(torch.randn(10), torch.randn(10))

if __name__ == "__main__":
    import depyf
    with depyf.prepare_debug("./dump_src_dir"):
        main()
+     with depyf.debug():
+         main()

Calling depyf.debug() will pause the program for you to set breakpoints, and then you can use debuggers to hit breakpoints in these files under the ./dump_src_dir directory you specified above.

If you have any question about depyf, feel free to open issues to reach out! Any discussion/issue report/PR is welcome. Or contact [email protected] if you have any other questions.