depyf

Warnings

• breaking depyf requires a recent version of PyTorch, specifically `>=2.2.0`, and often recommends using PyTorch nightly builds due to its close integration with the rapidly evolving `torch.compile` stack. Older PyTorch versions may lead to compatibility issues or incorrect decompilation.
  Fix: Upgrade PyTorch to version 2.2.0 or newer, preferably using a nightly build if encountering issues.
• gotcha depyf is a specialized decompiler focused on bytecode generated by `torch.compile`. It is not a general-purpose Python bytecode decompiler and may not fully support all Python syntax (e.g., `async/await`, complex `while` loops, or intricate `if/else` patterns outside of PyTorch's optimized graphs).
  Fix: Understand that depyf's scope is primarily for PyTorch-generated bytecode. For general Python decompilation, other tools might be necessary, though they often struggle with PyTorch bytecode.
• gotcha The output source code generated by depyf is semantically equivalent to the original bytecode but may not be syntactically identical. It often includes verbose details (e.g., explicit `return None`) that are typically implicit in human-written Python code.
  Fix: Be aware that the decompiled output serves as a detailed representation of the bytecode's operations, not a precise reconstruction of the original Python source. Focus on the logical flow rather than exact syntax.
• deprecated The `TORCH_COMPILE_DEBUG` environment variable, while providing debug information for `torch.compile`, produces logs that are generally much harder to parse and understand compared to `depyf`'s human-readable decompiled source. Relying solely on `TORCH_COMPILE_DEBUG` is inefficient for deep analysis.
  Fix: Prioritize using `depyf`'s context managers (`prepare_debug`, `debug`) for a clearer and more organized understanding of `torch.compile`'s internal workings.
• gotcha The `DEPYF_REMOVE_TEMP` environment variable, introduced in v0.20.0, can affect whether temporary files generated during decompilation are removed. If you need to inspect intermediate artifacts, ensure this is not set or set appropriately.
  Fix: Set `DEPYF_REMOVE_TEMP=0` in your environment if you wish to retain temporary files generated by depyf for further inspection.

Install

• pip install depyf Install stable release
• pip install git+https://github.com/thuml/depyf.git Install nightly version (recommended)

Imports

• decompile

  from depyf import decompile

• prepare_debug

  import depyf
  with depyf.prepare_debug(...)

• debug

  import depyf
  with depyf.debug()

Quickstart

This example demonstrates how to use `depyf` to inspect the code generated by `torch.compile`. Wrapping the `main` function (which calls the `torch.compile`d `toy_example`) with `depyf.prepare_debug` will decompile and dump the generated source code into the specified directory (`./debug_dir`). You can then examine these files to understand PyTorch's compiler optimizations. The `depyf.debug()` context manager can be used to pause execution for interactive debugging with breakpoints.

import torch
import depyf

@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))

# Wrap the code that triggers compilation within depyf.prepare_debug
# This will dump decompiled source code to './debug_dir'
with depyf.prepare_debug("./debug_dir"):
    main()

# Optional: Use depyf.debug() to pause execution and set breakpoints
# The program will pause here, allowing you to browse files in ./debug_dir
# and set breakpoints before continuing execution.
# with depyf.debug():
#    output = toy_example(torch.randn(10), torch.randn(10))

print("Decompiled code and debug info available in ./debug_dir")