Red Arrow Documentation
PyTorch Compiler

PyTorch Compiler

pytorch2 backend

TORCH_COMPILE_DEBUG=1 python my_script.py
  • TORCH_COMPILE_DEBUG=1 enables debug logging for torch.compile, providing insights into the compilation process, optimizations applied, and any potential issues encountered during graph capture and execution.
  • Generated triton kernel will be available in torchinductor/model_*/output_code.py file. It’ll also have fx graphs saved in torchinductor/model_*/fx_graphs.py file, which can be used to understand the transformations applied to the original PyTorch code during compilation.
  • we can also set logging options to control the verbosity and format of the debug output, such as:
import torch

torch._logging.set_logs(graph_code = True, graph_breaks = True)
torch._dynamo.reset()
  • torch._dynamo.reset() is a function that resets the state of the PyTorch Dynamo compiler, clearing any cached compiled graphs and allowing for a fresh start in subsequent compilations. This can be useful when you want to ensure that changes to your code are reflected in the compiled output without interference from previously cached results.