Dynamo API

This section documents the PyTorch Dynamo integration for tracing compiled models.

Dynamo Class

Module Instance

The module provides a pre-configured dynamo instance:

from dftracer.python import dynamo

# Use the dynamo decorator
@dynamo.compile
def forward(self, x):
    return x * 2

Constants

Internal Classes

CallStackRecord

Backend Functions

create_backend

The create_backend function creates a custom PyTorch compile backend with DFTracer instrumentation. This allows you to use DFTracer directly with torch.compile().

Example:

from dftracer.python.dynamo import create_backend
import torch

# Create a backend with custom parameters
backend = create_backend(
    name="my_model",
    epoch=0,
    step=0,
    enable=True,
    autograd=True
)

# Use with torch.compile
model = MyModel()
compiled_model = torch.compile(model, backend=backend)

Utility Functions

create_detailed_op_name

Usage Examples

See PyTorch Dynamo Integration for detailed usage examples and best practices.