PyTorch to ONNX Exporter

Experimental torch ONNX exporter. Compatible with torch>=2.1.

[!WARNING] This is an experimental project and is not designed for production use. Use torch.onnx.export for these purposes.

Installation

pip install --upgrade torch-onnx

Usage

import torch
import torch_onnx
from onnxscript import ir
import onnx

# Get an exported program with torch.export
exported = torch.export.export(...)
model = torch_onnx.exported_program_to_ir(exported)
proto = ir.to_proto(model)
onnx.save(proto, "model.onnx")

# Or patch the torch.onnx export API
# Set error_report=True to get a detailed error report if the export fails
torch_onnx.patch_torch(report=True, verify=True, profile=True)
torch.onnx.export(...)

# Use the analysis API to print an analysis report for unsupported ops
torch_onnx.analyze(exported)

Design

{dynamo/jit} -> {ExportedProgram} -> {torchlib} -> {ONNX IR} -> {ONNX}

• Use ExportedProgram
  • Rely on robustness of the torch.export implementation
  • Reduce complexity in the exporter
  • This does not solve dynamo limitations, but it avoids introducing additional breakage by running fx passes
• Flat graph; Scope info as metadata, not functions
  • Because existing tools are not good at handling them
• Eager optimization where appropriate
  • Because exsiting tools are not good at optimizing
• Drop in replacement for torch.onnx.export
  • Minimum migration effort
• Build graph eagerly in the exporter
  • Give the exporter full control over the graph being built

Why is this doable?

• We need to verify torch.export coverage on Huggingface Optimum https://github.com/huggingface/optimum/tree/main/optimum/exporters/onnx; and they are not patching torch.onnx itself.
• Patch torch.onnx.export such that packages do not need to change a single line to use dynamo
• We have all operators implemented and portable