📋 New Package Request for AIPCC

This form is for requesting new Python packages to be built and added to the AIPCC Package Index.

1. Requester Information

• Team / Project: ODH Notebooks / OpenDataHub Workbench Images

2. Package Details

• Package Name: onnxscript
• Package Source: PyPI: https://pypi.org/project/onnxscript/
• (Optional) Required package version: Latest stable version is acceptable.

3. Build & Platform Requirements

• Hardware Variants:
  Select all additional hardware platforms where this package is needed. A CPU-only variant is built by default for all requests.
• [x] CUDA (NVIDIA GPU)
• [x] ROCm (AMD GPU)
• [ ] TPU (Google TPU)
• [ ] Gaudi (Intel Gaudi)
• [ ] Spyre (IBM)
• [ ] Other

4. Justification & Testing

• Business Justification:

Summary: As of PyTorch 2.9+, onnxscript is no longer installed by default and must be explicitly added as a dependency for ONNX export functionality to work.

Background Investigation:

Our CI tests for PyTorch notebook images started failing with:

ERROR: test_convert_to_onnx (TestPytorchNotebook.test_convert_to_onnx)
ModuleNotFoundError: No module named 'onnxscript'

The failure occurs when calling torch.onnx.export() in the test notebook at jupyter/rocm/pytorch/ubi9-python-3.12/test/test_notebook.ipynb.

Cause analysis: PyTorch does not officially require onnxscript as a dependency. The torch.onnx module defers onnxscript imports to runtime (lazy import) rather than at initialization. This was an intentional design decision to handle the absence of the onnxscript module gracefully. See: https://github.com/pytorch/pytorch/issues/103764

Verification of AIPCC index availability: We checked all three AIPCC indexes and confirmed onnxscript is not available:

• • https://console.redhat.com/api/pulp-content/public-rhai/rhoai/3.3/cpu-ubi9/onnxscript/ → 404
  • https://console.redhat.com/api/pulp-content/public-rhai/rhoai/3.3/cuda12.9-ubi9/onnxscript/ → 404
  • https://console.redhat.com/api/pulp-content/public-rhai/rhoai/3.3/rocm6.4-ubi9/onnxscript/ → 404

Affected images: All PyTorch-based notebook and runtime images:

• • jupyter/pytorch/ubi9-python-3.12
  • jupyter/pytorch+llmcompressor/ubi9-python-3.12
  • jupyter/rocm/pytorch/ubi9-python-3.12
  • runtimes/pytorch/ubi9-python-3.12
  • runtimes/pytorch+llmcompressor/ubi9-python-3.12
  • runtimes/rocm-pytorch/ubi9-python-3.12

User impact: Without onnxscript, users cannot export PyTorch models to ONNX format using torch.onnx.export(), which is a common workflow for model deployment and interoperability.

• Release target: RHOAI 3.4 GA
• (Optional) Testing Requirements:

Default import test is sufficient:

import onnxscript

Additionally, the following validates the integration with PyTorch:

import torch
import torch.onnx

# If onnxscript is properly installed, this import chain should work
from torch.onnx._internal.exporter import _compat