DML provider crash / runtime error on attention Einsum (SCUNet-GAN.simpl.onnx)

[TheColorRed]

Title: DML provider crash / runtime error on attention Einsum (SCUNet-GAN.simpl.onnx)

Description

Running SCUNet-GAN.simpl.onnx (simplified from SCUNet-GAN.onnx) with the DirectML provider (ONNX Runtime DmlExecutionProvider) on Windows/AMD produces an access-violation crash (process exit code 3221225477) in my environment. Binary search isolation points to an Einsum in the MSA block; replacing it with Transpose+MatMul avoids the segfault but causes a runtime error in DML (MatMul: "The parameter is incorrect").

Files to reproduce

• abra/ai/super-resolution/models/SCUNet-GAN.simpl.onnx (simplified full model)
• deployment/tmp_prefix_65.onnx (minimal crashing prefix discovered by binary search)
• deployment/tmp_prefix_65_fixed.onnx (Einsum replaced with Transpose+MatMul; triggers MatMul runtime error on DML)
• deployment/onnx_dml_repro/run_repro.py (simple runner)
• deployment/onnx_dml_repro/logs.txt (captured outputs & environment info)

Environment

• OS: Windows 11 10.0.26200-SP0
• Python: 3.12.10
• onnx: 1.17.0
• onnxruntime: 1.23.2
• onnxruntime-directml: 1.21.1
• GPU: AMD (local machine)

Reproduction steps

1. Activate virtualenv (.venv\Scripts\activate) in the repo root.
2. (Quick test) Run: .venv\Scripts\python abra/ai/super-resolution/deployment/test_directml_safe.py abra/ai/super-resolution/models/SCUNet-GAN.simpl.onnx
   • Expected: process returns 3221225477 (access violation) in my earlier runs; current runs sometimes show NO_SUCHFILE referencing original model (see logs). Either way the run fails on DML.
3. (Isolated repro) Run: .venv\Scripts\python deployment/onnx_dml_repro/run_repro.py deployment/tmp_prefix_65.onnx
   • Observed: earlier binary search identified tmp_prefix_65.onnx as the minimal prefix that reproduces the problematic provider behavior.
4. (Einsum replacement experiment) Run: .venv\Scripts\python deployment/onnx_dml_repro/run_repro.py deployment/tmp_prefix_65_fixed.onnx
   • Observed: CPU runs OK; DML returns runtime error about MatMul parameters (see logs).

Observed logs

See deployment/onnx_dml_repro/logs.txt (contains MatMul runtime message and the earlier segfault exit code). Key excerpt:

• MatMul error (DML):
  "Non-zero status code returned while running MatMul node... The parameter is incorrect."

• Earlier segfault on simplified model:
  returncode: 3221225477

Notes & suggestions

• The problem appears tied to Einsum/attention matmul patterns used in the model's MSA block. Replacing Einsum removed the segfault, but DML failed to run MatMul on the replacement (shape/parameter issue).
• Minimal prefix (tmp_prefix_65.onnx) is fairly small (56 nodes) and should be helpful as a repro; SCUNet-GAN.simpl.onnx demonstrates the failure at production scale.

Would you like me to open this as an issue on the ONNX Runtime (runtime/directml provider) GitHub repo and attach these files? I can create the issue and attach the deployment/tmp_prefix_65.onnx and abra/ai/super-resolution/models/SCUNet-GAN.simpl.onnx files for triage.

[TheColorRed]

I've attached reproduction details in a public gist: https://gist.github.com/TheColorRed/46c7065ba31deb6b7a766a131aa32c74

The minimal crashing ONNX (base64-encoded) is here: https://gist.github.com/TheColorRed/255b3e5d55e920331e384053aee38c88 (decode with the instructions in the file).

If you'd like the full repro zip, I can provide it (it's available locally as deployment/onnx_dml_repro/repro_package.zip) — tell me how you'd prefer to receive it and I'll upload it.