bug: winml build shows misleading ValueError (opset domain) when disk is full during quantize step

[DingmaomaoBJTU]

Summary

When disk space is exhausted during winml build, the optimize step writes a corrupted/truncated ONNX file. The subsequent quantize step then fails with ValueError: Failed to find proper ai.onnx domain, giving no indication that the real cause is an out-of-disk-space condition. Users will chase a phantom code bug.

Repro

# Fill disk to near-zero free space, then:
winml build -c config.json -m microsoft/resnet-50 -o model_a_config/

Error

ValueError: Failed to find proper ai.onnx domain in
onnxruntime.quantization.quant_utils.get_opset_version

Stack trace: quantizer.py:166 → quantize.py:909 → quantize.py:702 → quant_utils.py:984 → quant_utils.py:977 → ValueError

Root Cause

The optimize step writes a truncated/zero-byte ONNX output due to OSError (disk full). The corrupted file is passed to the quantizer, which fails trying to parse the opset domain. The OSError is swallowed or not surfaced to the user.

Expected behavior

ModelKit should either:

1. Check available disk space before winml build starts and warn if below a threshold, or
2. Catch OSError during ONNX file writes and surface a clear error: "Insufficient disk space — unable to write output file."

Notes

• Severity: P2
• Reporter: Agency bug bash feedback
• Context: Full bug bash consumes significant disk — venv (~1.4 GB), HuggingFace cache (~14.4 GB), per-model artifacts (~200–400 MB each). Minimum ~20 GB free disk recommended.

[DingmaomaoBJTU]

Triage summary

• Component: Sys
• Priority: P2
• Owner(s): @chinazhangchao

winml build shows a misleading ValueError about opset domain when disk is full during the optimize/quantize step; catch OSError during ONNX file writes and surface a clear 'insufficient disk space' error.