microsoft/winml-cli v0.0.4

WinML CLI v0.0.4

Repository: microsoft/winml-cli

Tag: v0.0.4

Published: 2026-05-20T13:18:57Z

Prerelease: yes

Release notes:

🚀 Platform upgrades

• Windows ML → 2.0 (#441)
• Python → 3.11 (requires-python = ">=3.11,<3.12")

⚠️ Behavior changes

• Incompatible --ep / --device pairs are now rejected instead of silently overridden (#641, #661).
• winml config/build --device npu exits non-zero when no compatible NPU EP is available (#660).
• winml analyze --ep cpu resolves to CPU instead of falling back to NPU (#641).
• trust_remote_code now prints a bold-red stderr warning whenever it is honoured (#641).

✨ Improvements

• winml build writes analyze_result.json to the output folder (#673) and validates the config up-front (#675).
• Exported ONNX is auto-normalized via optimize_onnx() (#681).
• winml inspect distinguishes local-path-not-found from network errors (#679).

🐛 Fixes

• --run-unknown-op compile=false regression (#662).
• winml build --device npu failing with quant.task is required (#673).
• HF build path dropped explicit --ep on compile-less paths (#678).
• run_eval.py not forwarding --device to winml build (#674).
• Seq2seq decoder calibration crash on image-to-text models (#671).

🔧 Internals & tests

• Strong-typed EP parameters across analyze/compiler/optracing (#632).
• EP_SUPPORTED_DEVICES as single source of truth (#641).
• Expanded E2E / CLI surface tests for analyze, compile, inspect, catalog, and perf (#645, #652, #661, #665, #669, #672, #676).

Full changelog: https://github.com/microsoft/winml-cli/blob/release/v0.0.4/CHANGELOG.md