Repository navigation
This cycle introduces CGC workflows for the first time, adds text-generation benchmarking, GPU execution timing, and AMD NPU operator profiling, and improves runtime startup and memory reporting. See the behavior changes below.
⚠️ Behavior changes
- EP discovery - listing providers no longer downloads missing Catalog providers; acquisition requires an explicit EP/device request (#1508).
✨ Improvements
- CGC support (first release) - introduces CGIR export, compilation, inference, and evaluation through Windows ML Runtime and WinMLCG EP, including GPU timing, memory measurements, and experimental text-generation benchmarks for local GGUF and safetensors models (#1426, #1467, #1492, #1503, #1505, #1518).
winml perf- adds experimental GGUF text-generation benchmarks through the Windows ML Tasks llama.cpp backend (#1468).- Operator profiling - adds basic VitisAI tracing for AMD NPU with per-operator timings and raw trace artifacts (#1516).
- EP loading -
WINMLCLI_EP_PATHaccepts individual DLL paths as well as directories, including custom filenames with explicit EP selection (#1508). - CLI usability - improves help and console accessibility, reduces perf/eval startup waits, and lowers GGUF startup memory use (#1485, #1533, #1494).
🔧 Internals & CI
- E2E coverage - adds CGC cases and optional runtime-wheel overrides, fixes E2E failures, and shares OpenVINO GPU allocators within test sessions (#1532, #1527, #1531, #1480).
- Contributor workflows - refines model-support and auto-optimization skill boundaries and restores their evaluation inputs (#1495, #1496, #1497, #1501).
📦 Assets
winml_cli-0.5.0-py3-none-any.whlrules-v0.5.0.zip