WinML CLI v0.3.0

This cycle expands model preparation and evaluation across the CLI: precision-driven quantization, composite-model and dynamic-axis workflows, real-input perf/eval, optimization previews, and opt-in Dynamo export. It also introduces one-command Qwen3 onnxruntime-genai bundles, GenAI benchmarking, broader model recipes, and more reliable EP discovery, compilation, and monitoring. See the behavior changes below.

⚠️ Behavior changes

• Output-producing commands now refuse to replace existing files or non-empty directories unless --overwrite is passed; winml build retains its existing incremental --rebuild behavior (#970).
• winml quantize renames --model-name to --model-id, including the corresponding quantization-config field (#984).
• Compile configuration no longer silently defaults a missing execution provider to QNN; incomplete configurations now fail validation instead (#1026).
• winml inspect / winml perf hide third-party and native warning noise by default; use -v, -vv, or WINMLCLI_SHOW_ALL_WARNINGS=1 to restore diagnostics (#1232, #1246).

✨ Improvements

• Quantization — --precision selects FP16 conversion, RTN INT4, static QDQ, or calibration-free dynamic INT8; winml quantize can compose multiple precision passes such as INT4 followed by FP16 (#872, #985, #1047).
• winml build — --export-type optimized produces a complete Qwen3 onnxruntime-genai NPU/QNN bundle, including prefill/decode, embeddings, LM head, tokenizer, and manifest files (#836, #996, #1008, #1081, #1104).
• winml perf --runtime winml-genai — benchmarks prebuilt or automatically cached GenAI bundles with TTFT, token throughput, prompt-template controls, EP overrides, and isolated pre-compilation (#1015, #1042, #1046, #1054, #1109).
• Composite models — export and build automatically fan out pipeline components; export / build / perf support --submodel, and explicit composite tasks such as summarization and translation are accepted (#1031, #1037, #1058, #1071, #1089).
• Export controls — dynamic axes and symbolic input dimensions are supported while static TorchScript export remains the default; build, config, perf, and eval expose matching shape/input/export overrides (#1074, #1083, #1106, #1141, #1156, #1188).
• winml perf — real .npz inputs, time-budgeted --duration runs, cached per-module builds, actual dynamic dimensions, and QNN profiler ONNX metrics (#1004, #1055, #1066, #1102, #1168).
• QNN op tracing — per-model tracing can be enabled automatically; basic traces exclude warmup samples and null fields, while detail tracing accepts compile EP options and auto-compiles raw ONNX inputs when required (#1006, #1032, #1249, #1252).
• winml eval --mode compare — compares two ONNX models directly or uses real .npz samples against a Hugging Face reference; Qwen3 adds perplexity evaluation (#1139, #1209, #1221).
• winml optimize / winml analyze — --check-optim previews applicable rewrites and verifies their produced operators against the target EP; new rewrites cover static Split-to-Slice and Conv affine/BatchNormalization folding (#1142, #1167, #1171, #1238, #1257).
• EP discovery and monitoring — registration is isolated and failures are structured, startup remains lazy, op-tracing dispatch is unified, and provider-download progress is restored (#1019, #1239).
• CLI quality of life — EP/device and pipeline-stage flags are consistent across commands; --no-color disables ANSI output for one invocation (#923, #978, #992).
• Hub-hosted ONNX — commands accept <org>/<repo>/<path>.onnx references from Hugging Face Hub, enabling SAM 3 encoder/decoder workflows (#582).
• Keypoint detection — ViTPose supports config, build, and perf, plus COCO OKS-AP evaluation (#905, #949).
• Vision and document recipes — refreshed coverage adds DINOv2, SwinV2, OWL-ViT/OWLv2, BEiT, SegFormer, YOLOS, ViTPose/SynthPose, LayoutLM/LayoutLMv3, and document/question-answering models (#925, #1064, #1088, #1093, #1100, #1101, #1123, #1125, #1145, #1155, #1173, #1174, #1178, #1187, #1201, #1202, #1205, #1208).
• Language recipes — expanded BART, BERT, DeBERTa, DistilBERT, KoELECTRA, MiniLM, MPNet, Marian/OPUS, GTE reranker, feature-extraction, and entity-linking coverage (#1068, #1080, #1112, #1115, #1116, #1117, #1118, #1120, #1121, #1124, #1134, #1143, #1144, #1153, #1169, #1170, #1179, #1200, #1214).
• Audio recipes — expanded Wav2Vec2, HuBERT, AST, MMS, language/gender/music classification, forced alignment, and multilingual ASR coverage (#1094, #1095, #1114, #1131, #1148, #1154, #1176, #1186, #1206, #1207, #1211, #1225).

🐛 Fixes

• winml perf — throughput uses the batch size actually executed; analyzer EP resolution and op-trace paths now match the runtime target (#930, #941, #1000).
• winml build — honors explicit --ep, supports non-compiling cross-target builds, keeps ONNX caches distinct by resolved path and configuration, preserves configured model classes, reports disk-full failures clearly, and keeps CPU/GPU automatic precision at FP32 (#856, #947, #987, #997, #998).
• GenAI and composite export — fixed component export/build failures, compile fallback paths, accelerator selection, isolated EPContext preparation, and final Qwen3 bundle assembly (#1037, #1051, #1103, #1138, #1248).
• Task and model resolution — reconciled the task registry, corrected model-specific task listings and Hub pipeline_tag fallback, accepted composite tasks, and resolved CTC-based ASR model classes correctly (#724, #986, #1070, #1071, #1113, #1154).
• Depth and keypoint evaluation — fixed inference-time evaluator failures (#1023).
• Analyzer and optimizer rules — corrected coverage counting, aligned pattern checks with node support, consolidated recommendation metadata, and fixed dtype constraints and unknown-pattern handling (#922, #1020, #1063, #1130, #1162).
• EP / device resolution — WindowsML catalog providers register correctly; device listings retain hardware details without duplicate aliases; analyzer auto-selection prefers the strongest exact target; CPU bridge providers resolve safely; invalid EP/device pairs fail early (#1076, #1220, #1227, #1228, #1231, #1237).
• Native EP execution — hardened spawned-provider progress, prevented compiler-output deadlocks, released native sessions before process exit, and replaced pipe-backed warning capture with a bounded file spool to avoid EP compiler hangs (#1017, #1223, #1230, #1266, release cherry-pick #1267).
• Export and quantization — standalone quantization suppresses duplicate ORT preprocessing warnings; decoder KV-cache dimensions survive tracing; large external-data models can convert to FP16; and EPs that quantize internally no longer receive redundant WinML quantization (#956, #1176, #1235, #1242).
• QNN evaluation and tracing — repaired evaluation failures, detail-trace DLL/summary handling, and compile-time provider options (#1247, #1249).
• winml eval — default text datasets and sentiment recipes use fully qualified Hugging Face dataset IDs (#1262, release cherry-pick #1263).
• CLI help — winml --help shows the correct sys summary and concise, untruncated build / quantize descriptions (#1254).
• Telemetry — local Path model references no longer cause successful commands to fail during telemetry scrubbing (#1273).

🔧 Internals & CI

• Release pipelines — E2E aligns ModelKitArtifacts with the matching release branch, stable GitHub releases receive CHANGELOG notes and “Latest” status, and the official-build toolchain is pinned for reproducibility (#940, #967, #1268).
• Evaluation CI — recipe-driven build/eval supports per-EP matrices, pre-exported ONNX, reliable resume behavior, actual applied-precision reporting, unquantized-track EPs, and broader MIGraphX/TensorRT RTX coverage (#845, #902, #1009, #1039, #1086, #1160, #1163, #1226, #1243, #1286).
• Telemetry — action events record scrubbed model identifiers, while error events retain scrubbed root-cause details for diagnosis (#1108, #1111).
• Development environment — expanded type checking, added a tracked uv lockfile, selected CPU-only PyTorch wheels, and consolidated development dependencies (#932, #957, #1105, #1251, #1255).
• Documentation publishing — added and published the version-stamped model accuracy report from the current documentation site (#974, #975, #979, #1203).

📦 Assets

• winml_cli-0.3.0-py3-none-any.whl
• rules-v0.3.0.zip