feat(session): add GenaiSession for onnxruntime-genai pipeline inference

[DingmaomaoBJTU]

Closes #1011

Introduces GenaiSession, a generic wrapper around onnxruntime-genai (og.Model + og.Generator) for autoregressive text generation using multi-model decoder pipelines (ctx + iter + embeddings + lm_head), plus a winml-genai perf runtime so winml perf can benchmark these bundles.

Usage

Benchmark a genai bundle directory (one containing genai_config.json, as produced by a winml-cli export) with winml perf --runtime winml-genai:

# Mixed execution (default): ctx/iter route to the accelerator via each stage's
# session_options, embeddings/lm_head stay on CPU.
winml perf -m path/to/bundle --runtime winml-genai

# Pre-compile EPContext-capable stages (e.g. QNN) before benchmarking:
winml perf -m path/to/bundle --runtime winml-genai --compile \
  --prompt "Explain the theory of relativity in simple terms." \
  --max-new-tokens 128 --iterations 10 --warmup 2

genai-specific flags:

Flag	Default	Description
--runtime winml-genai	winml	Benchmark an onnxruntime-genai bundle (LLM generation) instead of single-shot ONNX inference.
--device	auto	auto→mixed, npu→qnn, gpu→dml, cpu→CPU-only.
--compile / --no-compile	off	Pre-compile EPContext-capable stages before the run (JIT otherwise).
--compile-timeout	300	Max seconds to compile each EPContext stage before falling back to the original ONNX.
--prompt	built-in	Prompt to generate from (wrapped in the bundle's own chat template before timing).
--max-new-tokens	128	New tokens generated per iteration.
--iterations / --warmup	10 / 2	Timed / warmup generations (genai-specific low defaults; a full generation is far costlier than one session.run()).
-o, --output	auto	JSON report path (~/.cache/winml/perf/<bundle>/<ts>.json otherwise).

Reported metrics come straight from onnxruntime-genai call boundaries — TTFT/prefill measured around append_tokens, decode tokens/sec and per-output-token latency (TPOT) around generate_next_token.

The benchmark prompt is wrapped in the bundle's own chat template (via og.Tokenizer.apply_chat_template) before timing, so prefill reflects realistic chat usage; bundles that ship no template benchmark the raw prompt.

Programmatic use:

from winml.modelkit.session import GenaiSession, GenerationConfig

with GenaiSession("path/to/bundle", ep="mixed") as sess:
    for tok in sess.generate_streaming("Hello, world!", GenerationConfig(max_new_tokens=64)):
        print(tok, end="", flush=True)

Changes

src/winml/modelkit/session/genai_session.py (new)

• GenaiSession — context manager wrapping og.Model + og.Generator
• generate_streaming() / generate() — streaming and full-string generation; generate_timed() returns a GenerationTiming with TTFT/TPOT sourced from ort-genai call boundaries
• EPContext pre-compilation via _prepare_compiled_bundle(); each pipeline stage's EP is resolved generically from its genai_config.json provider_options via WinMLCompileConfig.for_provider() (QNN, OpenVINO, VitisAI, …) — no provider name is hardcoded — and compiled in an isolated subprocess that delegates to the shared compiler component (compile_onnx) with a configurable compile_timeout (default 300 s). Stages on non-EPContext EPs (CPU, DML) stay on JIT. A .meta.json marker invalidates a cached stage when its provider options change or the source weights (.onnx / .data sidecar) are newer.
• WinML EP auto-registration for QNN/DML/mixed bundles; static KV cache length read from genai_config.json
• apply_chat_template() renders the prompt through the bundle's own chat template (model-driven via og.Tokenizer; no chat format hardcoded)
• _mirror_non_onnx_files() symlinks non-compiled assets into the _compiled/ subdir
• GenerationConfig / GenerationTiming dataclasses and GenaiSessionError / GenaiNotInstalledError / GenaiLoadError hierarchy

src/winml/modelkit/commands/_perf_genai.py (new) — the winml-genai perf runtime: GenaiPerfConfig, the generation benchmark loop, LLM metric aggregation (TTFT percentiles, decode tokens/sec, avg output-token latency), and rich/JSON reporting.

src/winml/modelkit/commands/perf.py — adds --runtime {winml,winml-genai}, --prompt, --max-new-tokens, and --compile-timeout. winml-genai takes a bundle directory and defaults --device auto to mixed execution so each stage's per-stage EP routing (ctx/iter on the accelerator, embeddings/lm_head on CPU) is honored. winml-only flags are warned-and-ignored; genai imports are function-local so perf --help stays fast.

src/winml/modelkit/session/__init__.py — exports GenaiSession, GenerationConfig, GenerationTiming, and the exception classes.

pyproject.toml — adds onnxruntime-genai-winml==0.14.1.

Tests — tests/unit/session/test_genai_session.py and tests/unit/commands/test_perf_genai.py (mock onnxruntime_genai, no hardware needed): init/load lifecycle, EP registration, streaming, chat-template application, generic per-EP compile dispatch, EPContext cache invalidation, _prepare_compiled_bundle, _mirror_non_onnx_files, _patch_stage_filename, compile_timeout propagation, timing math with an injected clock, and full CLI dispatch/validation.