Optimize analyze performance.

[chinazhangchao]

1. pattern/utils.py — make_hashable fast-path for sequences of simple types (avoids 57M unnecessary recursive calls)
2. analyze/core/runtime_checker_query.py — Replaced pandas .apply(make_hashable) with direct list comprehension over numpy array, bypassing pandas per-cell dispatch overhead (27s → ~2s)
3. analyze/core/doc_constraint_checker.py — Added skip_shape_inference parameter to avoid redundant infer_shapes() call (~3.3s saved)
4. analyze/core/information_engine.py + analyze/analyzer.py + analyze/core/runtime_checker.py — Pass shape-inferred model proto from RuntimeCheckerQuery to DocConstraintChecker
5. export/init.py — Lazy-load io and pytorch submodules (avoids importing optimum.exporters.tasks + transformers at startup, ~29s)
6. compiler/init.py — Lazy-load Compiler, stages (avoids importing session → torch, ~4s)
7. core/init.py — Lazy-load generate_dummy_inputs_from_specs (avoids importing torch, ~3.6s)
8. config/build.py — Lazy-import ONNXConfigNotFoundError (avoids importing export/io.py at module load)
9. analyze/utils/json_utils.py — Lazy-import jsonschema (~3.5s saved)

[DingmaomaoBJTU]

Code review

Found 1 issue:

1. export/__init__.py's __getattr__ never caches loaded symbols back into the module's __dict__ via globals().update(...), so every subsequent attribute access re-triggers __getattr__ instead of hitting the module dict directly. The same PR's compiler/__init__.py and core/__init__.py both call globals().update(...) / globals()[name] = val after loading to avoid this, but export/__init__.py does not.

https://github.com/microsoft/ModelKit/blob/2a263478058c40d3ef4d0cc2494ea9787a5d9dda/src/winml/modelkit/export/__init__.py#L29-L46

For comparison, compiler/__init__.py (same PR) correctly caches:

https://github.com/microsoft/ModelKit/blob/2a263478058c40d3ef4d0cc2494ea9787a5d9dda/src/winml/modelkit/compiler/__init__.py#L35-L47

🤖 Generated with Claude Code

_{- If this code review was useful, please react with 👍. Otherwise, react with 👎.}