perf(inspect): dedupe AutoConfig + gate processor Auto* lookups (#543)

[timenick]

Summary

Follow-up to #718 (banner + spinner). The spinner removed the perceived-hang UX problem but the total wall time was still ~23 s. This PR cuts the resolver-side redundancy to bring that to ~18 s.

Profile (warm cache, microsoft/resnet-50)

Before this PR (per-call counts inside _inspect_model_v2):

0.71s  AutoConfig.from_pretrained   ← explicit fetch in inspect.py
0.38s  AutoConfig.from_pretrained   ← duplicate inside resolve_loader_config
0.72s  resolve_io_specs
0.38s  hf_hub_download              ← preprocessor_config.json
0.31s  hf_hub_download              ← tokenizer_config.json (404)
0.70s  AutoTokenizer.from_pretrained
0.71s  AutoImageProcessor.from_pretrained
3.53s  AutoProcessor.from_pretrained
0.38s  AutoConfig.from_pretrained
0.70s  AutoTokenizer.from_pretrained
0.69s  AutoImageProcessor.from_pretrained
4.20s  AutoFeatureExtractor.from_pretrained
─────
12.33s total inspect work

After this PR:

… (same fast portion) …
0.69s  AutoImageProcessor.from_pretrained
3.49s  AutoProcessor.from_pretrained
0.70s  AutoTokenizer.from_pretrained
─────
9.29s total inspect work

AutoConfig calls 5 → 4, AutoFeatureExtractor entirely eliminated, duplicate AutoImageProcessor skipped.

End-to-end wall time (uv run winml inspect -m microsoft/resnet-50 --format json):

Branch	run 1	run 2	run 3
main	23.4 s	23.2 s	23.0 s
this PR	17.6 s	17.7 s	18.8 s

Change

1. Dedupe AutoConfig.from_pretrained

_inspect_model_v2 already loads the parent hf_config (it needs the un-narrowed config for I/O introspection — resolve_loader_config may narrow it to a sub-config for multimodal models like CLIP). Then resolve_loader_config loaded it again.

Added an hf_config= kwarg to resolve_loader_config — when supplied, step 1's AutoConfig.from_pretrained is skipped. Pass-through from _inspect_model_v2. Other callers default to hf_config=None, original behaviour preserved.

2. Gate processor Auto* lookups per-field

resolve_processor ran every Auto* class instantiation in Strategy 2, even when Strategies 0+1 had already populated the corresponding field. Added try_* kwargs to _resolve_processor_from_auto_classes and skip each Auto* whose field is already known.

For ResNet, Strategies 0+1 set image_processor_class from the HF registry, so try_image_processor=False skips the redundant AutoImageProcessor.from_pretrained. AutoProcessor succeeds and supplies feature_extractor_class as a side effect, so AutoFeatureExtractor (the 4.2s call) is skipped via the existing inner gate.

If Strategies 0+1 already populate all four fields (CLIP-like models), Strategy 2 is skipped entirely.

Why not further?

The remaining 9 s of resolver work is inside transformers' Auto* internals — AutoProcessor.from_pretrained does its own AutoConfig load and may construct sub-processors. Cutting deeper requires upstream restructuring, not a ModelKit-side fix.

Tests

• test_resolve_loader_config.py::test_hf_config_kwarg_skips_autoconfig_fetch — pins the dedupe by patching AutoConfig.from_pretrained and asserting call_count == 0 when hf_config= is supplied.
• test_resolve_processor_gating.py — three cases:
  • All four fields filled by Strategy 1 → _resolve_processor_from_auto_classes is not called at all
  • Partial fill → only the still-needed try_* flags are True
  • Nothing filled → every try_* flag is True

112 targeted tests pass.

Closes #543.

[DingmaomaoBJTU]

Overall this is a well-motivated performance optimization with clean implementation and good test coverage. Two comments inline.

[DingmaomaoBJTU]

Well-motivated perf optimization with clean implementation and solid test coverage. Two inline comments.