fix(ml): drop resume_download, removed from from_pretrained in transformers v5 (#15054)

[mrveiss]

Thinking Path

transformers is pinned at 5.15.1 (requirements-ci/ai-ml.txt:10 -- the issue's
body cited 5.15.0, but the requirements file at the base branch pins 5.15.1).
Before touching any call site, I read that exact version's own source
(.venv-python-suite/lib/python3.14/site-packages/transformers/ -- already
installed locally for an unrelated purpose, read-only, never imported/executed)
to establish what resume_download actually does now rather than assume from
memory:

• PreTrainedModel.from_pretrained (modeling_utils.py) never pops
  resume_download. It forwards unrecognized kwargs into
  config_class.from_pretrained(..., return_unused_kwargs=True, **kwargs), and
  the still-unused remainder becomes model_kwargs, which reaches
  cls(config, *model_args, **model_kwargs). A concrete model/processor class
  (CLIPModel, Wav2Vec2Model, WhisperForConditionalGeneration, ...) has no
  **kwargs catch-all on __init__, so this is a live TypeError --
  exactly the CI reproduction the issue cites: CLIPModel.__init__() got an unexpected keyword argument 'resume_download' (run 32839088246).
• AutoConfig.from_pretrained / AutoTokenizer.from_pretrained survive it
  instead: PretrainedConfig.__post_init__ stores any leftover kwarg as a bare
  instance attribute (configuration_utils.py -- "Additional attributes without
  default values"), and PreTrainedTokenizerBase.__init__ just leaves it sitting
  in init_kwargs, unused. Both silently swallow it -- latent debt, not a
  crash, on that path.

So the breakage is mixed by call shape, not uniform -- but the fix is the same
either way: resume_download has no replacement kwarg (resumption is automatic
in v5), so it is dropped everywhere rather than substituted.

Call-site count: a repo-wide grep for from_pretrained( (git-tracked .py,
prose-stripped to exclude a docstring mention in hf_quantizer.py:97) finds
22 real call sites, of which 18 carried resume_download -- not the
issue's 10. The issue's table missed
autobot-backend/multimodal_processor/processors/voice.py entirely (4 sites)
and only listed 1 of vision.py's 5. The other 4 (a pyannote Pipeline.from_pretrained
in media/audio/diarization_service.py, and 3 trust_remote_code-only calls in
autobot-npu-worker/resources/windows-npu-worker/app/npu_worker.py, which pins
transformers>=4.36.0 separately) never passed resume_download and are untouched.

What Changed

• Removed resume_download=True from all 18 call sites across
  ai_hardware_accelerator.py, code_embedding_generator.py,
  llm_shared/optimization/layer_inference.py,
  llm_shared/optimization/model_inspector.py,
  multimodal_processor/processors/vision.py, and
  multimodal_processor/processors/voice.py. No replacement kwarg -- v5
  resumption is automatic.
• ai_hardware_accelerator.py shrank by 2 lines from the deletion; lowered its
  ratchet ceiling from 1048 to 1046 in both
  scripts/python_file_size_known_large.py and
  repo_tests/python_file_size_ratchet_baseline.py (the mirrored pair the
  ratchet requires) -- a genuine shrink, not a raised ceiling.
• The vision/voice "load failure surfaced, not swallowed into a
  degraded-but-ready processor" acceptance criterion is already satisfied by
  existing code (vision.py:351's clip_model is None or blip_model is None
  gate from Implement real multimodal processing with NPU integration #466, and processor.py's get_stats() "model_availability"
  block from [Backend] Vision and audio model loading failures #675) -- verified by reading, not modified further.
• Added repo_tests/transformers_resume_download_guard_test.py: a durable guard
  that walks every git-tracked production .py file, finds every
  .from_pretrained(...) call (paren-depth tracked so a multi-line call is
  captured whole, prose stripped first so a docstring mention is never counted),
  and fails if any call span contains resume_download.

Verification

• Read transformers 5.15.1's own source
  (modeling_utils.py, configuration_utils.py, tokenization_utils_base.py,
  processing_utils.py) to establish the TypeError-vs-silently-swallowed split
  above -- no from_pretrained call was executed and no model was downloaded.
• grep -rn resume_download --include='*.py' . on the tree after the change: no
  matches.
• Reach floor: the guard's FROM_PRETRAINED_REACH_FLOOR = 15 sits between the
  tree's real total (22) and a broken walk's 0 -- test_the_from_pretrained_sweep_reaches_the_repo
  fails loudly if the sweep stops finding call sites instead of reporting a
  false "no offenders" having checked nothing.
• Contrast mutation: test_a_reintroduced_resume_download_is_caught reproduces
  the exact pre-fix vision.py CLIPModel.from_pretrained(...) call
  (resume_download=True on the call's second line) and asserts the detector
  flags it -- proving the guard isn't vacuously passing.
• The git pre-push hook ran pytest on every test file this PR's changes touch
  (code_embedding_generator_test.py, layer_inference_test.py,
  model_inspector_test.py, and the new guard) and reported all green -- the
  hook's own gate, not a manual pytest invocation, per the no-run-pytest
  constraint on this task.

Closes #15054

Model Used

Claude Sonnet 5

[github-actions]

✅ SSOT Configuration Compliance: Passing

🎉 No new hardcoded values of either class — ssot and other both block.

Known backlog in pipeline-scripts/hardcoded_values_baseline.txt is suppressed and tracked in #14371.