WIP: GHA test with model revisions

[ckadner]

Description

Cache models with revision for GitHub Action test runs.

TODOs:

• rebase this PR onto main (currently based on @prashantgupta24 branch)
  --> [CI] Enable model revisions in GHA test #523
• load model with revision in unit tests (not revision: main)
  https://github.com/vllm-project/vllm-spyre/actions/runs/18189810233/job/51782179337?pr=499#step:14:258

  > ☑️ Download HF models
  Downloading 'ibm-ai-platform/micro-g3.3-8b-instruct-1b' with revision '2714578f54cfb744ece40df9326ee0b47e879e03' ...
  
  
  > ❌ Runt tests
  ...
  head_call_error = OfflineModeIsEnabled("Cannot access file since 'local_files_only=True' as been set. (repo_id: ibm-ai-platform/micro-g3.3-8b-instruct-1b, repo_type: model, revision: main, filename: config.json)")
  

Related Issues

#497

[joerunde]

I think all of this can be simplified to a single function like:

from transformers import pipeline
from sentence_transformers import SentenceTransformer

try:
    pipeline(model='$1', revision='$2')
except RuntimeError:
    SentenceTransformer('$1', revision='$2')

I'd be fine sticking that in a script that we keep around here in the .github/workflows folder.

python3 ./just_download_the_dang_model.py --model foo --revision bar

That should get rid of a lot of the complexity here. Otherwise these changes all look great to me- putting the revisions in the cache key and splitting out the multi-model cases to have individual cache entries is exactly what we want. Sure hf_model_2 is a little wonky but I can live with it.

[ckadner]

✅ #523

[ckadner]

Need to rebase this PR onto main (currently it's based on top of Prashant's branch)

[ckadner]

Closing this. Changes from this PR are rebased onto main in Pr #523