transformers
9c64f0c6 - fix(tests): restore ModelOnTheFlyConversionTester after token revocation and Space API migration (#50075)

Commit
24 days ago
fix(tests): restore ModelOnTheFlyConversionTester after token revocation and Space API migration (#50075) * fix(tests): update ModelOnTheFlyConversionTester to use hf-transformers-bot token and hf-internal-testing org Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * fix thread name mismatch in test_absence_of_safetensors_triggers_conversion Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * fix safetensors Space API URL and increase thread join timeout - Update spawn_conversion() to use /gradio_api/call/run (Gradio v5 migration changed the endpoint from /call/run) - Increase thread join timeout from 10s to 60s to give the Space enough time to respond (typically needs 20-30s) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * fix(tests): unset DISABLE_SAFETENSORS_CONVERSION in ModelOnTheFlyConversionTester conftest.py globally sets DISABLE_SAFETENSORS_CONVERSION=true to avoid Hub traffic during regular test runs (PR #41194). ModelOnTheFlyConversionTester specifically exercises the conversion path, so it needs to unset the var in setUp and restore it in tearDown. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * fix(config): forward token kwarg to get_config_dict in PreTrainedConfig.from_pretrained token was an explicit parameter of from_pretrained but was never added to kwargs before calling get_config_dict, causing it to arrive as None in cached_files. This broke loading from private repos when token= was passed as a kwarg (it only worked if HF_TOKEN env var was set). Also update ModelOnTheFlyConversionTester: - re-enable private conversion tests (now work with the token fix) - fix assertion: private PR author is "hf-transformers-bot", not self.user - skip gated tests with a note: SFconvertbot cannot agree to gating on dynamically-created repos (needs further investigation, may have worked previously with huggingface-hub-ci org setup) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * skip private/gated conversion tests; revert author assertions Private repo conversion tests are skipped — no strong need to test this path since SFconvertbot opens PRs for public repos, and private repo behaviour is covered by the Space separately. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * re-enable private conversion tests (they pass with token fix) Private repo conversion works since configuration_utils.py now correctly forwards `token` to `get_config_dict`. PR author for private repos is `hf-transformers-bot` (the token owner). Only gated tests remain skipped (Space-side limitation). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * use TRANSFORMERS_HUB_BOT_HF_TOKEN (with HF_TOKEN fallback) in conversion tester Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * rename cls.user → cls.org (hf-internal-testing is an org, not a user) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * clarify token requirement and skip message in ModelOnTheFlyConversionTester Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * use hf-ci-ephemeral-repos org and HF_CI_EPHEMERAL_REPO_MANIPULATION_TOKEN Switch to a dedicated org and fine-grained token scoped only to ephemeral repo create/delete + discussions, instead of the broader TRANSFORMERS_HUB_BOT_HF_TOKEN. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * add inline comment explaining thread.join(timeout=60) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> --------- Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>
Author
Parents
Loading