[For maintainers] Suggested jobs to run (before merge)

run-slow: idefics2, idefics3, smolvlm

run-slow: idefics2, idefics3, smolvlm

This comment contains run-slow, running the specified jobs:

models: ['models/idefics2', 'models/idefics3', 'models/smolvlm']
quantizations: [] ...

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

-slow: idefics2, idefics3

I'm struggling a bit to understand where exactly the tests have failed . I can see a test_model_parallel_beam_search test has failed .
Is there anything else that has failed ? And any pointers on how I can work on fix that would be really appreciated.
Thank you @zucchini-nlp for looking into the PR

On my local machine after running python3 -m pytest -v -rsfE --make-reports=multi-gpu_run_models_gpu_models/smolvlm_test_reports tests/models/smolvlm

I got the result :
244 passed, 126 skipped, 18 warnings in 111.25s (0:01:51)

I don't have a multi gpu system so I'm not sure if that's why I have more skips

@i3hz the only failing test is tests/models/smolvlm/test_modeling_smolvlm.py::SmolVLMForConditionalGenerationIntegrationTest::test_integration_test_video which seems to be unrelated

Btw for running locally you also need to set RUN_SLOW=1 which runs slow tests as well. Can you check that the export test passes, then we can merge

tests/models/smolvlm/test_modeling_smolvlm.py::SmolVLMForConditionalGenerationIntegrationTest::test_export_smolvlm_connector
tests/models/smolvlm/test_modeling_smolvlm.py::SmolVLMForConditionalGenerationIntegrationTest::test_export_smolvlm_text_decoder
tests/models/smolvlm/test_modeling_smolvlm.py::SmolVLMForConditionalGenerationIntegrationTest::test_export_smolvlm_vision_encoder

@zucchini-nlp I ran the RUN_SLOW=1 pytest tests/models/smolvlm/test_modeling_smolvlm.py command .

With these results
8 failed, 176 passed, 93 skipped, 11 warnings in 182.77s (0:03:02)

The good news is that all three export tests passed successfully-

• test_export_smolvlm_connector
• test_export_smolvlm_text_decoder
• test_export_smolvlm_vision_encoder

As for the failures -

• test_integration_test_video
• The other 7 are regarding FlashAttention . They all failed with the same error: RuntimeError: cu_seqlens_q must have shape (batch_size + 1). I could be wrong but I think these are unrelated .

Let me know what you think .
Thanks