transformers
4cc2aa84 - [Qwen3VLMoe] Update expected text in test_small_model_integration_test_expand (#49301)

Commit
40 days ago
[Qwen3VLMoe] Update expected text in test_small_model_integration_test_expand (#49301) After #45783 ("Encode multimodal data only once"), the vision encoder now processes the image once (batch=1) and shares embeddings across beams, instead of processing a duplicated batch (batch=N beams). This changes bf16 layer-norm statistics and shifts beam ranking at the margin, producing a slightly different beam[1] output. Both outputs are semantically correct. Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>
Author
Parents
Loading