Hugging Face Transformers v5.19.0 released with EmbeddingGemma2 support, MoE router logits changes, and continuous batching updates

Hugging Face released version 5.19.0 of the Transformers library, adding support for Google’s EmbeddingGemma2, a multimodal embedding model that encodes text, images, audio, and video into a shared 768-dimensional vector space for tasks such as cross-modal retrieval, semantic similarity, clustering, and classification.

The update includes Matryoshka Representation Learning, enabling embeddings to be truncated to 512, 256, or 128 dimensions, and supports configurable visual and video token budgets with the ability to disable unused vision or audio towers at load time to reduce memory usage.

Breaking changes affect mixture-of-experts (MoE) models: when output_router_logits=True, routers now return logits following the Qwen3-MoE pattern, changing the output structure for code that previously relied on their absence. Object detection with Owlv2 now selects the query box with the highest objectness score, altering image-guided query embeddings and detections compared to earlier releases.

Attention mechanisms are updated: the 'paged|' prefix for SDPA and flash attention implementations is deprecated in favor of direct sdpa or flash_attention_2 for continuous batching, with regular implementations now supporting continuous batching directly. Cache update logic in continuous batching has been fused into a single call, changing behavior for custom code that calls separate update paths.

Expert parallelism gains a token-dispatch implementation selected via the new ep_dispatch_experts plan rule and is now the default for Qwen3 MoE and Mellum, removing the requirement that expert parallelism size equal tensor parallelism size. The Trainer was adapted to work with expert parallelism, and documentation notes that PEFT adapters support tensor parallelism.

Cache handling is improved: generate no longer mutates user cache_config, QuantizedLayer.reorder_cache is repaired, and per-layer cache configuration is added so DynamicCache and StaticCache initialize each layer from its own config to better support heterogeneous models. The deprecation cycle on mask and cache has also been removed.

Additional fixes include reverting a temporary CI change for AMD, correcting DataCollatorForLanguageModeling to respect seed=0, addressing explicit kernelization modes for evaluation models, and updating GPT-OSS configuration handling for SwiGLU parameters.