Transformers 5.19 adds EmbeddingGemma 2 support, stronger MoE expert parallelism and broader continuous batching

What changed

Hugging Face released Transformers v5.19.0 on October 6. The release adds support for Google's EmbeddingGemma 2, which maps text, images, audio and video into a shared 768-dimensional embedding space and supports Matryoshka truncation to 512, 256 or 128 dimensions. The framework also adds token dispatch for expert parallelism, makes it the default for Qwen3 MoE and related models, adapts Trainer to expert parallelism, broadens continuous batching to regular SDPA and Flash attention, and introduces per-layer cache configuration. The release contains breaking changes around MoE router logits and paged-attention behavior.