v0.6.0

Added

• Notebook about Retrieval-Augmented Generation (RAG) (#353)
• Phi model (#356)
• Gemma model (#358)
• Support for input streaming in Whisper serving (#361)
• Stable Diffusion ControlNet model (#359)
• Support for multiple EOS tokens in text generation (#368)
• Phi-3 model (#374)
• Support for Llama 3 (#387)
• Support for CLS token pooling in text embedding (#385)
• Support for HTTP proxies (#391)
• M2M100 and NLLB models (#392)
• Multilingual translation serving (Bumblebee.Text.translation/4) (#395)
• Swin model (#394)

Changed

• (Breaking) Renamed :use_qkv_bias spec option to :use_attention_bias in ViT, DeiT and DINOv2
• (Breaking) Changed spec options for optional outputs (:output_hidden_states, :output_attentions) to global layer options (#360)
• Whisper serving to stream input using ffmpeg when given a file path (#361)
• (Breaking) Changed model params to use %Axon.ModelState{}, matching Axon v0.7.0 (#375)

Fixed

• Token classification crashing when all tokens are special
• Batched text generation finishing too early
• Crash on empty list of entities when aggregating entities in token classification (#386)