vLLM v0.19.1
vLLM v0.19.1 releases as a patch update with a Transformers v5.5.3 upgrade and multiple bug fixes targeting Gemma4 streaming, tool calls, LoRA loading, and quantized MoE support.
This is a patch release on top of `v0.19.0` with Transformers v5.5.3 upgrade and bug fixes for Gemma4: Update to transformers v5 (#30566) [Bugfix] Fix invalid JSON in Gemma 4 streaming tool calls by stripping partial delimiters (#38992) [Bugfix][Frontend] Fix Gemma4 streaming HTML duplication after tool calls (#38909) [Bugfix] Fix Gemma4 streaming tool call corruption for split boolean/number values (#39114) [Tool] adjust_request to reasoning parser, and Gemma4 fixes (#39027) [Gemma4] Support quantized MoE (#39045) Add Gemma4 Eagle3 support (#39450) [Gemma4][Bugfix]: Enable Gemma4ForCasualLM to load lora adapters correctly (#38844) [Bugfix] Fix Gemma4 tool parser converting bare null to string "null" (#39679) [Model] Fix Gemma 4 token repetition by dynamic BOS injection for PT models (#39842) fix(kimi_k25): resolve media_placeholder_token_id from tokenizer (#39344)
- github.comvllm v0.19.1primary