fix(gemma2): Skip missing parameters during GGUF weight loading#30421

Closed

kitaekatt wants to merge 7 commits into

vllm-project:mainfrom

kitaekatt:fix/gemma2-gguf-qweight-type

Commits on Dec 15, 2025

fix: Lazy tokenizer init in StructuredOutputManager to prevent semaphore leak
kitaekatt
committed
fix(gguf): Skip lm_head mapping for models with tied word embeddings
kitaekatt
committed
fix(gguf): Ensure Gemma2 configs have hidden_act for backward compatibility
kitaekatt
committed
fix(gguf): Auto-select compatible dtype for GGUF models on Blackwell

kitaekatt
and
claude
committed
fix(nemotron_h): Add missing rotary positional embeddings to attention layers

kitaekatt
and
claude
committed
fix: Use model_config.dtype for rotary embeddings
kitaekatt
committed
fix(gemma2): Skip missing parameters during GGUF weight loading
kitaekatt
committed