#30209 — Skip generation config fallback for GGUF to prevent multi-process hang. Loading GGUF models in multi-process mode (V1 engine) caused an indefinite hang — both the EngineCore and APIServer processes tried to memory-map the same GGUF file. Fix skips the fallback entirely since GGUF files embed config in the file header.

#30407 — Add memory barriers for cross-process shared memory visibility. Shared memory broadcast caused data races across process boundaries in multi-process inference. Added ordering guarantees to ensure correct visibility of shared state.

#30408 — Disable bfloat16 for GGUF on Blackwell. GGUF models on Blackwell GPUs (RTX 5090, SM 120+) produced incorrect output because bfloat16 causes precision issues with quantized weights on this architecture. Fix defaults GGUF to float16 on Blackwell with a warning when bfloat16 is explicitly requested.

#30409 — Lazy tokenizer init to prevent GGUF semaphore leak. Repeated GGUF model loading/unloading exhausted system semaphores due to eager tokenizer initialization in StructuredOutputManager. Fix defers tokenizer init until first use.

#30410 — Auto-select compatible dtype for GGUF on Blackwell. Gemma2/Gemma3 GGUF models on Blackwell hit a dtype deadlock: float16 causes numerical instability in Gemma, bfloat16 causes precision issues with GGUF on Blackwell. Fix adds _resolve_dtype_conflict() to auto-select float32 when both are disallowed.

#30412 — Skip lm_head mapping for models with tied word embeddings. GGUF loading failed with RuntimeError: Failed to map GGUF parameters: ['lm_head.weight'] for models like Gemma2 that share weights between input embeddings and output projection. Fix adds lm_head.weight to sideload params when tie_word_embeddings=True.

#30434 — Use EOS token ID from GGUF metadata instead of HF tokenizer. Gemma 3 GGUF models never stopped generating — the model emitted <end_of_turn> (token 106) but vLLM waited for the HF tokenizer's EOS (token 1), resulting in repeated EOS tokens until max_tokens. Fix reads the correct EOS from GGUF metadata.

#30411 — Ensure Gemma2 configs have hidden_act for backward compatibility. Gemma2 GGUF loading hit AttributeError: 'Gemma2Config' has no attribute 'hidden_act' because Transformers uses hidden_activation while vLLM accesses hidden_act directly. Fix copies the value across.

#30413 — Add missing rotary positional embeddings to Nemotron-H attention layers. Nemotron-H models loaded successfully but generated corrupted output because the attention class had no RoPE initialization. Without positional information, attention scores were meaningless. Fix adds full rotary embedding support.

#30421 — Skip missing parameters during GGUF Gemma2 weight loading. GGUF loader yielded qweight_type metadata for all quantized tensors including embeddings, but VocabParallelEmbedding doesn't have those parameters, causing a KeyError. Fix adds a safety check matching the existing pattern in llama.py.

#30423 — Make GGUFMoEMethod.apply() parameters optional. GGUF MoE models (e.g., Qwen3-30B) failed because GGUFMoEMethod.apply() required top_k and renormalize arguments that were never passed by the caller and not used in the method body.

#30424 — Add quant_config to Gemma2 embedding layer for GGUF support. Gemma2 GGUF models loaded successfully but produced garbage output because the embedding layer lacked quant_config, causing F.embedding() to interpret quantized bytes as float values. Same bug previously fixed for DeepSeek in #12836.

#30427 — Extract attn_logit_softcapping from GGUF metadata. Gemma2 GGUF models produced garbage because FlashAttention used softcap=0 (disabled) without this parameter, causing numerical instability. The attention backends already supported softcap — this was a config mapping gap.

#30500 — Extract HF config from GGUF metadata for repos without config.json. GGUF repos like bartowski's caused vLLM to fail at model loading. Fix adds a GGUF config parser that constructs HuggingFace-compatible config from GGUF metadata fields.

#30699 — Skip missing parameters during GGUF Gemma2 weight loading. Targeted resubmission of #30421 fix — adds safety check in Gemma2Model.load_weights() to skip parameters not in params_dict.

#30702 — Handle missing config.json in speculator probe for GGUF models. The speculator probe tried to load config.json before GGUF handling ran, failing at engine init. More targeted fix than #30500, following reviewer feedback that Transformers already handles GGUF config extraction.

#31464 — Apply RMSNorm weight correction for Gemma2 GGUF models. Gemma2 GGUF models produced gibberish because llama.cpp adds 1 to RMSNorm weights during GGUF conversion, but vLLM expects original values. Fix subtracts 1 during loading, matching the correction already applied for Gemma3 in #26189. Tested on RTX 5090: coherent output, 40% MMLU accuracy, 344 tok/s.