vLLM supported models documentation

vLLM's official list of natively implemented model architectures, read 2026-10-03. Confirmed native rows: MambaForCausalLM, Mamba2ForCausalLM, JambaForCausalLM, FalconMambaForCausalLM, NemotronHForCausalLM, GraniteMoeHybridForCausalLM, Qwen3NextForCausalLM, KimiLinearForCausalLM, Lfm2ForCausalLM, Lfm2MoeForCausalLM, OlmoeForCausalLM, Qwen2MoeForCausalLM, GraniteMoeForCausalLM, and DiffusionGemmaForBlockDiffusion. No native LLaDA or Dream row. The docs also define four testing levels, of which 'Strict Consistency' (greedy-decode match vs HuggingFace) is the strongest.