vLLM supported models documentation, re-read for architecture-class and encoder-decoder statements

Re-read 2026-10-03 for the architecture x engine join; complements pap-arch-vllm-supported-models (same URL, earlier read). Four statements that matter and are not in any engine record. (1) Transformers modeling backend scope, verbatim: 'Currently, the Transformers modeling backend works for the following: Modalities: embedding models, language models, vision-language models* and audio-language models; Architectures: encoder-only, decoder-only, mixture-of-experts; Attention types: full attention and/or sliding attention', with '*Vision-language models currently accept only image inputs.' That is the whole fallback envelope: full or sliding attention ONLY, so no MLA, no hybrid SSM/linear-attention, no state-space and no diffusion path can be entered through the fallback. (2) Encoder-decoder is NOT native: 'For other model architectures not natively supported, in particular for Encoder-Decoder models, we recommend following a similar pattern by implementing support through the plugin system', and the only two plugin rows are BartForConditionalGeneration and Florence2ForConditionalGeneration (both via bart-plugin). (3) Diffusion: the only diffusion row in the whole registry is DiffusionGemmaForBlockDiffusion (DiffusionGemma, T + I+ + V+); there is no LLaDA and no Dream row, matching pap-arch-vllm-supported-models. (4) Native rows confirmed for the hybrid/SSM and MLA classes: MambaForCausalLM, Mamba2ForCausalLM, JambaForCausalLM, FalconMambaForCausalLM, NemotronHForCausalLM, GraniteMoeHybridForCausalLM, Qwen3NextForCausalLM, KimiLinearForCausalLM, Lfm2ForCausalLM, Lfm2MoeForCausalLM, OlmoHybridForCausalLM, Olmo3ForCausalLM, GptOssForCausalLM, Glm4MoeForCausalLM, GlmMoeDsaForCausalLM, DeepseekV2/V3/V32/V4, KimiK3ForConditionalGeneration, BailingMoeV3ForCausalLM, Step3p5ForCausalLM, Qwen3_5/Qwen3.5 hybrid rows, Gemma3nForCausalLM, Gemma4ForCausalLM, Llama4ForConditionalGeneration, MiniCPMV, MolmoForCausalLM, Molmo2ForConditionalGeneration, PaliGemmaForConditionalGeneration, SmolVLMForConditionalGeneration, InternVLChatModel, Glm4vForConditionalGeneration, VoxtralForConditionalGeneration, WhisperForConditionalGeneration. Encoder-side (Bert/XLMRoberta/Roberta/ModernBert/NomicBert/Gte/Gemma text) architectures are on the pooling pages, not here, and are covered by embed-vllm-pooling-models-docs. Also documented: Gemma3nForConditionalGeneration 'is only supported on V1 due to shared KV caching and it depends on timm>=1.0.17 to make use of its MobileNet-v5 vision backbone' - a stated hard dependency of one vision tower.