SGLang supported models: embedding, rerank, classify and reward pages

Read 2026-10-03 (embedding_models.md, rerank_models.md, classify_models.md, reward_models.md). Embedding page: 'Native encoder embedding architectures and google/embeddinggemma-300m are detected automatically. Decoder-style embedding models require --is-embedding'; on CUDA SGLang 'automatically uses breakable CUDA graph (BCG) for its full encoder prefill and DISABLES incompatible radix-cache and chunked-prefill behavior'. Multimodal embedding needs a chat template (--chat-template gme-qwen2-vl for gme-Qwen2-VL). Rerank page states the two-class taxonomy explicitly: 'Cross-encoder rerank models: run with --is-embedding (embedding runner)' vs 'Decoder-only rerank models: run WITHOUT --is-embedding and use next-token logprob scoring (yes/no)', text-only (Qwen3-Reranker) and multimodal (Qwen3-VL-Reranker). BGE-reranker-v2-m3 row carries a HARD KERNEL CONSTRAINT: 'Currently only support attention-backend triton and torch_native', and its example launch command passes --attention-backend triton, --disable-radix-cache and --chunked-prefill-size -1. No ColBERT / late-interaction / token_embed endpoint appears on any of the four pages. Classify page example is jason9693/Qwen2.5-1.5B-apeach; reward page lists internlm2-7b-reward, Qwen2.5-Math-RM-72B and two Skywork reward models.