TensorRT-LLM supported models and feature matrices (PyTorch backend)

Read at main 2026-10-03. THE ONLY ENGINE REGISTRY IN THIS REPO THAT PUBLISHES A PER-FEATURE MATRIX BY ARCHITECTURE, which makes it the strongest available evidence for the architecture x engine join. Rows confirmed in the PyTorch-backend architecture table: DeepseekV3ForCausalLM (DeepSeek-V3, Kimi-K2), DeepseekV32ForCausalLM, DeepseekV4ForCausalLM (footnote: Blackwell SM100+ only), GlmMoeDsaForCausalLM (GLM-5), Glm4MoeForCausalLM, GptOssForCausalLM, KimiK3ForConditionalGeneration and KimiLinearForCausalLM (footnote: Blackwell SM100 family only), NemotronHForCausalLM, Qwen3NextForCausalLM, Qwen3_5MoeForCausalLM, Qwen4ExpForCausalLM, MiniCPMV4_6ForConditionalGeneration, Step3p7ForConditionalGeneration, Llama4ForConditionalGeneration, WhisperForConditionalGeneration, MllamaForConditionalGeneration, T5ForConditionalGeneration, BertForSequenceClassification, Qwen3ForTextEmbedding, Qwen2For{Reward,ProcessReward}Model, Mistral3ForConditionalGeneration (in the multimodal matrix), Phi4MMForCausalLM, Qwen2VL/Qwen2_5_VL/Qwen3VL/Qwen3VLMoe, LlavaNextForConditionalGeneration. ABSENT from the architecture table entirely: MambaForCausalLM, Mamba2ForCausalLM, FalconH1ForCausalLM, FalconMambaForCausalLM, JambaForCausalLM, Lfm2ForCausalLM, Lfm2MoeForCausalLM, GraniteMoeHybridForCausalLM, Olmo3ForCausalLM, OlmoHybridForCausalLM, any diffusion/ masked-diffusion text architecture (LLaDA, Dream, SDAR), RWKV, Molmo, PaliGemma, InternVL, Voxtral. 'Sliding Window Attention' column in the key-models feature matrix reads: Yes for DeepseekV4ForCausalLM, Gemma4ForConditionalGeneration, Gemma4UnifiedForConditionalGeneration and Step3p7ForConditionalGeneration; No for Qwen3NextForCausalLM; N/A for every DeepSeek-V3/V3.2/GLM/GPT-OSS/Llama-4/Qwen3.5/Kimi-K3/Nemotron-H row. 'KV Cache Reuse' reads No for both Gemma 4 rows and both Step-3.7 rows. Encoder-decoder section states: use the TRTLLM attention backend for encoder-decoder models, tensor parallelism requires attention head counts divisible by TP size, and chunked prefill is NOT supported for the encoder phase so the complete encoder input must fit in the iteration token budget.