InternVL3-8B

TWO-PART MODEL. (1) Vision: InternViT (vision_config._name_or_path InternViT-6B-448px-V1-5), 24 layers, hidden 1024, 16 heads, patch 14, image_size 448. The vision_config still carries the training-time MoE scaffolding (num_experts 8, num_routed_experts 4, num_shared_experts 4, use_moe FALSE) -- the released encoder is dense, the MoE fields are inert. (2) LLM: Qwen2-7B, 28 layers, hidden 3584, 28 heads / 4 KV, head_dim 128 (k_proj [512,3584] confirmed). ARCHITECTURE = dense (both parts dense). PROJECTOR: MLP + PIXEL-SHUFFLE. downsample_ratio 0.5 means a 2x2 spatial downsample before the MLP; ps_version 'v2'. Each 448x448 tile = (448/14)^2 = 1,024 ViT patches, and pixel-shuffle 0.5 quarters that to 256 LLM tokens per tile. This is the key efficiency move: 4x token reduction for a small projector cost, unlike InternVL2's naive per-patch MLP. DYNAMIC TILING (decides prefill cost): dynamic_image_size true, force_image_size 448, min_dynamic_patch 1, max_dynamic_patch 12, use_thumbnail true. So one image becomes between 2 and 13 ViT forward passes: up to 12 tiles of 448x448 plus a 448x448 thumbnail. VISION TOKENS: 256 per tile, so 512 tokens for a single-tile image, up to 13*256 = 3,328 tokens at the 12-tile maximum. DOMINANCE: VISION DOMINATES decisively. At the default high-res setting (12 tiles) 3,328 vision tokens swamp a ~100-token text prompt by 30:1, and the ViT runs 13 separate attention passes. Vision encoder FLOPs, not LLM prefill, are the first thing to profile on InternVL3. KV CACHE: 28 * 4 * 128 * 2 * 2 = 57,344 B/token. params_b 7.944 = HF API safetensors total (7,944,373,760). SERVICEABILITY JOIN (added 2026-10-03; the architecture x engine matrix this record was missing, with a citation per cell in docs/13-model-compatibility.md): FAMILY=multimodal. REQUIRED KERNEL CLASS: an encoder tower (ViT / SigLIP / ConvNet / audio encoder) running on a SEPARATE schedule from the decoder, plus a projector into the decoder's embedding space, plus - for the mrope and DeepStack families - a multi-axis rotary scheme in the decoder rather than plain 1-D RoPE. Three engine-side consequences, all cited. (1) The encoder tower is a separate compute path, so its throughput is a different problem from decode throughput and engines that disclaim encoder throughput are disclaiming this. vLLM says so in its own words: 'We currently support pooling models primarily for convenience. This is not guaranteed to provide any performance improvements over using Hugging Face Transformers or Sentence Transformers directly' [[vllm-pooling-models]] - that is the general statement this class inherits. TRT-LLM's answer is a documented mechanism: a Multimodal Encoder Side Stream that prefetches encoder work on a separate CUDA stream plus a per-model Multimodal Embeddings Cache, currently available only for seven architectures (Gemma4, Gemma4Unified, Mistral3, Qwen3VL, Qwen3VLMoe, Qwen3_5, Qwen3_5Moe) [[mdl2-trtllm-supported-models]]. (2) Several towers carry hard external dependencies: Gemma3nForConditionalGeneration 'depends on timm>=1.0.17 to make use of its MobileNet-v5 vision backbone' and 'is only supported on V1 due to shared KV caching' [[mdl2-vllm-supported-models-2026-10]]; MiniCPM-V 4.6 requires transformers>=5.7.0 because it was upstreamed as a native model type [[mdl2-trtllm-supported-models]]. (3) Hybrid-only multimodal models expose a memory knob unique to this class: vLLM's --language-model-only flag sets all supported multimodal modalities to 0 so the multimodal modules are not loaded, 'to free up more GPU memory for KV cache' for Llama-4, Step3, Mistral-3 and Qwen-3.5 [[mdl2-vllm-supported-models-2026-10]].SUPPORTED BY (cited): vLLM native rows include InternVLChatModel and InternVLForConditionalGeneration, LlavaOnevisionForConditionalGeneration, LlavaNextForConditionalGeneration, MolmoForCausalLM and Molmo2ForConditionalGeneration, PaliGemmaForConditionalGeneration, SmolVLMForConditionalGeneration, Glm4vForConditionalGeneration, Qwen2VL/Qwen3VL, MiniCPMV, Mllama, Phi4MM, Gemma4, NemotronH_Nano_VL_V2, VoxtralForConditionalGeneration (which requires mistral-common[audio]) and Qwen3OmniMoeThinkerForConditionalGeneration [[mdl2-vllm-supported-models-2026-10]]. SGLang's multimodal table covers InternLM2-VL, LLaVA-OneVision, LLaVA-NeXT(-Video), Ling-3.0-flash-VL, LFM2.5-VL, MiniCPM-V-2, GLM-4.5V, GLM-OCR, Qwen3-VL 235B-A22B and 30B-A3B, Qwen3-Omni-30B-A3B, Kimi-VL-A3B, MiMo-VL-7B-RL, Step3-VL-10B, dots.ocr, PaddleOCR-VL, Cosmos3, deepseek-vl2, Janus-Pro-7B, NVILA-8B and gemma-3-4b-it, plus openai/whisper-large-v3 and Qwen3-ASR-1.7B on the audio route [[mdl2-sglang-multimodal-language-models]]. TRT-LLM's multimodal matrix covers a narrower set with an explicit per-row modality string and per-row Untested cells [[mdl2-trtllm-supported-models]]. Older GGUF-side engines reach the class through llama.cpp's projector graph, which llama.cpp-embedding-server states it cannot serve on the embedding route [[llama-cpp-embedding-server]]. PER-MODALITY CAVEATS CITED: TRT-LLM marks Gemma4Unified and MiniCPM-V 4.6 Untested for CUDA Graph / chunked prefill, MiniMax-M3 Untested for attention DP and KV cache reuse, Qwen4Exp Untested across the board, and notes audio requires a checkpoint with a sound_config and is supported only on the full non-disaggregated path [[mdl2-trtllm-supported-models]]. Step-3.7's vision tower runs in BF16 even when the text decoder is quantized [[mdl2-trtllm-supported-models]] - a precision asymmetry that silently costs memory. vLLM's fallback path accepts image inputs only, with video 'to be added in a future release' [[mdl2-vllm-supported-models-2026-10]]. CONTEXT_SCALING_METHOD (native): config nests the LM under a language_model sub-config; no rope_scaling declared there.