SmolLM3-3B

CONTEXT SCALING: null is a VERIFIED value, checked against config rather than left blank. config.json declares rope_scaling null, rope_theta 5000000, max_position_embeddings 65536. Trained with the smoltalk2 mixture (HuggingFaceTB's 2025 SFT dataset); config declares layer_types all full_attention, so no sliding-window or hybrid mix. The unusual serving-relevant detail is positional: no_rope_layer_interval 4 with no_rope_layers showing 1 on three of every four layers, i.e. 27 of 36 layers carry no rotary embedding at all (NoPE) and only 9 are RoPE'd. Kernel selection cannot assume uniform RoPE across layers. rope_theta 5000000, no rope_scaling, 65536 native context. tie_word_embeddings true at vocab 128256 keeps the parameter count low for the context length. head_dim derived as 2048 / 16 = 128; gqa_ratio = 16 / 4 = 4.KV bytes/token at bf16 (2 bytes/element): 36 layers x 4 KV heads x 128 head_dim x 2 (K and V) x 2 bytes = 73728 B/token = 72.0 KiB/token. At 64k context that is 4.50 GiB of KV per sequence at bf16 - still 1.5x the model's own 6.15 GB bf16 weight footprint. 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=dense-gqa. REQUIRED KERNEL CLASS: ordinary causal attention plus GEMM/GEMV. This is the baseline every engine in this repo implements, so the join here is a floor rather than a filter: no engine record in data/engines/ states that dense GQA is unsupported. SUPPORTED BY (cited): vLLM LlamaForCausalLM / Qwen2 / Qwen3 / Mistral / Phi3ForCausalLM rows [[mdl2-vllm-supported-models-2026-10]]; TensorRT-LLM LlamaForCausalLM, Qwen2ForCausalLM, Qwen3ForCausalLM, Phi3ForCausalLM rows [[mdl2-trtllm-supported-models]]; SGLang generative table (Llama, Qwen, Mistral families) [[mdl2-sglang-generative-models]]; llama.cpp LLM_ARCH_LLAMA [[pap-arch-llamacpp-arch-registry]]. HARD CONSTRAINTS THAT DO APPLY even to dense GQA, and they are model-shape rather than architecture-class: tensor-parallel size must divide the attention head count or the engine refuses to start [[mgpu-tp-size-must-divide-attention-heads]] (affects vllm and sglang); and a GQA KV cache is the thing paged-attention machinery exists for, so an engine with no paged KV (llama.cpp's slot-based pool [[llama-cpp]], exllamav2 [[exllamav2]]) loses the concurrency lever rather than the ability to run. ONE HARDWARE-DEPENDENT KERNEL GATE THAT IS NOT ABOUT GQA: on RDNA, vLLM's custom paged-attention kernel is gated on head_size == 128 AND block_size == 16, so a dense GQA model with head_dim 256 falls to the Triton 2D fallback and decode falls off a cliff - measured 12.1 tok/s at 518 tokens of context to 4.2 tok/s at 32k, with kernel_paged_attention_2d growing 28.3x [[rdna-triton-paged-attn-decode-cliff]]. Check head_dim before blaming the card. CONTEXT_SCALING_METHOD (native): rope_scaling null at theta 5e6 over 65536; NoPE on every 4th layer, no extension.