SmolLM2-1.7B-Instruct

CONTEXT SCALING: null is a VERIFIED value, checked against config rather than left blank. config.json declares rope_scaling null, rope_theta 130000, max_position_embeddings 8192. Included mainly to show the regression SmolLM3 fixed: SmolLM2 is model_type 'llama' with plain MHA (32 query = 32 KV heads) and head_dim only 64. head_dim derived = 2048 / 32 = 64. kv_cache_bytes_per_token = 24 * 32 * 64 * 2 * 2 = 196608 bytes/token at bf16. Read that next to SmolLM3-3B's 73728 bytes/token: SmolLM3 is 1.8x the parameters but uses 2.7x LESS KV per token, because halving head_dim (128 -> 64) while moving to 4 KV heads cuts per-layer KV by 4x. At 8k context SmolLM2-1.7B holds 1.5 GiB of KV — a large fraction of its own 1.7B weights at bf16, and the single biggest reason 8k-context serving of this model is memory-bound rather than compute-bound. tie_word_embeddings true, vocab 49152, intermediate 8192, rope_theta 130000. Official GGUF (HuggingFaceTB/SmolLM2-1.7B-Instruct-GGUF) and ONNX ports exist; Apache-2.0. 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-mha. 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. MHA SPECIFIC: gqa_ratio is 1.0, so the KV cache is num_layers x num_kv_heads x head_dim x 2 x 2 with num_kv_heads == num_attention_heads. There is no engine-side special case: MHA is the shape every paged-attention kernel was written for, so the join is strictly easier than for GQA and the only cost is the KV bytes themselves. Tensor parallelism is unchanged - the constraint is on the QUERY head count, so MHA with many heads is the most permissive case for head divisibility [[mgpu-tp-size-must-divide-attention-heads]].