Llama 4 Scout (17B active / 109B total)

PARAMS: 108.6B is the safetensors weight count; Meta's card says 109B total. Not contested, rounding. ACTIVE: 17B is Meta's published figure (and is in the model name). KV: 48 * 8 * 128 * 2 * 2 bytes = 196,608 bytes/token = 192 KiB/token. The 10,485,760 max_position_embeddings in config.json is a nominal training/rope-range figure, NOT a usable serving context; Meta's own table lists 10M context but real deployments land at 10M for Scout only with heavy KV offload. Treat the advertised context as aspirational. Note experts_per_token = 1 with 16 experts: this is a very low top-1 routing ratio, which makes expert batching during decode awkward (each token touches a different expert), unlike Mixtral's 2-of-8 or gpt-oss's 4-of-128. SOURCE CAVEAT: config values from the unsloth mirror; the 17B/109B and 10M-context figures are Meta's own card. 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 (llama3): SCHEME RECLASSIFIED (2026-10-04): was 'llama3 rope scaling (factor 16 from an 8192 base)'. config confirms factor 16.0 over original_max_position_embeddings 8192 - a 16x extrapolation of an 8K base reaching 10,485,760 declared positions, NOT native.