Qwen3-8B

CONTEXT SCALING: null is a VERIFIED value here, checked against config rather than left blank. config.json (Qwen/Qwen3-8B) declares rope_scaling null, rope_theta 1000000, max_position_embeddings 40960. Verified null - Qwen3 ships 40960 as a native length, it is not extrapolated. Dense GQA, the baseline serving case - no MoE, no MLA, no hybrid attention. kv_cache_bytes_per_token = 36 layers x 8 kv_heads x head_dim 128 x 2 (K and V) x 2 bytes (bf16) = 36 x 8 x 128 x 2 x 2 = 147,456 B/token (144 KiB). head_dim 128 is explicit in config (and equals 4096/32). gqa_ratio 4. Params: safetensors total 8.19B; our arithmetic from config (hidden 4096, 36 layers, intermediate 12288, vocab 151936) gives 8.2B, matching. Context: max_position_embeddings 40960, rope_theta 1000000, rope_scaling null (native, no YaRN). Apache-2.0, which makes this the licence-easiest model in the slice - Qwen's older 2.5 line used the bespoke Qwen License Agreement, so 3.x is a real improvement for on-prem use. Thinking vs non-thinking is a serving mode (enable_think in the chat template), not a different architecture; the -2507 refresh is the same shape. Qwen3's hybrid thinking mode has no KV implication. 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.