Qwen3-235B-A22B

CONTEXT SCALING: null is a VERIFIED value, checked against config rather than left blank. config.json (Qwen/Qwen3-235B-A22B) declares rope_scaling null, rope_theta 1000000, max_position_embeddings 40960. Verified native. Vendor: 235B total / 22B activated; safetensors total 235.09B; our config arithmetic gives 235.0B / 22.1B, both matching the vendor figures to ~0.5%. kv_cache_bytes_per_token = 94 layers x 4 kv_heads x head_dim 128 x 2 x 2 = 94 x 4 x 128 x 2 x 2 = 192,512 B/token (188 KiB). GQA, not MLA. That 188 KiB is the headline number in this record: DeepSeek-V3 does 37B active in 68.6 KiB/token while this does 22B active in 188 KiB/token - 2.8x the KV for 60% of the compute. On a long-context workload the MoE advantage nearly disappears. MoE: 128 experts, 8 per token, no shared expert. hidden 4096, intermediate 12288, moe_intermediate_size 1536, all 94 layers MoE (no first_k_dense_replace), gqa_ratio 16 (64 heads / 4 kv heads). Context 40960 with rope_theta 1000000, no rope_scaling. Apache-2.0. Weights are ~470 GB in bf16, so this is a multi-node or heavily quantized model. 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.