Qwen3.5-35B-A3B

CONTEXT SCALING: null is a VERIFIED value, checked against config rather than left blank. config.json declares rope_scaling null, max_position_embeddings 262144. layer_types over 40 layers = 30 linear_attention + 10 full_attention, full_attention_interval 4. First Qwen generation to abandon all-full attention. Architecture config is now nested (text_config + vision_config) under Qwen3_5MoeForConditionalGeneration. layer_types over 40 layers: 30 linear_attention + 10 full_attention, full_attention_interval 4, linear_conv_kernel_dim 4, linear_num_value_heads 32. This is the Gated DeltaNet / gated-linear-attention pattern, and it is the single most serving-relevant change in the Qwen line. kv_cache_bytes_per_token = 10 full-attention layers x 2 kv_heads x head_dim 256 x 2 x 2 bytes = 10 x 2 x 256 x 2 x 2 = 20,480 B/token (20 KiB). The 30 linear-attention layers hold a fixed-size recurrent state instead of a growing cache, so KV becomes nearly length-independent. If all 40 layers cached like a normal GQA stack this model would need 40 x 2 x 256 x 2 x 2 = 81,920 B/token - 4x the actual figure. Compare with Qwen3-30B-A3B (the previous generation's small MoE): nearly identical KV per token (20 vs 96 KiB... note 20 KiB vs 96 KiB, a 4.8x reduction) at similar active size, and similar total params. MoE: 256 experts, 8 per token, PLUS a shared expert (shared_expert_intermediate_size 512) - the first Qwen MoE to add one, matching DeepSeek's convention. Vendor: 35B total / 3B activated; safetensors total 35.95B. hidden 2048, moe_intermediate_size 512, gqa_ratio 8 (16 heads / 2 kv heads), context 262144, rope_theta 1000000. Apache-2.0. Multimodal (vision_config present). 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=hybrid-attention-ssm. REQUIRED KERNEL CLASS: two kernel families in one graph - a chunked linear-attention / delta-rule kernel for the SSM or linear-attention layers, and an ordinary or MLA decode kernel for the full-attention minority - plus a fixed-size recurrent state allocator. Three independent records say the recurrent branch is the hard part: flop/ssm-recurrent-state-update (the state is a fixed per-sequence tax that never amortises, so it scales with concurrency and NOT with context); flop/ssm-sequential-scan-tensor-core-mismatch (the recurrence is a multiplicative chain that tensor cores cannot express, so it must run as a sequential scan; the published SSD-vs-fused-scan gap is 2-8x for identical mathematics); and llama.cpp's own dispatch table, which puts every hybrid and SSM architecture in its supports_sm_tensor == false list [[mdl2-llamacpp-arch-class-registry]] - the engine agreeing with the arithmetic. That same list also shows llama.cpp needed ONE shared hybrid code path (llm_arch_is_hybrid, 18 entries: JAMBA, FALCON_H1, PLAMO2, GRANITE_HYBRID, LFM2, LFM2MOE, NEMOTRON_H, NEMOTRON_H_MOE, QWEN3NEXT, KIMI_LINEAR, BAILINGMOE3, KIMI_K3, GLM5_NEXT, QWEN35, QWEN35MOE, QWEN4EXP, DEEPSEEK4, MINIMAX_01) covering exactly the families vLLM and SGLang each ship as separate rows, which is why hybrid support arrives per engine release rather than at once. SUPPORTED BY (cited): vLLM native rows MambaForCausalLM, Mamba2ForCausalLM, JambaForCausalLM, FalconMambaForCausalLM, FalconH1ForCausalLM, NemotronHForCausalLM, GraniteMoeHybridForCausalLM, Qwen3NextForCausalLM, KimiLinearForCausalLM, Lfm2ForCausalLM, Lfm2MoeForCausalLM, OlmoHybridForCausalLM, KimiK3ForConditionalGeneration, BailingMoeV3ForCausalLM, Qwen3_5 rows [[mdl2-vllm-supported-models-2026-10]]; SGLang generative table names Granite 4.0 micro as 'hybrid Mamba-MoE ... combining gated short convolutions with a small number of grouped query attention blocks', Kimi Linear as 'hybrid linear attention model ... Kimi Delta Attention (KDA) for up to 6x faster decoding and 75% KV cache reduction vs full attention', Nemotron Nano as 'hybrid Mamba-Transformer architecture', Ling as 'hybrid MoE model: per-layer mix of DeepSeek-style MLA full attention and Qwen3 Gated-Delta-Net (GDN) linear attention', LFM2 as 'hybrid backbone', Mamba-Codestral as 'Mamba2 ... holds no KV cache', and its multimodal page carries the descriptor 'hybrid KDA/MLA' [[mdl2-sglang-generative-models]] [[mdl2-sglang-multimodal-language-models]]. TensorRT-LLM has rows for NemotronHForCausalLM, Qwen3NextForCausalLM, Qwen3_5MoeForCausalLM, Qwen4ExpForCausalLM, KimiLinearForCausalLM and KimiK3ForConditionalGeneration but NO row at all for MambaForCausalLM, Mamba2ForCausalLM, FalconH1ForCausalLM, Lfm2ForCausalLM or GraniteMoeHybridForCausalLM [[mdl2-trtllm-supported-models]]. PER-FAMILY FEATURE TRAPS, all from the TRT-LLM key-models matrix, which is the only per-architecture feature matrix in this repo [[mdl2-trtllm-supported-models]]: Qwen3NextForCausalLM has Sliding Window Attention = No and Speculative Decoding = No; Kimi K3 has no MTP or EAGLE-3 head at all (spec method DSpark only) and needs Blackwell SM100; MiniMax-M3 sparse-attention path supports neither KV cache reuse nor MTP in this release; DeepSeek-V4 and Step-3.7 are 'Untested' for disaggregated serving. KERNEL BUGS THAT ARE ARCHITECTURE-SPECIFIC AND REPRODUCED: (1) vLLM's EAGLE last-block drop (drop_eagle_block) costs 6-10 points of prefix-cache hit rate, and on a hybrid GDN/Mamba layout the scheduler alignment unit is 1,648 tokens instead of a 16-token dense block, so the same protection costs ~100x more cache hit - 97.5% to 86.9-91.7% hit rate and 30-40% batch throughput on prefix-reusing workloads [[eagle-prefix-cache-last-block-drop]]. (2) llama.cpp decode collapses on hybrid gated-deltanet GGUF models at high KV position: 33 t/s at 68K to 1.4 t/s at 91K with the model fully resident, across KV quant types and with FA on or off, reproduced on Windows and WSL2 on master and release [[gated-deltanet-decode-collapse-long-context]]. (3) llama.cpp's SYCL path has a regression specific to hybrid linear-attention graphs - qwen3next/qwen3_35 architecture models crash in ggml_sycl_op_mul_mat or return empty output on 2x Arc Pro B60, bracketed between builds b8477 (works) and b9479 (fails), while dense-attention models on the same cards are unaffected [[sycl-hybrid-linear-attention-arch-crash-intel-arc]]. (4) FlashInfer rejects large head sizes at engine init for the head geometry these hybrids introduce [[flashinfer-rejects-large-head-dim]]. Hybrid Gated DeltaNet with full attention every 4th layer. vLLM and TRT-LLM both have native rows (Qwen3_5 / Qwen3_5Moe), and TRT-LLM's key-models matrix marks Qwen3_5MoeForCausalLM with MTP speculative decoding and KV cache reuse Yes [[mdl2-vllm-supported-models-2026-10]] [[mdl2-trtllm-supported-models]]. The class-level traps apply in full: the eagle-prefix-cache-last-block-drop 1,648-token alignment is measured on exactly this hybrid GDN layout (30-40% batch throughput on prefix-reusing workloads), and llama.cpp's SYCL path has a bisected regression crashing qwen3_35-architecture models on Arc Pro B60 while dense models are unaffected [[eagle-prefix-cache-last-block-drop]] [[sycl-hybrid-linear-attention-arch-crash-intel-arc]]. FlashInfer also rejects large head sizes, and this checkpoint's head_dim is 256 [[flashinfer-rejects-large-head-dim]]. ik-llama-cpp records fused delta-net (Gated Delta Net) as a feature that landed there first and was later upstreamed [[ik-llama-cpp]]. CONTEXT_SCALING_METHOD (mrope): rope_parameters carries mrope_interleaved true with mrope_section [11,11,10] - multimodal interleaved rope, the same family as the record's sibling Qwen3-VL entries. LAYER LAYOUT (qwen3-5-35b-a3b): layer_types read verbatim from config, one entry per layer; attention_layer_indices are the 0-based positions of the full-attention entries.