Gemma 2 9B

PARAMS: 9.24B is the safetensors count (9B nominal). KV: 42 * 8 * 256 * 2 * 2 bytes = 344,064 bytes/token = 336 KiB/token, which is 2.6x the Llama 3.3 70B figure because head_dim is 256 rather than 128 and there are 42 rather than 80 layers. Gemma's wide-head GQA is expensive per token of context. SLIDING WINDOW: sliding_window=4096 on alternating local/global layers. Real resident KV at long context is therefore about half the naive figure, since local layers only keep 4096 tokens. LOGIT SOFTCAPPING is the numerically important part and it is NOT an attention change: attn_logit_softcapping=50.0 clamps the pre-softmax attention logits, and final_logit_softcapping=30.0 clamps the output logits, with query_pre_attn_scalar=256 scaling queries before the clamp. This changes the numerics of every attention and every sampling step, so it is not a drop-in for a plain Llama-shaped kernel, and it is why the model needed explicit stability work to train at 8k. SOURCE CAVEAT: unsloth mirror of Google weights. 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=sliding-window. REQUIRED KERNEL CLASS: a local/windowed attention decode kernel plus per-layer-type cache management. The kernel requirement is not cosmetic: flop/sliding-window-hybrid-attention records that a local layer's decode cost is constant in context (it reads exactly w KV entries) so its arithmetic intensity stays at ~1 flop/byte forever, while the global layers' bytes still grow linearly - the two layer classes have opposite scaling and one kernel must serve both. SUPPORTED BY (cited): vLLM lists sliding attention inside its fallback envelope - 'Attention types: full attention and/or sliding attention' - but ALSO has native rows for the specific SWA architectures: GptOssForCausalLM, Gemma3ForCausalLM, Gemma4ForCausalLM, Olmo3ForCausalLM, GraniteMoeSWAForCausalLM, GraniteSWAForCausalLM [[mdl2-vllm-supported-models-2026-10]]. SGLang's generative table names GPT-OSS as 'hybrid sliding-window-attention MoE model ... with per-layer-type RoPE (YARN on full-attention layers, default on sliding-attention layers) and a softplus per-head attention gate', and lists granite-swash SWA models [[mdl2-sglang-generative-models]]. llama.cpp exposes SWA as first-class KV keys rather than a flag: %%s.attention.sliding_window and %%s.attention.sliding_window_pattern exist as distinct keys per architecture class, alongside SWA-specific width keys, which is direct evidence that SWA changes the cache layout rather than adding an attention flag [[mdl2-llamacpp-arch-class-registry]]. THE PER-CELL CONSEQUENCE: an engine WITHOUT local-attention support does not fail, it silently allocates a full-context cache for every layer. docs/13-model-compatibility.md tabulates this as the KV overpay table; for this model the naive figure in kv_cache_bytes_per_token is the worst case and the effective figure is what the windowed layers actually hold. TWO HARD GATES THAT ARE NOT ABOUT WINDOW SIZE. (1) TensorRT-LLM publishes a per-architecture 'Sliding Window Attention' column and it is NOT uniformly Yes: Yes for DeepseekV4ForCausalLM, Gemma4ForConditionalGeneration, Gemma4UnifiedForConditionalGeneration and Step3p7ForConditionalGeneration, and explicitly No for Qwen3NextForCausalLM, while the DeepSeek-V3/V3.2/GLM/GPT-OSS/Llama-4/Qwen3.5/Kimi-K3/Nemotron-H rows read N/A [[mdl2-trtllm-supported-models]]. (2) On B200 with vLLM 0.23.1rc1.dev843 + flashinfer-python 0.6.14, a 50-of-60-layer SWA Gemma-4 checkpoint with --kv-cache-dtype fp8_e4m3 and MTP spec-decode (k=7) produces fluent output containing sporadic wrong tokens and short repetition loops once seq_len passes the 1024-token window, because flashinfer's trtllm spec-as-decode kernel mis-computes sliding-window attention for BF16-Q + FP8-KV at certain seq_len/page alignments and vLLM's scale bookkeeping double-applies k/v scales - gotcha-flashinfer-bf16q-fp8kv-spec-decode-corrupts-swa, severity BLOCKER. FP8-Q + spec is clean; BF16-Q + no spec is clean; either drop quantized KV, drop spec decode, or drop the disable_flashinfer_q_quantization flag. This is the cleanest statement in the repo that quantized KV and speculative decoding are not independently safe on an SWA model. (3) FlashInfer also rejects large head sizes at engine init, which is a separate SM120 gap from the page-block one: head_size not supported for gemma-4-31b-it's head_dim 256 / global_head_dim 512 on RTX PRO 6000 [[flashinfer-rejects-large-head-dim]]. (4) On Hopper, vLLM 0.30.0's vllm_flash_attn silently ignores num_splits on SM90, costing 48% per-token decode at batch 1 for Gemma-4 26B-A4B-it global layers (1 query token, 2 KV heads, 2 of an H100's 132 SMs with work), with TRITON_ATTN recovering ~70% [[fa4-num-splits-ignored-sm90]]. (5) vLLM's own scheduler granularity interacts with SWA groups: issue #53786 asks for fine-grained prefix hits for sliding-window groups, and hybrid reconciliation already caps the combined hit at the shorter boundary [[eagle-prefix-cache-last-block-drop]]. (6) On Apple Silicon the class is a cache-CLASS problem, not a kernel problem: mlx-lm's sliding-window layers use RotatingKVCache, which had no quantized implementation, so --kv-bits passes a hasattr presence check and raises NotImplementedError on the first request instead of at startup; the fix (PR 1584) adds RotatingQuantizedKVCache, and the commonly-recommended monkeypatch silently gives up the max_kv_size window bound and yields an unbounded cache [[mlx-kv-bits-crashes-on-first-request-for-hybrid-attention]] [[mdl2-mlxlm-pr1584-rotating-quantized-kv]]. KV OVERPAY: naive 42*8,192 = 344,064 B/token; effective with local support = (7*8192*c + 35*8192*4096)/c = 57,344 + 1,174,405,120/c - 1.71x at its 8,192 max context for the same reason as Gemma 2 2B. CONTEXT_SCALING_METHOD (native): rope_scaling null at theta 1e4 over 8192; the alternating sliding/global pattern is attention geometry, not context extension. LAYER LAYOUT (gemma-2-9b): layer_types read verbatim from config, one entry per layer; attention_layer_indices are the 0-based positions of the full-attention entries.