From 96c2ebc58ba15edd3f75d8ac7e2daab766f0c630 Mon Sep 17 00:00:00 2001 From: zijiexia <37504505+zijiexia@users.noreply.github.com> Date: Tue, 14 Jul 2026 00:04:47 -0700 Subject: [PATCH] [docs] Note the default dsa-topk-backend on all DSA-model cookbook pages (#31124) Co-authored-by: Claude Fable 5 --- .../cookbook/autoregressive/DeepSeek/DeepSeek-Math-V2.mdx | 4 ++++ docs_new/cookbook/autoregressive/DeepSeek/DeepSeek-V3_2.mdx | 4 ++++ docs_new/cookbook/autoregressive/GLM/GLM-5.1.mdx | 4 ++++ docs_new/cookbook/autoregressive/GLM/GLM-5.2.mdx | 4 ++++ docs_new/cookbook/autoregressive/GLM/GLM-5.mdx | 4 ++++ docs_new/cookbook/autoregressive/Meituan/LongCat-2.0.mdx | 4 ++++ 6 files changed, 24 insertions(+) diff --git a/docs_new/cookbook/autoregressive/DeepSeek/DeepSeek-Math-V2.mdx b/docs_new/cookbook/autoregressive/DeepSeek/DeepSeek-Math-V2.mdx index bbb08460f..539f0d3e6 100644 --- a/docs_new/cookbook/autoregressive/DeepSeek/DeepSeek-Math-V2.mdx +++ b/docs_new/cookbook/autoregressive/DeepSeek/DeepSeek-Math-V2.mdx @@ -38,6 +38,10 @@ This section provides deployment configurations optimized for different hardware + + DeepSeek-Math-V2 is built on DeepSeek-V3.2 and uses DSA sparse attention. All recipes here run the DSA indexer top-k on the default `--dsa-topk-backend sgl-kernel`. Other top-k backend choices have not been fully validated on this model. + + ### 3.2 Configuration Tips **Hardware Requirements:** diff --git a/docs_new/cookbook/autoregressive/DeepSeek/DeepSeek-V3_2.mdx b/docs_new/cookbook/autoregressive/DeepSeek/DeepSeek-V3_2.mdx index cba3154e7..5d821666b 100644 --- a/docs_new/cookbook/autoregressive/DeepSeek/DeepSeek-V3_2.mdx +++ b/docs_new/cookbook/autoregressive/DeepSeek/DeepSeek-V3_2.mdx @@ -57,6 +57,10 @@ import { DeepSeekV32Deployment } from "/src/snippets/autoregressive/deepseek-v32 + + All recipes here run the DSA indexer top-k on the default `--dsa-topk-backend sgl-kernel`. Other top-k backend choices have not been fully validated on DeepSeek-V3.2. + + ### 3.2 Configuration Tips - **Short-sequence MHA prefill (adaptive):** For prefill sequences shorter than 2048 tokens (default threshold), the DSA backend automatically switches to standard MHA (using FlashAttention variable-length on SM90, TRT-LLM ragged MHA on SM100). To extend this to longer sequences set env var `SGLANG_DSA_PREFILL_DENSE_ATTN_KV_LEN_THRESHOLD` to a larger value (potential minor accuracy trade-off). diff --git a/docs_new/cookbook/autoregressive/GLM/GLM-5.1.mdx b/docs_new/cookbook/autoregressive/GLM/GLM-5.1.mdx index 2d1824912..39445e63e 100644 --- a/docs_new/cookbook/autoregressive/GLM/GLM-5.1.mdx +++ b/docs_new/cookbook/autoregressive/GLM/GLM-5.1.mdx @@ -30,6 +30,10 @@ import { GLM51Deployment } from '/src/snippets/autoregressive/glm-51-deployment. + + All recipes here run the DSA indexer top-k on the default `--dsa-topk-backend sgl-kernel`. Other top-k backend choices have not been fully validated on GLM-5.1. + + ### 3.2 Configuration Tips - Speculative decoding (MTP) can significantly reduce latency for interactive use cases. diff --git a/docs_new/cookbook/autoregressive/GLM/GLM-5.2.mdx b/docs_new/cookbook/autoregressive/GLM/GLM-5.2.mdx index f21ec1094..acf599e11 100644 --- a/docs_new/cookbook/autoregressive/GLM/GLM-5.2.mdx +++ b/docs_new/cookbook/autoregressive/GLM/GLM-5.2.mdx @@ -52,6 +52,10 @@ import { benchmarks } from "/src/snippets/configs/zai-org/glm-5.2-benchmarks.jsx + + All recipes here run the DSA indexer top-k on the default `--dsa-topk-backend sgl-kernel`. Other top-k backend choices have not been fully validated on GLM-5.2. + + Speed numbers are measured with `--random-range-ratio 1.0`, `--flush-cache`, on `main @ 09ca4fc`. Spec cells pin the EAGLE acceptance length via the serve env `SGLANG_SIMULATE_ACC_LEN` (low-latency 5-1-6 = 3.5, balanced 2-1-3 = 2); high-throughput has no spec. diff --git a/docs_new/cookbook/autoregressive/GLM/GLM-5.mdx b/docs_new/cookbook/autoregressive/GLM/GLM-5.mdx index 444770c7d..3f89e593d 100644 --- a/docs_new/cookbook/autoregressive/GLM/GLM-5.mdx +++ b/docs_new/cookbook/autoregressive/GLM/GLM-5.mdx @@ -41,6 +41,10 @@ import { GLM5Deployment } from '/src/snippets/autoregressive/glm-5-deployment.js + + All recipes here run the DSA indexer top-k on the default `--dsa-topk-backend sgl-kernel`. Other top-k backend choices have not been fully validated on GLM-5. + + ### 3.2 Configuration Tips - Speculative decoding (MTP) can significantly reduce latency for interactive use cases. diff --git a/docs_new/cookbook/autoregressive/Meituan/LongCat-2.0.mdx b/docs_new/cookbook/autoregressive/Meituan/LongCat-2.0.mdx index d1b8fd5a0..b3db0f884 100644 --- a/docs_new/cookbook/autoregressive/Meituan/LongCat-2.0.mdx +++ b/docs_new/cookbook/autoregressive/Meituan/LongCat-2.0.mdx @@ -59,6 +59,10 @@ import { benchmarks } from "/src/snippets/configs/meituan-longcat/longcat-2.0-be + + All recipes here run the LongCat sparse-attention indexer top-k on the default `--dsa-topk-backend sgl-kernel`. Other top-k backend choices have not been fully validated on LongCat-2.0. + + The B300 single-node recipe was validated end-to-end with CUDA graph capture enabled. H200, B200, and H20 are shown as 2-node recipes because LongCat-2.0-FP8 needs 16 ranks for those GPU memory profiles.