[{"data":1,"prerenderedAt":52},["Reactive",2],{"news-story-news-1792384474854-fl9ix":3},{"id":4,"title_zh":5,"title_en":6,"summary_zh":7,"summary_en":8,"takeaways_zh":9,"takeaways_en":15,"content_zh":21,"content_en":22,"external_links":23,"source":30,"category":37,"impact_score":38,"views_count":39,"likes_count":40,"related_entities":41,"tags":45,"image_url":51},"news-1792384474854-fl9ix","SGLang v0.5.22 重磅发布：DeepSeek-V4.1 在 H200 吞吐提升 2.3 倍、MI355X 提速 1.9 倍，InstantTensor 载入提速 10 倍","SGLang v0.5.22 Released: DeepSeek-V4.1 Throughput Jumps 2.3x on H200 and 1.9x on MI355X, InstantTensor Loads Models 10x Faster","SGLang 团队正式发布 v0.5.22 大版本更新（凝聚 245 位贡献者提交的 950 个 PR）：深度优化 DeepSeek-V4.1 推理性能，在 H200 单卡吞吐狂飙 2.3 倍，并通过 DSpark 首次在 AMD MI350X/MI355X 实现 1.9 倍吞吐提升；首创 InstantTensor 加载器将大模型冷启动显存载入提速 10 倍；B200 上 Qwen3.8-Flash-Next 解码提速 12%、B300 上 GLM-5.3-Flash 提速 10% 且 MiMo-V2.6-Pro 解码吞吐达 1.85 倍；并原生支持苹果 Apple Silicon 与 ComfyUI 扩散生态。","The SGLang team released v0.5.22 featuring 950 PRs from 245 contributors: DeepSeek-V4.1 delivers up to 2.3x more throughput per GPU on NVIDIA H200 and 1.9x on AMD MI350X/MI355X via DSpark. The new InstantTensor loader cuts model loading time by 10x. B200 decodes Qwen3.8-Flash-Next 12% faster, while B300 speeds up GLM-5.3-Flash by 10% and yields 1.85x decode throughput for MiMo-V2.6-Pro. SGLang also adds native Apple Silicon PyTorch support and ComfyUI integration.",[10,11,12,13,14],"SGLang v0.5.22 正式发布，汇聚 245 位贡献者的 950 个 PR","DeepSeek-V4.1 在 NVIDIA H200 单卡吞吐暴涨 2.3 倍，并通过 DSpark 首次在 AMD MI350X/MI355X 实现 1.9 倍吞吐提升","全新 InstantTensor 加载器将千亿模型冷启动载入速度提升 10 倍（10x faster）","B200 上 Qwen3.8-Flash-Next 解码提速 12%，B300 上 GLM-5.3 提速 10% 且 MiMo-V2.6-Pro 解码达 1.85 倍","PD 分离架构使长 Prompt 首字延迟 p95 TTFT 减半（-50%），吞吐净增 27%，原生支持 Apple Silicon 与 ComfyUI",[16,17,18,19,20],"SGLang v0.5.22 released with 950 PRs from 245 contributors across heterogeneous platforms","DeepSeek-V4.1 achieves up to 2.3x throughput per GPU on NVIDIA H200 and 1.9x on AMD MI350X/MI355X via DSpark","New InstantTensor loader speeds up large model checkpoint ingestion by 10x","Qwen3.8-Flash-Next decodes 12% faster on B200; MiMo-V2.6-Pro reaches 1.85x decode throughput on B300","PD disaggregation cuts p95 TTFT in half (-50%) on long prompts, alongside native Apple Silicon and ComfyUI diffusion support","### 核心背景与行业痛点\n随着 DeepSeek-V4.1、Qwen3.8 与 GLM-5.3 等超大规模稀疏 MoE 模型的快速普及，企业私有化部署与公有云服务商在算力利用率上遭遇了多重严峻挑战：第一，千亿级参数模型在英伟达 H200 及 AMD MI350X/MI355X 等新一代异构加速卡上的显存带宽利用率与多步投机推演效率受制于算子通信瓶颈；第二，超大模型冷启动加载常耗时数分钟，制约了弹性扩缩容；第三，长文本长 Prompt 场景下的首字延迟（TTFT）与解码延迟随并发急剧恶化。开源推理服务框架 **SGLang** 正式发布了 **`v0.5.22`** 核心里程碑版本，给出了覆盖全硬件生态的系统级性能跃升。\n\n### 架构亮点与底层机制\n1. **DeepSeek-V4.1 与异构算力深度融合加速（PR #41251, #41308）**：\n   - 针对 DeepSeek-V4.1 引入全新多层显存复用与通信隐藏算子，在 NVIDIA H200 上实现单卡吞吐飙升 2.3 倍；\n   - 通过与 AMD 合作引入 `DSpark` 算子库，首次在 AMD Instinct MI350X / MI355X 旗舰集群上打通端到端低延迟服务，吞吐提升 1.9 倍；\n2. **InstantTensor 极速加载器（PR #40453）**：\n   - 彻底重构权重反序列化与页锁定内存传输通道，利用零拷贝直接映射消除多重内存拷贝与格式转换，实现模型权重载入速度提升 10 倍（10x faster）；\n3. **PD 分离架构与投机解码升级（PR #40703, #43735）**：\n   - Prefill-Decode（PD）分离架构在长 Prompt 工作负载下将吞吐提升 27%，并将 p95 TTFT 延迟砍半（降低 50%）；GLM-5.2 与 GLM-5.3 首次支持解码上下文并行与 MTP 联合运行，GLM-5.3 解码加速高达 2.1 倍；\n4. **全平台拓扑扩展**：\n   - 首次原生支持苹果 Apple Silicon（M 系列芯片）标准 PyTorch 后端；ComfyUI 原生支持 SGLang Diffusion（支持 MiniMax-H3、Qwen-Image 2.1、GGUF 权重与 LoRA）。\n\n### 权威 Benchmark 与实测跑分对比\n根据 SGLang 官方在 NVIDIA H200、B200、B300 及 AMD MI355X 上的全量 Benchmark 实测：\n- **DeepSeek-V4.1 吞吐**：NVIDIA H200 单卡吞吐达到原版的 **2.3 倍**；AMD MI355X 借助 DSpark 实现 **1.9 倍** 吞吐飞跃；\n- **B200 / B300 顶级芯片解码**：\n  - B200 上 `Qwen3.8-Flash-Next` 解码速度提升 **12%**；\n  - B300 上 `GLM-5.3-Flash` 配合 DFlash 解码速度提升 **10%**；\n  - B300 上 `MiMo-V2.6-Pro` 配合 DFlash 解码吞吐达到 **1.85 倍**；\n- **长文本首字延迟**：PD 分离使长 Prompt 场景的 **p95 TTFT 降低 50%**，吞吐净增 **+27%**；\n- **全新模型 Day-0 支持**：全面上线 Clef/Clef Flash、Perplexity PPLX-Decider v1/v1.1、Google EmbeddingGemma 2、LLaDA2.2、混元 HunyuanImage-3.0-Instruct 与 SANA-Video 2.0。\n\n### 开发者实战落地与开箱指南\n- **快速安装与更新**：\n  ```bash\n  pip install --upgrade \"sglang[all]>=0.5.22\"\n  ```\n- **生产部署建议**：高并发生产集群建议开启 `--enable-instant-tensor` 实现秒级容器冷启动，并在长文档 Agent 检索场景中配置 `--enable-pd-disaggregation`，以极低显存代价将首字延迟压缩一半。","### 核心背景与行业痛点\nAs next-generation MoE foundation models like DeepSeek-V4.1, Qwen3.8, and GLM-5.3 scale into production, enterprise serving stacks face severe memory bandwidth bottlenecks and slow initialization times across heterogeneous hardware clusters (NVIDIA H200/B200/B300 and AMD MI355X). Furthermore, long-context prompt prefilling severely degrades interactive decoding latencies without architectural isolation. The open-source serving engine **SGLang** released **`v0.5.22`**, featuring 950 PRs from 245 contributors that systematically overhaul serving efficiency across the stack.\n\n### 架构亮点与底层机制\n1. **DeepSeek-V4.1 Kernel and DSpark Acceleration (PR #41251, #41308)**:\n   - Delivers up to 2.3x higher throughput per GPU on NVIDIA H200 through optimized memory reuse kernels, and up to 1.9x throughput on AMD Instinct MI350X/MI355X via DSpark integration;\n2. **InstantTensor 10x Loader (PR #40453)**:\n   - Introduces zero-copy pinned memory weight ingestion, slashing large model load times by 10x for rapid container autoscale;\n3. **PD Disaggregation & Speculative Decoding (PR #40703, #43735)**:\n   - Prefill-Decode disaggregation lifts long-prompt throughput by 27% and cuts p95 TTFT by 50%. GLM-5.3 combines decode context parallelism with MTP for up to 2.1x faster token generation;\n4. **Broad Platform Support**:\n   - Supports Apple Silicon Macs natively via standard PyTorch alongside MLX. Adds ComfyUI diffusion integration for MiniMax-H3 and Qwen-Image 2.1 with GGUF and LoRA support.\n\n### 权威 Benchmark 与实测跑分对比\nOfficial benchmarks across NVIDIA H200/B200/B300 and AMD MI355X confirm:\n- **DeepSeek-V4.1 Throughput**: **2.3x per GPU on H200**; **1.9x throughput on AMD MI355X** via DSpark;\n- **Next-Gen Accelerators**: `Qwen3.8-Flash-Next` decodes **12% faster on B200**; `GLM-5.3-Flash` runs **10% faster on B300**, and `MiMo-V2.6-Pro` achieves **1.85x decode throughput with DFlash on B300**;\n- **Interactive Latencies**: PD disaggregation slashes **p95 TTFT by 50%** while increasing throughput by **+27%** on long prompts;\n- **Model Coverage**: Day-0 serving for Clef, PPLX-Decider v1.1, EmbeddingGemma 2, LLaDA2.2, HunyuanImage-3.0-Instruct, and SANA-Video 2.0.\n\n### 开发者实战落地与开箱指南\n- **Upgrade Command**:\n  ```bash\n  pip install --upgrade \"sglang[all]>=0.5.22\"\n  ```\n- **Production Tip**: Enable `--enable-instant-tensor` to slash Kubernetes pod restart times by 90%, and deploy `--enable-pd-disaggregation` on long-context agent pipelines to cut p95 response latencies in half.",[24,27],{"title":25,"url":25,"type":26},"https://github.com/sgl-project/sglang/releases/tag/v0.5.22","github",{"title":28,"url":28,"type":29},"https://docs.sglang.io/","docs",{"platform":31,"author_name":32,"author_handle":33,"author_avatar":34,"tweet_url":35,"published_at":36},"twitter","SGLang Project","@huggingface","https://unavatar.io/x/huggingface","","2026-10-19T04:34:34.853Z","Inference",94,50,0,{"models":42,"agents":43,"vendors":44},[],[],[],[46,47,48,49,50,37],"SGLang","DeepSeek-V4.1","AMD MI355X","B200","B300","https://pbs.twimg.com/media/HU6sLh-bQAAMRVN.jpg",1792890704270]