[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-80bc5e25-d24d-45a1-9c2b-534ebcae39f9":3},{"id":4,"title":5,"summary":6,"original_url":7,"source_id":8,"tags":9,"published_at":23,"created_at":24,"modified_at":25,"is_published":26,"publish_type":27,"image_url":13,"view_count":28},"80bc5e25-d24d-45a1-9c2b-534ebcae39f9","腾讯 WeDLM 开源：让扩散 LLM 在标准因果注意力下跑出 3-6× vLLM 加速","扩散语言模型（DLLM）的并行生成潜力一直被「双向注意力破坏 KV 缓存」这堵工程墙挡在门外：理论上吞吐再高，落到生产里跑不过 vLLM 这种被 FlashAttention、PagedAttention、CUDA Graphs 三件套武装到牙齿的自回归推理引擎。腾讯 WeChat AI 团队这次给出的答案 WeDLM 选择了不寻常的路径：保留标准因果注意力，用「Topological Reordering」把已观测 token 搬到物理前缀而保留其逻辑位置，从而在严格 causal mask 下做并行解码。\n\n效果立竿见影：WeDLM-8B-Instruct 基于 Qwen3-8B 微调，在 GSM8K、MATH 等数学推理任务上对 vLLM 优化版 Qwen3-8B 取得 3-6× 加速；顺序生成、计数题这类低熵场景最高能到 10×。7B\u002F8B 两档模型、推理引擎 wedlm、Docker 镜像和评测脚本一同开源，Apache 2.0 协议。\n\n更关键的是部署侧零摩擦：FlashAttention、PagedAttention、CUDA Graphs 一条不漏地复用，Prefix Cache 天然兼容，工程团队不用重写内核就能接入。这与同期 LLaDA、Dream-7B 等仍坚持双向注意力的路线形成鲜明对比——它们在论文里赢了 benchmark，部署时却要为 KV 缓存重新设计。\n\nDLLM 阵营过去一年「理论上赢、工程上输」的尴尬，根源是沉醉于双向注意力的理论优雅而牺牲了部署效率。WeDLM 反其道而行：把因果 mask 当作工程资产而非设计缺陷，是务实的取舍。这条路径如果被更多团队跟进，扩散 LLM 进入生产环境的门槛会显著降低，而 KV 缓存兼容性会成为评估生成范式的新基线。","https:\u002F\u002Fgithub.com\u002Ftencent\u002FWeDLM","d46ec0a7-501b-4ef8-9c89-2391b2701b3b",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"7b67033c-19e6-4052-a626-e681bba64c7a","diffusion",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":18,"name":19,"slug":19,"description":13,"color":13},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"id":21,"name":22,"slug":22,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm","2026-06-16T20:00:00Z","2026-06-16T20:10:17.565861Z","2026-06-16T20:10:17.565869Z",true,"agent",4]