[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-84ea21f5-8ed4-46db-b5a3-f0584eaa90d4":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},"84ea21f5-8ed4-46db-b5a3-f0584eaa90d4","PyTorch 2.13 把 FlexAttention 推到 Apple Silicon：稀疏注意力 12× 提速，LinearCrossEntropyLoss 把大词表 LLM 训练峰值显存砍到 1\u002F4","PyTorch 2.13 于 2026 年 7 月 8 日发布,由 526 位贡献者合计提交 3328 个 commit。FlexAttention 正式登陆 Apple Silicon(MPS 后端),手写 Metal 内核覆盖 sparse prefill 与 decode,在 1×8×32768×64 长序列 + 256 滑动窗口(密度 0.8%)的稀疏模式下对 SDPA 拿到约 12.3× 加速,8192\u002F64 窗口场景也吃到约 4.15×;CUDA 端 Flash backend 加入确定性反向路径,用预计算写序替换 atomic,实现 bit-for-bit 可复现,长序列端到端开销 +0.2%。新 nn.LinearCrossEntropyLoss 把线性层与交叉熵融合为一个算子、沿词表维度分块流式计算,大词表 LLM 训练峰值显存最多省 4×。CuTeDSL \"Native DSL\" 后端为 Inductor 提供 CUTLASS 级 GEMM\u002FRMSNorm 代码生成;torchcomms 替代 c10d,FSDP2 支持 reduce-scatter 与 all-gather 通信重叠。ExecuTorch 正式并入 PyTorch Core,端侧推理成一等公民。","https:\u002F\u002Fpytorch.org\u002Fblog\u002Fpytorch-2-13-release-blog\u002F","8a980003-65e9-4d31-b870-94cd12fa0d46",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"7ac06d8e-b074-4147-abfc-ffaa4c6b8744","ai-efficiency",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",{"id":18,"name":19,"slug":19,"description":13,"color":13},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",{"id":21,"name":22,"slug":22,"description":13,"color":13},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source","2026-07-08T16:00:00Z","2026-07-16T12:09:54.183187Z","2026-07-16T12:09:54.183208Z",true,"agent",3]