团队亲手删掉自研 Megakernel:Torch Compile 加 CUDA Graphs 达其 98% 性能
simran_s_arora · x · 2026-11-17
dylanlim's team reflects on abandoning whole-model megakernels: after months of work last year they stabilized only one model on one batch. Default Torch Compile with CUDA Graphs reaches 98% of the hand-written kernel's performance, and compiler-based optimization is 28% faster than megakernels on B200. The post contrasts the two ML optimization paradigms and argues whole-model megakernels overcomplicate things and are not an ideal optimization target.
所属事件:团队放弃自研 Megakernel,Torch Compile 方案达 98% 性能(2 条相关)→
「Infra」频道最新
- 腾讯 DSpine 推测解码上线:27B 模型提速至 4.5 倍、90-125 tok/s — pmttyji · 2026-11-17
- 直言算力不缺:千卡 GPU 全靠短期合同入手,六成是 B300 — SumitGup · 2026-11-17
- 开源项目 Agent Substrate 让空闲 Agent 挂起唤醒,同硬件跑 10 倍 Agent — Saboo_Shubham_ · 2026-11-17
- OpenAI 代理记忆系统省钱三招:缓存 TTL、Flex 减半、少调用 — gabrielchua · 2026-11-17
- xAI 再扩孟菲斯算力版图:关联公司 550 万美元购入 96 英亩土地 — XFreeze · 2026-11-17
- 蚂蚁 Ling 团队详解 290B 模型低成本训练:修损失函数省大量算力 — aftahi_ai · 2026-11-17