deepseek-ai/FlashMLA
FlashMLA: Efficient Multi-head Latent Attention Kernels
C++ · ★ 12,423 · Fork 973 · 在 GitHub 查看
在手机上浏览源码、订阅每日 AI 趋势日报与项目架构深度解读:获取 CopoHub 客户端(iOS & HarmonyOS)→
GitHub Trending 上榜记录
deepseek-ai/FlashMLA 已在 GitHub Trending 每日榜单上榜 3 天,最高排名第 5,首次上榜 2026-01-22,最近一次上榜 2026-01-24。
| 日期 | 排名 | Star | 当日增长 |
|---|---|---|---|
| 2026-01-24 | #9 | 12,359 | +184 |
| 2026-01-23 | #9 | 12,307 | +184 |
| 2026-01-22 | #5 | 12,114 | +25 |
同期上榜的项目
- remotion-dev/remotion – 🎥 Make videos programmatically with React
- VectifyAI/PageIndex – 📑 PageIndex: Document Index for Vectorless, Reasoning-based RAG
- OpenBMB/UltraRAG – UltraRAG v3: A Low-Code MCP Framework for Building Complex and Innovative RAG Pipelines
- browser-use/browser-use – Open-source Python library for AI browser agents, plus Browser Use Cloud: stealth browsers and a hosted agent API.
- ai-dynamo/dynamo – A Datacenter Scale Distributed Inference Serving Framework
- block/goose – an open source, extensible AI agent that goes beyond code suggestions - install, execute, edit, and test with any LLM
FlashMLA 项目深度剖析与商业评估报告
FlashMLA 项目深度剖析与商业评估报告
报告日期:2026-08-22 分析对象:deepseek-ai/FlashMLA v(基于 2026-07-28 提交) 许可证:MIT License 语言构成:C++(CUDA 内核主导) 社区热度:12.9k Stars / 1.1k Forks / 122 Open Issues
---
1. 核心定位与商业落地价值
1.1 根本痛点:MLA 推理的“内存墙”与“算力墙”双重瓶颈
FlashMLA 解决的核心问题是 Multi-head Latent Attention(MLA)在长上下文推理场景下的双重瓶颈:
内存墙:MLA 通过低秩压缩将 KV Cache 从 [seqlen, nheads, headdim] 压缩为 [seqlen, dlatent](DeepSeek-V3 中 dlatent=576),但即便如此,在 128K 上下文下,单请求的 KV Cache 仍可达数十 GB。FlashMLA 通过 FP8 量化 + Paged KV Cache + 稀疏注意力 三层手段,将单 token 的 KV 存储压缩至 656 Bytes(对比 BF16 密集存储的 1,152 Bytes,压缩率约 43%)。 算力墙:MLA 的 decode 阶段是典型的 memory-bound 操作(每 token 仅做一次矩阵乘),而 prefill 阶段是 compute-bound。FlashMLA 通过 Split-K 调度 + Warp 级并行 在两种模式下均逼近硬件峰值。
1.2 商业落地场景
在 README 中展示上榜徽章
[](https://github.fq6825.top/repo/deepseek-ai/FlashMLA)