Unverified50% confidenceFactExact time
DeepEP对应NCCL，负责多卡之间的通信；FlashMLA对应FlashAttention，用于超长文本处理；TileKernels对应CUTLASS；DeepSelect对应DALI这类数据管线
1
Sources
50%
Confidence
Long-term
Relevance
10/4/2026
First Seen
Sources
Related Entities
Related Claims
Unverified六个组件分工：DeepGEMM加速通用矩阵运算，DeepEP负责大规模跨设备通信，TileKernels提供常规向量计算和访存算子，FlashMLA提供稀疏注意力算子，DeepSelect做高效数据筛选78% similarUnverifiedDeepSeek 4.1 Flash 采用编码器-解码器结构，编码器创建共享的全局记忆，解码器从中读取信息72% similarUnverifiedDeepSeek发布了DeepSeek-V4.1-Flash，该方法对KV缓存所需内存进行了大幅压缩69% similarUnverified百度的 DeepSpeech 系列采用 RNN-CTC 架构69% similarUnverifiedDeepSeek开源了面向华为昇腾的算子和通信组件，与此前的GPU版本一一对应68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/973402API
curl https://kongchang.com/api/v1/knowledge/claims/973402MCP
get_claim(id=973402)