Skip to content

Latest commit

 

History

History
13 lines (10 loc) · 605 Bytes

README.md

File metadata and controls

13 lines (10 loc) · 605 Bytes

Kernel releases

只放通过正确性、模型 shape 集和重复性能测试的 HIP/Triton kernels;过程树在 working/kernels/。

Component Coverage
attention/ FA / MLA / PA / SageAttention
gemm-dequant/ GEMM/GEMV、量化与 dequant
moe/ routing、expert dispatch 与 GEMM
norm-softmax/ RMSNorm、Softmax 与融合算子
topk/ TopK / sampling-related kernels

每个主题必须有 README、REPORT 与 benchmark YAML,并更新 kernel dashboard。