Yuqiang Ge

icon company Cornell University icon location Beijing, China icon location Master in Electrical and Computer Engineering@Cornell University

Results 12 comments of Yuqiang Ge

Remove hard-coded hardware in sdpa

/re-run all-failed

Remove hard-coded hardware in sdpa

/re-run all-failed

Remove hard-coded hardware in sdpa

/re-run all-failed

Remove hard-coded hardware in sdpa

/re-run all-failed

Remove hard-coded hardware in sdpa

/re-run win openblas

Remove hard-coded hardware in sdpa

/re-run win openblas

Remove hard-coded hardware in sdpa

/re-run all-failed

Remove hard-coded hardware in sdpa

/re-run all-failed

Enable AVX in Linux-CPU / Build

/re-run all-failed

【启航计划】PaddlePaddle PHI算子库CUDA Kernel规范化

## 示例修复 以 **`paddle/phi/kernels/legacy/gpu/moe_gate_dispatch_kernel.cu`** 为例。 ### 1. 问题现状 在 **PaddleCustomDevice 仓库** 中可以看到以下两个注册文件直接 `#include` 了 `.cu` 源文件： - `backends/metax_gpu/kernels/ernie_core/moe_gate_dispatch_kernel_register.cu` - `backends/iluvatar_gpu/kernels/ernie_core/moe_gate_dispatch_kernel_register.cc` #### 错误示例 ```cpp #include "paddle/phi/core/kernel_registry.h" #include "paddle/phi/kernels/legacy/gpu/moe_gate_dispatch_kernel.cu" //NOLINT PD_CUSTOM_KERNEL_REGISTER(moe_gate_dispatch, metax_gpu,...