mirror of
https://github.com/ml-explore/mlx.git
synced 2025-12-16 01:49:05 +08:00
add cuda sm 90
This commit is contained in:
@@ -108,7 +108,7 @@ endif()
|
||||
# Compute capability 7 is required for synchronization between CPU/GPU with
|
||||
# managed memory. TODO: Add more architectures for potential performance gain.
|
||||
set(MLX_CUDA_ARCHITECTURES
|
||||
"70;80"
|
||||
"70;80;90"
|
||||
CACHE STRING "CUDA architectures")
|
||||
message(STATUS "CUDA architectures: ${MLX_CUDA_ARCHITECTURES}")
|
||||
set_target_properties(mlx PROPERTIES CUDA_ARCHITECTURES
|
||||
|
||||
@@ -49,11 +49,7 @@ inline __device__ void atomic_add(__half* out, __half val) {
|
||||
}
|
||||
|
||||
inline __device__ void atomic_add(complex64_t* out, complex64_t val) {
|
||||
#if __CUDA_ARCH__ < 900
|
||||
atomic_add_general(out, val);
|
||||
#else
|
||||
atomicAdd(out, val);
|
||||
#endif
|
||||
}
|
||||
|
||||
inline __device__ void atomic_add(__nv_bfloat16* out, __nv_bfloat16 val) {
|
||||
|
||||
Reference in New Issue
Block a user