zhangyiss/mlx
1
0
Fork 0
mirror of https://github.com/ml-explore/mlx.git synced 2025-10-19 16:48:10 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

Add more CUDA architectures for PyPi package (#2427)

Browse Source 
* add cuda sm 90

* add more archs
This commit is contained in:
Awni Hannun
2025-07-28 12:35:15 -07:00
committed by GitHub
parent ab0e608862
commit 641be9463b
4 changed files with 13 additions and 30 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
mlx/backend/cuda/CMakeLists.txt
View File

@@ -105,11 +105,11 @@ if(CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 12.8.0)
mlx PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:--compress-mode=size>")
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"
CACHE STRING "CUDA architectures")
# Compute capability >= 7.0 is required for synchronization between CPU/GPU with
# managed memory.
if(NOT DEFINED MLX_CUDA_ARCHITECTURES)
set(MLX_CUDA_ARCHITECTURES "native")
endif()
message(STATUS "CUDA architectures: ${MLX_CUDA_ARCHITECTURES}")
set_target_properties(mlx PROPERTIES CUDA_ARCHITECTURES
"${MLX_CUDA_ARCHITECTURES}")

17
mlx/backend/cuda/device.cpp
View File

@@ -97,23 +97,6 @@ CommandEncoder::CaptureContext::~CaptureContext() {
if (discard) {
return;
}
// Extract and add as single kernel node when possible.
size_t num_nodes;
CHECK_CUDA_ERROR(cudaGraphGetNodes(graph, NULL, &num_nodes));
if (num_nodes == 1) {
cudaGraphNode_t captured_node;
CHECK_CUDA_ERROR(cudaGraphGetNodes(graph, &captured_node, &num_nodes));
cudaGraphNodeType type;
CHECK_CUDA_ERROR(cudaGraphNodeGetType(captured_node, &type));
if (type == cudaGraphNodeTypeKernel) {
CUDA_KERNEL_NODE_PARAMS params;
CHECK_CUDA_ERROR(cuGraphKernelNodeGetParams(captured_node, &params));
enc.add_kernel_node(params);
return;
}
}
// Otherwise add the captured graph as subgraph.
enc.add_graph_node(graph);
}

4
mlx/backend/cuda/device/atomic_ops.cuh
View File

@@ -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) {