zhangyiss/mlx
1
0
Fork 0
mirror of https://github.com/ml-explore/mlx.git synced 2025-12-16 01:49:05 +08:00
Code Issues Actions 1 Packages Projects Releases Wiki Activity

[CUDA] Switch to CUDA graphs (#2317)

Browse Source 
* cuda graph prototype

fix signal bug + start to add dependencies

capture more

capture more ops

remaining ops

fix reduce and rope deps

add concurrent context

try update, but not working

cosistent topology order

use node api

use node api directly to reduce overhead

fix bug

use kernels in unary

cache graph

format

fix synchronization

format

* comment
This commit is contained in:
Awni Hannun
2025-07-02 15:59:13 -07:00
committed by GitHub
parent e76e9b87f0
commit ec0d5db67b
36 changed files with 1461 additions and 1212 deletions
Download Patch File Download Diff File Expand all files Collapse all files

36
mlx/backend/cuda/reduce/col_reduce.cu
View File

@@ -214,26 +214,24 @@ void col_reduce_looped(
encoder.set_input_array(in);
encoder.set_output_array(out);
encoder.launch_kernel([&](cudaStream_t stream) {
dispatch_all_types(in.dtype(), [&](auto type_tag) {
dispatch_reduce_ops(reduce_type, [&](auto reduce_type_tag) {
dispatch_reduce_ndim(args.reduce_ndim, [&](auto reduce_ndim) {
using OP = MLX_GET_TYPE(reduce_type_tag);
using T = cuda_type_t<MLX_GET_TYPE(type_tag)>;
using U = typename cu::ReduceResult<OP, T>::type;
dispatch_all_types(in.dtype(), [&](auto type_tag) {
dispatch_reduce_ops(reduce_type, [&](auto reduce_type_tag) {
dispatch_reduce_ndim(args.reduce_ndim, [&](auto reduce_ndim) {
using OP = MLX_GET_TYPE(reduce_type_tag);
using T = cuda_type_t<MLX_GET_TYPE(type_tag)>;
using U = typename cu::ReduceResult<OP, T>::type;
// Cub doesn't like const pointers for vectorized loads. (sigh)
T* indata = const_cast<T*>(in.data<T>());
// Cub doesn't like const pointers for vectorized loads. (sigh)
T* indata = const_cast<T*>(in.data<T>());
constexpr int N_READS = 4;
constexpr int BM = 32;
constexpr int BN = 32;
dim3 grid = output_grid_for_col_reduce(out, args, BN);
int blocks = BM * BN / N_READS;
auto kernel =
cu::col_reduce_looped<T, U, OP, reduce_ndim(), BM, BN, N_READS>;
kernel<<<grid, blocks, 0, stream>>>(indata, out.data<U>(), args);
});
constexpr int N_READS = 4;
constexpr int BM = 32;
constexpr int BN = 32;
dim3 grid = output_grid_for_col_reduce(out, args, BN);
int blocks = BM * BN / N_READS;
auto kernel =
cu::col_reduce_looped<T, U, OP, reduce_ndim(), BM, BN, N_READS>;
encoder.add_kernel_node(
kernel, grid, blocks, indata, out.data<U>(), args);
});
});
});