Skip to content

Commit 790be8a

Browse files
authored
Fix CUDA all-reduce planning for large inputs (ml-explore#3603)
1 parent 2e6632e commit 790be8a

1 file changed

Lines changed: 6 additions & 6 deletions

File tree

‎mlx/backend/cuda/reduce/all_reduce.cu‎

Lines changed: 6 additions & 6 deletions
Original file line numberDiff line numberDiff line change
@@ -68,12 +68,12 @@ void all_reduce(
6868

6969
out.set_data(cu::malloc_async(out.nbytes(), encoder));
7070

71-
auto get_args = [](int size, int N) {
72-
int threads = std::min(512, (size + N - 1) / N);
71+
auto get_args = [](size_t size, size_t N) {
72+
int threads =
73+
static_cast<int>(std::min(size_t{512}, cuda::ceil_div(size, N)));
7374
threads = ((threads + WARP_SIZE - 1) / WARP_SIZE) * WARP_SIZE;
74-
int reductions_per_step = threads * N;
75-
size_t steps_needed =
76-
(size + reductions_per_step - 1) / reductions_per_step;
75+
size_t reductions_per_step = threads * N;
76+
size_t steps_needed = cuda::ceil_div(size, reductions_per_step);
7777

7878
int blocks;
7979
if (steps_needed < 32) {
@@ -88,7 +88,7 @@ void all_reduce(
8888
blocks = 1024;
8989
}
9090

91-
size_t steps_per_block = (steps_needed + blocks - 1) / blocks;
91+
size_t steps_per_block = cuda::ceil_div(steps_needed, blocks);
9292
size_t block_step = steps_per_block * reductions_per_step;
9393

9494
return std::make_tuple(blocks, threads, block_step);

0 commit comments

Comments
 (0)