Skip to content

Launch fast.cross_entropy in slices past the Metal grid limit #4776

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
zcbenz merged 1 commit into from
Oct 31, 2026
+23 −21
Merged

Launch fast.cross_entropy in slices past the Metal grid limit #4776

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
44 changes: 23 additions & 21 deletions mlx/backend/metal/cross_entropy.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13,20 +13,16 @@ namespace mlx::core::fast {
namespace {

// Threadgroup shrinks with the axis so a small vocabulary needs one iteration.
inline std::pair<MTL::Size, MTL::Size> cross_entropy_dims(
inline size_t cross_entropy_threadgroup_size(
MTL::ComputePipelineState* kernel,
int axis_size,
int n_rows) {
int axis_size) {
constexpr size_t simd_size = 32;
size_t tg_max =
(kernel->maxTotalThreadsPerThreadgroup() / simd_size) * simd_size;
size_t needed =
(axis_size + CROSS_ENTROPY_N_READS - 1) / CROSS_ENTROPY_N_READS;
size_t threadgroup_size = std::clamp(
return std::clamp(
((needed + simd_size - 1) / simd_size) * simd_size, simd_size, tg_max);
return {
MTL::Size(n_rows * threadgroup_size, 1, 1),
MTL::Size(threadgroup_size, 1, 1)};
}

array ensure_row_contiguous(
Expand Down Expand Up @@ -66,18 +62,21 @@ void CrossEntropy::eval_gpu(
out.set_data(allocator::malloc(out.nbytes()));

int axis_size = in.shape().back();
int n_rows = in.data_size() / axis_size;
int64_t n_rows = in.data_size() / axis_size;

std::string kernel_name = "cross_entropy_" + type_to_name(in);
auto kernel = get_cross_entropy_kernel(d, kernel_name, in);
auto [grid_dims, group_dims] = cross_entropy_dims(kernel, axis_size, n_rows);
size_t threadgroup_size = cross_entropy_threadgroup_size(kernel, axis_size);

compute_encoder.set_compute_pipeline_state(kernel);
compute_encoder.set_input_array(in, 0);
compute_encoder.set_input_array(targets, 1);
compute_encoder.set_output_array(out, 2);
compute_encoder.set_bytes(axis_size, 3);
compute_encoder.dispatch_threads(grid_dims, group_dims);
const int64_t row_bytes = int64_t(axis_size) * in.itemsize();
dispatch_threadgroups_in_slices(
compute_encoder, n_rows, threadgroup_size, [&](int64_t row) {
compute_encoder.set_input_array(in, 0, row * row_bytes);
compute_encoder.set_input_array(targets, 1, row * targets.itemsize());
compute_encoder.set_output_array(out, 2, row * out.itemsize());
});
}

void CrossEntropyVJP::eval_gpu(
Expand Down Expand Up @@ -109,20 +108,23 @@ void CrossEntropyVJP::eval_gpu(
}

int axis_size = in.shape().back();
int n_rows = in.data_size() / axis_size;
int64_t n_rows = in.data_size() / axis_size;

std::string kernel_name = "cross_entropy_vjp_" + type_to_name(in);
auto kernel = get_cross_entropy_kernel(d, kernel_name, in);
auto [grid_dims, group_dims] = cross_entropy_dims(kernel, axis_size, n_rows);
size_t threadgroup_size = cross_entropy_threadgroup_size(kernel, axis_size);

compute_encoder.set_compute_pipeline_state(kernel);
compute_encoder.set_input_array(in, 0);
compute_encoder.set_input_array(targets, 1);
compute_encoder.set_input_array(loss, 2);
compute_encoder.set_input_array(cotan, 3);
compute_encoder.set_output_array(out, 4);
compute_encoder.set_bytes(axis_size, 5);
compute_encoder.dispatch_threads(grid_dims, group_dims);
const int64_t row_bytes = int64_t(axis_size) * in.itemsize();
dispatch_threadgroups_in_slices(
compute_encoder, n_rows, threadgroup_size, [&](int64_t row) {
compute_encoder.set_input_array(in, 0, row * row_bytes);
compute_encoder.set_input_array(targets, 1, row * targets.itemsize());
compute_encoder.set_input_array(loss, 2, row * loss.itemsize());
compute_encoder.set_input_array(cotan, 3, row * cotan.itemsize());
compute_encoder.set_output_array(out, 4, row * row_bytes);
});
}

} // namespace mlx::core::fast
Loading