Skip to content

Respect MLX_ENABLE_TF32 in gated_delta_update #4792

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
Nov 3, 2026
Merged

Respect MLX_ENABLE_TF32 in gated_delta_update #4792

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
22 changes: 16 additions & 6 deletions mlx/backend/metal/gated_delta_update.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15,14 +15,24 @@ namespace mlx::core::fast {

namespace {

inline int gated_delta_chunk_size(int T) {
bool nax = metal::is_nax_available();
inline int gated_delta_chunk_size(Dtype dtype, bool vectorized_g, int T) {
if (vectorized_g) {
// Only sequential kernel supports vectorized g.
return 1;
}
bool nax =
metal::is_nax_available() && (env::enable_tf32() || dtype != float32);
int C = config::get("MLX_GATED_DELTA_CHUNK", T > 8 ? (nax ? 16 : 8) : 1);
return (C == 16 && !nax) ? 8 : C;
}

inline int gated_delta_chunk_size_vjp() {
bool nax = metal::is_nax_available();
inline int gated_delta_chunk_size_vjp(Dtype dtype, bool vectorized_g) {
if (vectorized_g) {
// Only sequential kernel supports vectorized g.
return 1;
}
bool nax =
metal::is_nax_available() && (env::enable_tf32() || dtype != float32);
return nax ? config::get("MLX_GATED_DELTA_CHUNK_VJP", 16) : 1;
}

Expand Down Expand Up @@ -106,7 +116,7 @@ void GatedDeltaUpdate::eval_gpu(
int Dv = v.shape(3);

bool vectorized_g = (g.ndim() == 4);
int C = vectorized_g ? 1 : gated_delta_chunk_size(T);
int C = gated_delta_chunk_size(q.dtype(), vectorized_g, T);
const int ckpt = gated_delta_ckpt(C);

std::string suffix =
Expand Down Expand Up @@ -262,7 +272,7 @@ void GatedDeltaUpdateVJP::eval_gpu(

// 16 = chunked NAX, anything else = sequential.
bool vectorized_g = (g.ndim() == 4);
int C = vectorized_g ? 1 : gated_delta_chunk_size_vjp();
int C = gated_delta_chunk_size_vjp(q.dtype(), vectorized_g);
const bool chunked = (C == 16);

const int n_chunks = (T + 15) / 16;
Expand Down
Loading