Skip to content

Fix precision in Metal fused attention #3119

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
awni merged 1 commit into from
Feb 10, 2026
+8 −13
Merged

Fix precision in Metal fused attention #3119

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
21 changes: 8 additions & 13 deletions mlx/backend/metal/kernels/steel/attn/kernels/steel_attention.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15,16 +15,6 @@ constant bool has_mask [[function_constant(300)]];
constant bool do_causal [[function_constant(301)]];
constant bool has_sinks [[function_constant(302)]];

template <typename T>
struct TransformScale {
T scale;
METAL_FUNC TransformScale(T scale_) : scale(scale_) {}

METAL_FUNC T apply(T x) const {
return scale * x;
}
};

struct MaxOp {
template <typename T>
METAL_FUNC static constexpr T apply(T x, T y) {
Expand Down Expand Up @@ -173,7 +163,7 @@ template <
VBlockLoader loader_v(
V, params->V_strides[2], Vs, simd_group_id, simd_lane_id);

TransformScale<T> ts(static_cast<T>(params->scale * M_LOG2E_F));
const AccumType scale = params->scale * M_LOG2E_F;

// Prepare MMA tiles
constexpr short kFragSize = 8; // MMAFrag size
Expand Down Expand Up @@ -216,13 +206,12 @@ template <

threadgroup_barrier(mem_flags::mem_threadgroup);

// Load Q blocks apply scale
// Load Q blocks
if (!align_Q && int(tid.x) == (params->NQ_aligned)) {
loader_q.load_safe(short2(BD, params->qL_rem));
} else {
loader_q.load_unsafe();
}
loader_q.apply_inplace_op(ts);

// Init row reduction variables
constexpr short kRowsPT = decltype(Stile)::kRowsPerThread;
Expand Down Expand Up @@ -281,6 +270,12 @@ template <
tile_matmad(Stile, Qtile, Ktile, Stile);
}

// Apply scale in float32
STEEL_PRAGMA_UNROLL
for (short ii = 0; ii < decltype(Stile)::kElemsPerTile; ii++) {
Stile.elems()[ii] *= scale;
}

// Mask out length sequence
if (!align_K && kb == (params->NK_aligned)) {
using stile_t = decltype(Stile);
Expand Down
Loading