Skip to content

fast: add fused_qsdpa primitive (decode-shape QSDPA on quantized KV) #1

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
benjamin-levin merged 5 commits into from
May 19, 2026
Merged

fast: add fused_qsdpa primitive (decode-shape QSDPA on quantized KV) #1

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
a5f4672
Handle non-multiple-of-8 spatial dims in depthwise conv2d Metal path …
qflen May 17, 2026
ff381c3
Remove reference to groups in Conv3D extra_repr (#3559)
GriffinMB May 18, 2026
f831bdf
Add buffer caching to no_gpu CPU allocator (#3554)
dhiltgen May 18, 2026
2414e5d
Fix steel GEMM safe load offset (#3560)
sofinvalery May 18, 2026
202d4ce
fast: add fused_qsdpa primitive with optional left_padding
May 18, 2026
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
5 changes: 3 additions & 2 deletions mlx/backend/metal/conv.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -962,7 +962,9 @@ void depthwise_conv_2D_gpu(

MTL::Size group_dims = MTL::Size(tc, tw, th);
MTL::Size grid_dims = MTL::Size(
conv_params.C / tc, conv_params.oS[1] / tw, (conv_params.oS[0] / th) * N);
conv_params.C / tc,
(conv_params.oS[1] + tw - 1) / tw,
((conv_params.oS[0] + th - 1) / th) * N);

compute_encoder.dispatch_threadgroups(grid_dims, group_dims);
}
Expand All @@ -986,7 +988,6 @@ void dispatch_conv_2D_gpu(
if (C_per_group == 1 && O_per_group == 1 && is_kdil_one &&
conv_params.wS[0] <= 7 && conv_params.wS[1] <= 7 &&
conv_params.str[0] <= 2 && conv_params.str[1] <= 2 &&
conv_params.oS[0] % 8 == 0 && conv_params.oS[1] % 8 == 0 &&
conv_params.wt_strides[1] == conv_params.wS[1] &&
conv_params.C % 16 == 0 && conv_params.C == conv_params.O) {
return depthwise_conv_2D_gpu(s, d, in, wt, out, conv_params);
Expand Down
1 change: 1 addition & 0 deletions mlx/backend/metal/kernels/CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -54,6 +54,7 @@ build_kernel(random)
build_kernel(rms_norm)
build_kernel(rope)
build_kernel(scaled_dot_product_attention sdpa_vector.h)
build_kernel(fused_qsdpa fused_qsdpa.h)
if(MLX_METAL_VERSION GREATER_EQUAL 320)
build_kernel(fence)
endif()
Expand Down
6 changes: 5 additions & 1 deletion mlx/backend/metal/kernels/conv.metal
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -206,7 +206,7 @@ template <typename T>

threadgroup T ins[TGH * TGW * TGC];

const int n_tgblocks_h = params.oS[0] / th;
const int n_tgblocks_h = (params.oS[0] + th - 1) / th;
const int n = tid.z / n_tgblocks_h;
const int tghid = tid.z % n_tgblocks_h;
const int oh = tghid * th + lid.z;
Expand Down Expand Up @@ -277,6 +277,10 @@ template <typename T>
}
threadgroup_barrier(mem_flags::mem_none);

if (oh >= params.oS[0] || ow >= params.oS[1]) {
return;
}

out += n * params.out_strides[0] + oh * params.out_strides[1] +
ow * params.out_strides[2];
out[c] = static_cast<T>(o);
Expand Down
Loading
Loading