Skip to content

perf(moss): avoid redundant CUDA projection copies #2209

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
leejet merged 4 commits into from
Oct 25, 2026
Merged

perf(moss): avoid redundant CUDA projection copies #2209

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
627e126
perf(moss): avoid redundant CUDA projection copies
mirek190 Oct 24, 2026
042de50
docs(moss): remove unnecessary optimization documentation
mirek190 Oct 24, 2026
d819682
refactor: simplify MOSS QKV views and value preparation
leejet Oct 25, 2026
fdaf518
fix: restore capability check for direct value cast
leejet Oct 25, 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
14 changes: 8 additions & 6 deletions src/core/ggml_extend.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -678,21 +678,23 @@ ggml_tensor* ggml_ext_attention_ext(ggml_context* ctx,
// CUDA FA MMA starts at 64 channels; keep the original head's attention scale.
q_in = ggml_pad(ctx, q_in, static_cast<int>(64 - d_head), 0, 0, 0);
k_in = ggml_pad(ctx, k_in, static_cast<int>(64 - d_head), 0, 0, 0);
v_in = ggml_pad(ctx, v_in, static_cast<int>(64 - d_head), 0, 0, 0);
}
if (kv_scale != 1.0f) {
k_in = ggml_ext_scale(ctx, k_in, kv_scale);
}
k_in = ggml_cast(ctx, k_in, GGML_TYPE_F16);

v_in = ggml_ext_cont(ctx, ggml_permute(ctx, v_in, 0, 2, 1, 3));
v_in = ggml_reshape_3d(ctx, v_in, d_head, L_k, n_kv_head * N);
if (pad_head) {
v_in = ggml_pad(ctx, v_in, static_cast<int>(64 - d_head), 0, 0, 0);
}
if (kv_scale != 1.0f) {
v_in = ggml_ext_scale(ctx, v_in, kv_scale);
}
v_in = ggml_cast(ctx, v_in, GGML_TYPE_F16);
v_in = ggml_permute(ctx, v_in, 0, 2, 1, 3);
auto v_f16 = ggml_cast(ctx, v_in, GGML_TYPE_F16);
if (!ggml_backend_supports_op(backend, v_f16)) {
v_f16 = ggml_cast(ctx, ggml_ext_cont(ctx, v_in), GGML_TYPE_F16);
}
v_in = v_f16;
v_in = ggml_reshape_3d(ctx, v_in, v_in->ne[0], L_k, n_kv_head * N);

if (mask_in != nullptr) {
// ggml_flash_attn_ext expects the mask as a contiguous F16 tensor shaped
Expand Down
3 changes: 2 additions & 1 deletion src/model/diffusion/moss_image.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -122,7 +122,8 @@ namespace MOSSImage {
y = ggml_add(c, ggml_add(c, y, ggml_mul(c, y, part(1))), part(0));
auto qkv = linear("proj.qkv", y);
auto split = [&](int i) {
return ggml_reshape_4d(c, ggml_ext_slice(c, qkv, 0, i * config.dim, (i + 1) * config.dim), config.head_dim, config.dim / config.head_dim, x->ne[1], x->ne[2]);
return ggml_view_4d(c, qkv, config.head_dim, config.dim / config.head_dim, x->ne[1], x->ne[2],
config.head_dim * qkv->nb[0], qkv->nb[1], qkv->nb[2], i * config.dim * qkv->nb[0]);
};
auto q = norm("q_norm", split(0));
auto k = norm("k_norm", split(1));
Expand Down
Loading