DarthAffe/stable-diffusion.cpp
1
0 Fork 0
mirror of https://github.com/leejet/stable-diffusion.cpp.git synced 2026-11-11 11:26:34 +00:00
Code Issues Packages Projects Releases Wiki Activity

perf: use ggml rope apply op on supported backends (#2113)

Browse Source
This commit is contained in:
leejet authored and GitHub committed 2026-10-09 04:32:36 +08:00
1 parent e16d26a96a
commit 228c707fde
7 files changed
+36 -21
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

src/model/diffusion/minimax_h3.hpp
+5 -4
View File
@@ -160,12 +160,13 @@ namespace MiniMaxH3 {
return ggml_reshape_3d(ctx, x, x->ne[0], x->ne[1], x->ne[2] * x->ne[3]);
}
static ggml_tensor* apply_partial_rope(ggml_context* ctx,
static ggml_tensor* apply_partial_rope(GGMLRunnerContext* runner_ctx,
ggml_tensor* x,
ggml_tensor* pe) {
auto ctx = runner_ctx->ggml_ctx;
int64_t rot_dim = pe->ne[2] * 2;
GGML_ASSERT(rot_dim <= x->ne[0]);
auto rotated = Rope::apply_rope(ctx,
auto rotated = Rope::apply_rope(runner_ctx,
ggml_ext_slice(ctx, x, 0, 0, rot_dim),
pe,
false);
@@ -209,8 +210,8 @@ namespace MiniMaxH3 {
q = q_norm->forward(ctx, q);
k = k_norm->forward(ctx, k);
if (pe != nullptr) {
q = apply_partial_rope(ctx->ggml_ctx, q, pe);
k = apply_partial_rope(ctx->ggml_ctx, k, pe);
q = apply_partial_rope(ctx, q, pe);
k = apply_partial_rope(ctx, k, pe);
} else {
q = attention_layout(ctx->ggml_ctx, q);
k = attention_layout(ctx->ggml_ctx, k);