DarthAffe/stable-diffusion.cpp
1
0 Fork 0
mirror of https://github.com/leejet/stable-diffusion.cpp.git synced 2026-11-11 19:36:33 +00:00
Code Issues Packages Projects Releases Wiki Activity

refactor: unify ggml operation helpers around runner context (#2314)

Browse Source
This commit is contained in:
leejet authored and GitHub committed 2026-11-10 01:58:04 +08:00
1 parent 73fef08224
commit a3c49b13be
76 files changed
+1983 -1831
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

src/model/diffusion/krea2.hpp
+14 -16
View File
@@ -185,9 +185,9 @@ namespace Krea2 {
ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
ggml_tensor* scale = params["scale"];
if (ctx->weight_adapter) {
scale = ctx->weight_adapter->patch_weight(ctx->ggml_ctx, ctx->backend, scale, prefix + "scale.weight");
scale = ctx->weight_adapter->patch_weight(ctx, scale, prefix + "scale.weight");
}
scale = ggml_add(ctx->ggml_ctx, scale, ggml_ext_ones(ctx->ggml_ctx, scale->ne[0], 1, 1, 1));
scale = ggml_add(ctx->ggml_ctx, scale, ggml_ext_ones(ctx, scale->ne[0], 1, 1, 1));
x = ggml_rms_norm(ctx->ggml_ctx, x, eps);
x = ggml_mul_inplace(ctx->ggml_ctx, x, scale);
return x;
@@ -273,9 +273,7 @@ namespace Krea2 {
auto knorm = std::dynamic_pointer_cast<KreaRMSNorm>(blocks["qknorm.knorm"]);
auto wo = std::dynamic_pointer_cast<Linear>(blocks["wo"]);
if (sd_backend_is(ctx->backend, "Vulkan") || sd_backend_is(ctx->backend, "ROCm")) {
wo->set_force_prec_f32(true);
}
wo->set_precision_flags(GGML_PRECISION_F32_VULKAN | GGML_PRECISION_F32_ROCM);
int64_t L = x->ne[1];
int64_t N = x->ne[2];
@@ -316,11 +314,11 @@ namespace Krea2 {
std::vector<ggml_tensor*> forward(GGMLRunnerContext* ctx, ggml_tensor* vec) {
auto lin = params["lin"];
if (ctx->weight_adapter) {
lin = ctx->weight_adapter->patch_weight(ctx->ggml_ctx, ctx->backend, lin, prefix + "lin.weight");
lin = ctx->weight_adapter->patch_weight(ctx, lin, prefix + "lin.weight");
}
lin = ggml_repeat(ctx->ggml_ctx, lin, vec);
auto out = ggml_add(ctx->ggml_ctx, vec, lin);
return ggml_ext_chunk(ctx->ggml_ctx, out, 6, 0);
return ggml_ext_chunk(ctx, out, 6, 0);
}
};
@@ -342,10 +340,10 @@ namespace Krea2 {
std::vector<ggml_tensor*> forward(GGMLRunnerContext* ctx, ggml_tensor* vec) {
auto lin = params["lin"];
if (ctx->weight_adapter) {
lin = ctx->weight_adapter->patch_weight(ctx->ggml_ctx, ctx->backend, lin, prefix + "lin.weight");
lin = ctx->weight_adapter->patch_weight(ctx, lin, prefix + "lin.weight");
}
auto out = ggml_add(ctx->ggml_ctx, lin, vec);
return ggml_ext_chunk(ctx->ggml_ctx, out, 2, 1);
return ggml_ext_chunk(ctx, out, 2, 1);
}
};
@@ -541,7 +539,7 @@ namespace Krea2 {
auto linear_0 = std::dynamic_pointer_cast<Linear>(blocks["0"]);
auto linear_2 = std::dynamic_pointer_cast<Linear>(blocks["2"]);
x = linear_0->forward(ctx, x);
x = ggml_ext_gelu(ctx->ggml_ctx, x, false);
x = ggml_ext_gelu(ctx, x, false);
x = linear_2->forward(ctx, x);
return x;
}
@@ -555,7 +553,7 @@ namespace Krea2 {
ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
auto linear_1 = std::dynamic_pointer_cast<Linear>(blocks["1"]);
x = ggml_ext_gelu(ctx->ggml_ctx, x, false);
x = ggml_ext_gelu(ctx, x, false);
x = linear_1->forward(ctx, x);
return x;
}
@@ -575,7 +573,7 @@ namespace Krea2 {
auto linear_3 = std::dynamic_pointer_cast<Linear>(blocks["3"]);
x = norm->forward(ctx, x);
x = linear_1->forward(ctx, x);
x = ggml_ext_gelu(ctx->ggml_ctx, x, true);
x = ggml_ext_gelu(ctx, x, true);
x = linear_3->forward(ctx, x);
return x;
}
@@ -656,7 +654,7 @@ namespace Krea2 {
int64_t ref_len = img->ne[1] - img_len;
img = first->forward(ctx, img);
auto t = ggml_ext_timestep_embedding(ctx->ggml_ctx, timestep, static_cast<int>(config.timestep_dim), 10000, 1000.f);
auto t = ggml_ext_timestep_embedding(ctx, timestep, static_cast<int>(config.timestep_dim), 10000, 1000.f);
t = tmlp->forward(ctx, t);
t = ggml_reshape_3d(ctx->ggml_ctx, t, t->ne[0], 1, t->ne[1]);
auto tvec = tproj->forward(ctx, t);
@@ -665,7 +663,7 @@ namespace Krea2 {
if (ref_latents.size() > 0 && zero_timestep_refs) {
// "index_timestep_zero" mode: use timestep = 0 for ref latents
auto timestep_0 = ggml_scale(ctx->ggml_ctx, timestep, 0.0f);
auto t_0 = ggml_ext_timestep_embedding(ctx->ggml_ctx, timestep_0, static_cast<int>(config.timestep_dim), 10000, 1000.f);
auto t_0 = ggml_ext_timestep_embedding(ctx, timestep_0, static_cast<int>(config.timestep_dim), 10000, 1000.f);
t_0 = tmlp->forward(ctx, t_0);
t_0 = ggml_reshape_3d(ctx->ggml_ctx, t_0, t_0->ne[0], 1, t_0->ne[1]);
tvec_0 = tproj->forward(ctx, t_0);
@@ -683,9 +681,9 @@ namespace Krea2 {
sd::ggml_graph_cut::mark_graph_cut(hidden_states, "krea2.blocks." + std::to_string(i), "hidden_states");
}
hidden_states = ggml_ext_slice(ctx->ggml_ctx, hidden_states, 1, txt_len, txt_len + img_len);
hidden_states = ggml_ext_slice(ctx, hidden_states, 1, txt_len, txt_len + img_len);
hidden_states = last->forward(ctx, hidden_states, t);
hidden_states = DiT::unpatchify_and_crop(ctx->ggml_ctx, hidden_states, H, W, config.patch_size, config.patch_size, true);
hidden_states = DiT::unpatchify_and_crop(ctx, hidden_states, H, W, config.patch_size, config.patch_size, true);
return hidden_states;
}
};