DarthAffe/stable-diffusion.cpp
1
0 Fork 0
mirror of https://github.com/leejet/stable-diffusion.cpp.git synced 2026-11-12 11:56:36 +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/wan_audio.hpp
+11 -11
View File
@@ -27,7 +27,7 @@ namespace WAN {
ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
// Replicate the first sample for causal left padding.
if (kernel_size_ > 1) {
auto first = ggml_ext_slice(ctx->ggml_ctx, x, 0, 0, 1);
auto first = ggml_ext_slice(ctx, x, 0, 0, 1);
for (int i = 0; i < kernel_size_ - 1; i++) {
x = ggml_concat(ctx->ggml_ctx, first, x, 0);
}
@@ -57,7 +57,7 @@ namespace WAN {
x = std::dynamic_pointer_cast<LayerNorm>(blocks[norm_key])->forward(ctx, x);
x = ggml_silu(ctx->ggml_ctx, x);
if (to_conv_layout) {
x = ggml_ext_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 1, 0, 2, 3));
x = ggml_ext_cont(ctx, ggml_permute(ctx->ggml_ctx, x, 1, 0, 2, 3));
}
return x;
}
@@ -87,11 +87,11 @@ namespace WAN {
auto norm1 = std::dynamic_pointer_cast<LayerNorm>(blocks["norm1"]);
std::vector<ggml_tensor*> tokens;
// Each token group is normalized independently over channels.
for (auto& group : ggml_ext_chunk(ctx->ggml_ctx, local, num_token_, 1)) {
for (auto& group : ggml_ext_chunk(ctx, local, num_token_, 1)) {
ggml_tensor* s = ggml_permute(ctx->ggml_ctx, group, 1, 0, 2, 3);
s = norm1->forward(ctx, s);
s = ggml_silu(ctx->ggml_ctx, s);
s = ggml_ext_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, s, 1, 0, 2, 3));
s = ggml_ext_cont(ctx, ggml_permute(ctx->ggml_ctx, s, 1, 0, 2, 3));
s = conv_norm_silu(ctx, s, "conv2", "norm2", true);
s = conv_norm_silu(ctx, s, "conv3", "norm3", false);
tokens.push_back(ggml_reshape_3d(ctx->ggml_ctx, s, s->ne[0], 1, s->ne[1]));
@@ -99,7 +99,7 @@ namespace WAN {
auto padding = ggml_reshape_3d(ctx->ggml_ctx, params["padding_tokens"], hidden_dim_, 1, 1);
padding = ggml_repeat(ctx->ggml_ctx, padding, tokens[0]);
tokens.push_back(padding);
ggml_tensor* local_out = ggml_ext_vec_concat(ctx->ggml_ctx, tokens, 1);
ggml_tensor* local_out = ggml_ext_vec_concat(ctx, tokens, 1);
if (!need_global_) {
return {local_out, nullptr};
@@ -141,10 +141,10 @@ namespace WAN {
auto x = ggml_mul(ctx->ggml_ctx, features, ggml_reshape_3d(ctx->ggml_ctx, weights, 1, 1, num_layers_));
x = ggml_div(ctx->ggml_ctx, x, ggml_sum(ctx->ggml_ctx, weights));
// Move the layer axis to ggml dimension 0 for reduction.
x = ggml_ext_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, x, 2, 0, 1, 3));
x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 2, 0, 1, 3));
x = ggml_sum_rows(ctx->ggml_ctx, x);
x = ggml_reshape_2d(ctx->ggml_ctx, x, x->ne[1], x->ne[2]);
x = ggml_ext_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, x, 1, 0, 2, 3));
x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 1, 0, 2, 3));
return std::dynamic_pointer_cast<WanMotionEncoder>(blocks["encoder"])->forward(ctx, x);
}
};
@@ -187,12 +187,12 @@ namespace WAN {
auto temb = ggml_silu(ctx->ggml_ctx, audio_global);
temb = adain_linear->forward(ctx, temb);
auto shift = ggml_ext_slice(ctx->ggml_ctx, temb, 0, 0, dim_);
auto scale = ggml_ext_slice(ctx->ggml_ctx, temb, 0, dim_, dim_ * 2);
auto shift = ggml_ext_slice(ctx, temb, 0, 0, dim_);
auto scale = ggml_ext_slice(ctx, temb, 0, dim_, dim_ * 2);
shift = ggml_reshape_3d(ctx->ggml_ctx, shift, dim_, 1, T);
scale = ggml_reshape_3d(ctx->ggml_ctx, scale, dim_, 1, T);
auto x_vid = ggml_ext_slice(ctx->ggml_ctx, x, 1, 0, seq_len);
auto x_vid = ggml_ext_slice(ctx, x, 1, 0, seq_len);
auto h = ggml_reshape_3d(ctx->ggml_ctx, x_vid, dim_, n_tok, T);
h = adain_norm->forward(ctx, h);
h = ggml_add(ctx->ggml_ctx, h, ggml_mul(ctx->ggml_ctx, h, scale));
@@ -203,7 +203,7 @@ namespace WAN {
auto x_head = ggml_add(ctx->ggml_ctx, x_vid, res);
if (seq_len < n_token) {
auto x_tail = ggml_ext_slice(ctx->ggml_ctx, x, 1, seq_len, n_token);
auto x_tail = ggml_ext_slice(ctx, x, 1, seq_len, n_token);
return ggml_concat(ctx->ggml_ctx, x_head, x_tail, 1);
}
return x_head;