DarthAffe/stable-diffusion.cpp
1
0
Fork 0
mirror of https://github.com/leejet/stable-diffusion.cpp.git synced 2026-09-24 20:20:37 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix: avoid narrowing conversion in SigVQ patch embedding and format code

Browse Source
This commit is contained in:
leejet 2026-09-21 01:27:31 +08:00
parent 187b2561ea
commit b56c68617d
4 changed files with 19 additions and 19 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

24
src/model/diffusion/qwen_image_2_1.hpp
View File

@ -193,16 +193,16 @@ namespace Qwen {
class QwenImage21TransformerBlock : public GGMLBlock { class QwenImage21TransformerBlock : public GGMLBlock {
public: public:
QwenImage21TransformerBlock(const QwenImage21Config& config) { QwenImage21TransformerBlock(const QwenImage21Config& config) {
blocks["img_norm1"] = std::make_shared<LayerNorm>(config.hidden_size, 1e-6f, false); blocks["img_norm1"] = std::make_shared<LayerNorm>(config.hidden_size, 1e-6f, false);
blocks["img_norm2"] = std::make_shared<LayerNorm>(config.hidden_size, 1e-6f, false); blocks["img_norm2"] = std::make_shared<LayerNorm>(config.hidden_size, 1e-6f, false);
blocks["attn"] = std::make_shared<QwenImage21Attention>(config); blocks["attn"] = std::make_shared<QwenImage21Attention>(config);
if (config.fused_mlp) { if (config.fused_mlp) {
blocks["img_mlp.gate_up"] = std::make_shared<Linear>(config.hidden_size, 2 * config.intermediate_size, false); blocks["img_mlp.gate_up"] = std::make_shared<Linear>(config.hidden_size, 2 * config.intermediate_size, false);
} else { } else {
blocks["img_mlp.proj"] = std::make_shared<Linear>(config.hidden_size, config.intermediate_size, false); blocks["img_mlp.proj"] = std::make_shared<Linear>(config.hidden_size, config.intermediate_size, false);
blocks["img_mlp.gate_layer"] = std::make_shared<Linear>(config.hidden_size, config.intermediate_size, false); blocks["img_mlp.gate_layer"] = std::make_shared<Linear>(config.hidden_size, config.intermediate_size, false);
} }
blocks["img_mlp.out"] = std::make_shared<Linear>(config.intermediate_size, config.hidden_size, false); blocks["img_mlp.out"] = std::make_shared<Linear>(config.intermediate_size, config.hidden_size, false);
} }
static ggml_tensor* modulate(ggml_context* ctx, ggml_tensor* x, ggml_tensor* params, int64_t prefix_length, bool gate = false) { static ggml_tensor* modulate(ggml_context* ctx, ggml_tensor* x, ggml_tensor* params, int64_t prefix_length, bool gate = false) {
@ -220,12 +220,12 @@ namespace Qwen {
} }
ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, const std::vector<ggml_tensor*>& modulation, ggml_tensor* pe, const QwenImage21Layout& layout, const std::vector<ggml_tensor*>& masks) { ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, const std::vector<ggml_tensor*>& modulation, ggml_tensor* pe, const QwenImage21Layout& layout, const std::vector<ggml_tensor*>& masks) {
auto h = std::dynamic_pointer_cast<LayerNorm>(blocks["img_norm1"])->forward(ctx, x); auto h = std::dynamic_pointer_cast<LayerNorm>(blocks["img_norm1"])->forward(ctx, x);
h = modulate(ctx->ggml_ctx, h, modulation[0], layout.prefix_length); h = modulate(ctx->ggml_ctx, h, modulation[0], layout.prefix_length);
h = std::dynamic_pointer_cast<QwenImage21Attention>(blocks["attn"])->forward(ctx, h, pe, layout.segments, masks); h = std::dynamic_pointer_cast<QwenImage21Attention>(blocks["attn"])->forward(ctx, h, pe, layout.segments, masks);
x = ggml_add(ctx->ggml_ctx, x, modulate(ctx->ggml_ctx, h, modulation[1], layout.prefix_length, true)); x = ggml_add(ctx->ggml_ctx, x, modulate(ctx->ggml_ctx, h, modulation[1], layout.prefix_length, true));
h = std::dynamic_pointer_cast<LayerNorm>(blocks["img_norm2"])->forward(ctx, x); h = std::dynamic_pointer_cast<LayerNorm>(blocks["img_norm2"])->forward(ctx, x);
h = modulate(ctx->ggml_ctx, h, modulation[2], layout.prefix_length); h = modulate(ctx->ggml_ctx, h, modulation[2], layout.prefix_length);
ggml_tensor* gate; ggml_tensor* gate;
auto fused = blocks.find("img_mlp.gate_up"); auto fused = blocks.find("img_mlp.gate_up");
if (fused != blocks.end()) { if (fused != blocks.end()) {
@ -237,8 +237,8 @@ namespace Qwen {
gate = std::dynamic_pointer_cast<Linear>(blocks["img_mlp.gate_layer"])->forward(ctx, h); gate = std::dynamic_pointer_cast<Linear>(blocks["img_mlp.gate_layer"])->forward(ctx, h);
h = std::dynamic_pointer_cast<Linear>(blocks["img_mlp.proj"])->forward(ctx, h); h = std::dynamic_pointer_cast<Linear>(blocks["img_mlp.proj"])->forward(ctx, h);
} }
h = ggml_mul(ctx->ggml_ctx, h, ggml_silu(ctx->ggml_ctx, gate)); h = ggml_mul(ctx->ggml_ctx, h, ggml_silu(ctx->ggml_ctx, gate));
h = std::dynamic_pointer_cast<Linear>(blocks["img_mlp.out"])->forward(ctx, h); h = std::dynamic_pointer_cast<Linear>(blocks["img_mlp.out"])->forward(ctx, h);
return ggml_add(ctx->ggml_ctx, x, modulate(ctx->ggml_ctx, h, modulation[3], layout.prefix_length, true)); return ggml_add(ctx->ggml_ctx, x, modulate(ctx->ggml_ctx, h, modulation[3], layout.prefix_length, true));
} }
}; };

2
src/model/te/llada_image_te.hpp
View File

@ -265,7 +265,7 @@ namespace LLaDAImageTE {
struct SigVQConfig { struct SigVQConfig {
int64_t image_size = 2048; int64_t image_size = 2048;
int64_t patch_size = 16; int patch_size = 16;
int64_t in_channels = 3; int64_t in_channels = 3;
int64_t hidden_size = 1536; int64_t hidden_size = 1536;
int64_t intermediate_size = 6144; int64_t intermediate_size = 6144;

10
src/model/vae/wan_vae.hpp
View File

@ -1070,11 +1070,11 @@ namespace WAN {
} }
if (version == VERSION_QWEN_IMAGE_2_1) { if (version == VERSION_QWEN_IMAGE_2_1) {
wan2_2 = true; wan2_2 = true;
dec_dim = 144; dec_dim = 144;
z_dim = 64; z_dim = 64;
input_channels = 4; input_channels = 4;
dim_mult = {1, 2, 4, 8, 8}; dim_mult = {1, 2, 4, 8, 8};
} }
if (is_2D) { if (is_2D) {

2
src/model_loader_files.cpp
View File

@ -147,7 +147,7 @@ bool ModelLoader::add_file_impl(const std::string& path, const std::string& pref
} }
if (tensor.index_in_zip < 0) { if (tensor.index_in_zip < 0) {
const auto& stamp = physical_files[tensor.file_index]; const auto& stamp = physical_files[tensor.file_index];
if (tensor.offset > stamp.size || static_cast<uint64_t>(tensor.nbytes_to_read()) > stamp.size - tensor.offset) { //kcpp int8 fp8 if (tensor.offset > stamp.size || static_cast<uint64_t>(tensor.nbytes_to_read()) > stamp.size - tensor.offset) { // kcpp int8 fp8
LOG_ERROR("tensor '%s' extends beyond its model file", tensor.name.c_str()); LOG_ERROR("tensor '%s' extends beyond its model file", tensor.name.c_str());
return false; return false;
} }