DarthAffe/stable-diffusion.cpp
1
0
Fork 0
mirror of https://github.com/leejet/stable-diffusion.cpp.git synced 2025-12-13 05:48:56 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix the issue that occurs when using CUDA with k-quants weights

Browse Source
This commit is contained in:
leejet 2025-10-12 15:41:40 +08:00
parent 6ea2a75929
commit 98d6e71492
2 changed files with 13 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
common.hpp
View File

@ -245,7 +245,7 @@ public:
Activation activation = Activation::GEGLU, Activation activation = Activation::GEGLU,
bool force_prec_f32 = false) { bool force_prec_f32 = false) {
int64_t inner_dim = dim * mult; int64_t inner_dim = dim * mult;
SD_UNUSED(force_prec_f32);
if (activation == Activation::GELU) { if (activation == Activation::GELU) {
blocks["net.0"] = std::shared_ptr<GGMLBlock>(new GELU(dim, inner_dim)); blocks["net.0"] = std::shared_ptr<GGMLBlock>(new GELU(dim, inner_dim));
} else { } else {
@ -253,7 +253,7 @@ public:
} }
// net_1 is nn.Dropout(), skip for inference // net_1 is nn.Dropout(), skip for inference
blocks["net.2"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, dim_out, true, false, force_prec_f32)); blocks["net.2"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, dim_out));
} }
struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) { struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
@ -264,7 +264,13 @@ public:
auto net_2 = std::dynamic_pointer_cast<Linear>(blocks["net.2"]); auto net_2 = std::dynamic_pointer_cast<Linear>(blocks["net.2"]);
x = net_0->forward(ctx, x); // [ne3, ne2, ne1, inner_dim] x = net_0->forward(ctx, x); // [ne3, ne2, ne1, inner_dim]
x = net_2->forward(ctx, x); // [ne3, ne2, ne1, dim_out] // The purpose of the scale here is to prevent NaN issues in certain situations.
// For example, when using Vulkan without enabling force_prec_f32,
// or when using CUDA but the weights are k-quants.
float scale = 1.f / 128.f;
x = ggml_scale(ctx, x, scale);
x = net_2->forward(ctx, x); // [ne3, ne2, ne1, dim_out]
x = ggml_scale(ctx, x, 1.f / scale);
return x; return x;
} }
}; };

4
ggml_extend.hpp
View File

@ -56,6 +56,10 @@
#define __STATIC_INLINE__ static inline #define __STATIC_INLINE__ static inline
#endif #endif
#ifndef SD_UNUSED
#define SD_UNUSED(x) (void)(x)
#endif
__STATIC_INLINE__ void ggml_log_callback_default(ggml_log_level level, const char* text, void*) { __STATIC_INLINE__ void ggml_log_callback_default(ggml_log_level level, const char* text, void*) {
switch (level) { switch (level) {
case GGML_LOG_LEVEL_DEBUG: case GGML_LOG_LEVEL_DEBUG: