common.hpp

@@ -243,8 +243,9 @@ public:
                 int64_t dim_out,
                 int64_t mult          = 4,
                 Activation activation = Activation::GEGLU,
-                bool precision_fix    = false) {
+                bool force_prec_f32   = false) {
         int64_t inner_dim = dim * mult;
+
         if (activation == Activation::GELU) {
             blocks["net.0"] = std::shared_ptr<GGMLBlock>(new GELU(dim, inner_dim));
         } else {
@@ -252,14 +253,7 @@ public:
         }
 
         // net_1 is nn.Dropout(), skip for inference
-        float scale = 1.f;
+        blocks["net.2"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, dim_out, true, false, force_prec_f32));
-        if (precision_fix) {
-            scale = 1.f / 128.f;
-        }
-        // 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.
-        blocks["net.2"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, dim_out, true, false, false, scale));
     }
 
     struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {

ggml_extend.hpp

@@ -56,10 +56,6 @@
 #define __STATIC_INLINE__ static inline
 #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*) {
     switch (level) {
         case GGML_LOG_LEVEL_DEBUG:
@@ -941,18 +937,11 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_nn_linear(struct ggml_context* ctx,
                                                      struct ggml_tensor* x,
                                                      struct ggml_tensor* w,
                                                      struct ggml_tensor* b,
-                                                     bool force_prec_f32 = false,
+                                                     bool force_prec_f32 = false) {
-                                                     float scale         = 1.f) {
-    if (scale != 1.f) {
-        x = ggml_scale(ctx, x, scale);
-    }
     x = ggml_mul_mat(ctx, w, x);
     if (force_prec_f32) {
         ggml_mul_mat_set_prec(x, GGML_PREC_F32);
     }
-    if (scale != 1.f) {
-        x = ggml_scale(ctx, x, 1.f / scale);
-    }
     if (b != NULL) {
         x = ggml_add_inplace(ctx, x, b);
     }
@@ -1966,7 +1955,6 @@ protected:
     bool bias;
     bool force_f32;
     bool force_prec_f32;
-    float scale;
 
     void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") {
         enum ggml_type wtype = get_type(prefix + "weight", tensor_types, GGML_TYPE_F32);
@@ -1985,14 +1973,12 @@ public:
            int64_t out_features,
            bool bias           = true,
            bool force_f32      = false,
-           bool force_prec_f32 = false,
+           bool force_prec_f32 = false)
-           float scale         = 1.f)
         : in_features(in_features),
           out_features(out_features),
           bias(bias),
           force_f32(force_f32),
-          force_prec_f32(force_prec_f32),
+          force_prec_f32(force_prec_f32) {}
-          scale(scale) {}
 
     struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
         struct ggml_tensor* w = params["weight"];
@@ -2000,7 +1986,7 @@ public:
         if (bias) {
             b = params["bias"];
         }
-        return ggml_nn_linear(ctx, x, w, b, force_prec_f32, scale);
+        return ggml_nn_linear(ctx, x, w, b, force_prec_f32);
     }
 };
 

qwen_image.hpp

@@ -97,10 +97,7 @@ namespace Qwen {
             blocks["to_out.0"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, out_dim, out_bias));
             // to_out.1 is nn.Dropout
 
-            float scale = 1.f / 32.f;
+            blocks["to_add_out"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, out_context_dim, out_bias));
-            // The purpose of the scale here is to prevent NaN issues in certain situations.
-            // For example when using CUDA but the weights are k-quants (not all prompts).
-            blocks["to_add_out"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, out_context_dim, out_bias, false, false, scale));
         }
 
         std::pair<ggml_tensor*, ggml_tensor*> forward(struct ggml_context* ctx,