diff --git a/clip.hpp b/clip.hpp
index 7a6ebe9..4eec024 100644
--- a/clip.hpp
+++ b/clip.hpp
@@ -479,9 +479,9 @@ public:
 
         x = fc1->forward(ctx, x);
         if (use_gelu) {
-            x = ggml_gelu_inplace(ctx->ggml_ctx, x);
+            x = ggml_ext_gelu(ctx->ggml_ctx, x, true);
         } else {
-            x = ggml_gelu_quick_inplace(ctx->ggml_ctx, x);
+            x = ggml_ext_gelu_quick(ctx->ggml_ctx, x, true);
         }
         x = fc2->forward(ctx, x);
         return x;
diff --git a/common.hpp b/common.hpp
index 13ab103..7183eb8 100644
--- a/common.hpp
+++ b/common.hpp
@@ -200,7 +200,7 @@ public:
 
         gate = ggml_cont(ctx->ggml_ctx, gate);
 
-        gate = ggml_gelu_inplace(ctx->ggml_ctx, gate);
+        gate = ggml_ext_gelu(ctx->ggml_ctx, gate, true);
 
         x = ggml_mul(ctx->ggml_ctx, x, gate);  // [ne3, ne2, ne1, dim_out]
 
@@ -220,7 +220,7 @@ public:
         auto proj = std::dynamic_pointer_cast<Linear>(blocks["proj"]);
 
         x = proj->forward(ctx, x);
-        x = ggml_gelu_inplace(ctx->ggml_ctx, x);
+        x = ggml_ext_gelu(ctx->ggml_ctx, x, true);
         return x;
     }
 };
@@ -536,8 +536,8 @@ public:
         // image_only_indicator is always tensor([0.])
         float alpha = get_alpha();
         auto x      = ggml_add(ctx->ggml_ctx,
-                               ggml_scale(ctx->ggml_ctx, x_spatial, alpha),
-                               ggml_scale(ctx->ggml_ctx, x_temporal, 1.0f - alpha));
+                               ggml_ext_scale(ctx->ggml_ctx, x_spatial, alpha),
+                               ggml_ext_scale(ctx->ggml_ctx, x_temporal, 1.0f - alpha));
         return x;
     }
 };
diff --git a/esrgan.hpp b/esrgan.hpp
index 961e84f..f740c2b 100644
--- a/esrgan.hpp
+++ b/esrgan.hpp
@@ -51,7 +51,7 @@ public:
         x_cat      = ggml_concat(ctx->ggml_ctx, x_cat, x4, 2);
         auto x5    = conv5->forward(ctx, x_cat);
 
-        x5 = ggml_add(ctx->ggml_ctx, ggml_scale(ctx->ggml_ctx, x5, 0.2f), x);
+        x5 = ggml_add(ctx->ggml_ctx, ggml_ext_scale(ctx->ggml_ctx, x5, 0.2f), x);
         return x5;
     }
 };
@@ -76,7 +76,7 @@ public:
         out      = rdb2->forward(ctx, out);
         out      = rdb3->forward(ctx, out);
 
-        out = ggml_add(ctx->ggml_ctx, ggml_scale(ctx->ggml_ctx, out, 0.2f), x);
+        out = ggml_add(ctx->ggml_ctx, ggml_ext_scale(ctx->ggml_ctx, out, 0.2f), x);
         return out;
     }
 };
diff --git a/flux.hpp b/flux.hpp
index 83a4a22..ff8c189 100644
--- a/flux.hpp
+++ b/flux.hpp
@@ -153,7 +153,7 @@ namespace Flux {
             if (use_mlp_silu_act) {
                 x = ggml_ext_silu_act(ctx->ggml_ctx, x);
             } else {
-                x = ggml_gelu_inplace(ctx->ggml_ctx, x);
+                x = ggml_ext_gelu(ctx->ggml_ctx, x, true);
             }
             x = mlp_2->forward(ctx, x);
             return x;
@@ -511,7 +511,7 @@ namespace Flux {
             } else if (use_mlp_silu_act) {
                 mlp = ggml_ext_silu_act(ctx->ggml_ctx, mlp);
             } else {
-                mlp = ggml_gelu_inplace(ctx->ggml_ctx, mlp);
+                mlp = ggml_ext_gelu(ctx->ggml_ctx, mlp, true);
             }
             auto attn_mlp = ggml_concat(ctx->ggml_ctx, attn, mlp, 0);  // [N, n_token, hidden_size + mlp_hidden_dim]
             auto output   = linear2->forward(ctx, attn_mlp);           // [N, n_token, hidden_size]
diff --git a/ggml_extend.hpp b/ggml_extend.hpp
index f61f7c6..fedab38 100644
--- a/ggml_extend.hpp
+++ b/ggml_extend.hpp
@@ -949,6 +949,49 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_ext_group_norm_32(struct ggml_context
     return ggml_group_norm(ctx, a, 32, eps);
 }
 
+__STATIC_INLINE__ struct ggml_tensor* ggml_ext_scale(struct ggml_context* ctx,
+                                                     struct ggml_tensor* x,
+                                                     float factor,
+                                                     bool inplace = false) {
+    if (!ggml_is_contiguous(x)) {
+        x = ggml_cont(ctx, x);
+    }
+    if (inplace) {
+        x = ggml_scale_inplace(ctx, x, factor);
+    } else {
+        x = ggml_scale(ctx, x, factor);
+    }
+    return x;
+}
+
+__STATIC_INLINE__ struct ggml_tensor* ggml_ext_gelu(struct ggml_context* ctx,
+                                                    struct ggml_tensor* x,
+                                                    bool inplace = false) {
+    if (!ggml_is_contiguous(x)) {
+        x = ggml_cont(ctx, x);
+    }
+    if (inplace) {
+        x = ggml_gelu_inplace(ctx, x);
+    } else {
+        x = ggml_gelu(ctx, x);
+    }
+    return x;
+}
+
+__STATIC_INLINE__ struct ggml_tensor* ggml_ext_gelu_quick(struct ggml_context* ctx,
+                                                          struct ggml_tensor* x,
+                                                          bool inplace = false) {
+    if (!ggml_is_contiguous(x)) {
+        x = ggml_cont(ctx, x);
+    }
+    if (inplace) {
+        x = ggml_gelu_quick_inplace(ctx, x);
+    } else {
+        x = ggml_gelu_quick(ctx, x);
+    }
+    return x;
+}
+
 __STATIC_INLINE__ struct ggml_tensor* ggml_ext_linear(struct ggml_context* ctx,
                                                       struct ggml_tensor* x,
                                                       struct ggml_tensor* w,
@@ -956,10 +999,7 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_ext_linear(struct ggml_context* ctx,
                                                       bool force_prec_f32 = false,
                                                       float scale         = 1.f) {
     if (scale != 1.f) {
-        if (!ggml_is_contiguous(x)) {
-            x = ggml_cont(ctx, x);
-        }
-        x = ggml_scale(ctx, x, scale);
+        x = ggml_ext_scale(ctx, x, scale);
     }
     if (x->ne[2] * x->ne[3] > 1024) {
         // workaround: avoid ggml cuda error
@@ -978,7 +1018,7 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_ext_linear(struct ggml_context* ctx,
         }
     }
     if (scale != 1.f) {
-        x = ggml_scale(ctx, x, 1.f / scale);
+        x = ggml_ext_scale(ctx, x, 1.f / scale);
     }
     if (b != nullptr) {
         x = ggml_add_inplace(ctx, x, b);
@@ -1047,7 +1087,7 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_ext_conv_2d(struct ggml_context* ctx,
                                                        bool circular_y = false,
                                                        float scale     = 1.f) {
     if (scale != 1.f) {
-        x = ggml_scale(ctx, x, scale);
+        x = ggml_ext_scale(ctx, x, scale);
     }
     if (w->ne[2] != x->ne[2] && ggml_n_dims(w) == 2) {
         w = ggml_reshape_4d(ctx, w, 1, 1, w->ne[0], w->ne[1]);
@@ -1065,7 +1105,7 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_ext_conv_2d(struct ggml_context* ctx,
         x = ggml_conv_2d(ctx, w, x, s0, s1, p0, p1, d0, d1);
     }
     if (scale != 1.f) {
-        x = ggml_scale(ctx, x, 1.f / scale);
+        x = ggml_ext_scale(ctx, x, 1.f / scale);
     }
     if (b != nullptr) {
         b = ggml_reshape_4d(ctx, b, 1, 1, b->ne[0], 1);
@@ -1163,7 +1203,7 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_ext_full(struct ggml_context* ctx,
                                                     int64_t ne2,
                                                     int64_t ne3) {
     auto one = ggml_get_tensor(ctx, "ggml_runner_build_in_tensor:one");
-    auto t   = ggml_scale(ctx, one, value);                 // [1,]
+    auto t   = ggml_ext_scale(ctx, one, value);             // [1,]
     t        = ggml_repeat_4d(ctx, t, ne0, ne1, ne2, ne3);  // [ne0, ne1, ne2, ne3]
     return t;
 }
@@ -1263,7 +1303,7 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_ext_attention_ext(struct ggml_context
             k_in = ggml_pad(ctx, k_in, 0, kv_pad, 0, 0);
         }
         if (kv_scale != 1.0f) {
-            k_in = ggml_scale(ctx, k_in, kv_scale);
+            k_in = ggml_ext_scale(ctx, k_in, kv_scale);
         }
         k_in = ggml_cast(ctx, k_in, GGML_TYPE_F16);
 
@@ -1273,7 +1313,7 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_ext_attention_ext(struct ggml_context
             v_in = ggml_pad(ctx, v_in, 0, kv_pad, 0, 0);
         }
         if (kv_scale != 1.0f) {
-            v_in = ggml_scale(ctx, v_in, kv_scale);
+            v_in = ggml_ext_scale(ctx, v_in, kv_scale);
         }
         v_in = ggml_cast(ctx, v_in, GGML_TYPE_F16);
 
@@ -1305,7 +1345,7 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_ext_attention_ext(struct ggml_context
         auto out = ggml_flash_attn_ext(ctx, q_in, k_in, v_in, mask_in, scale / kv_scale, 0, 0);
         ggml_flash_attn_ext_set_prec(out, GGML_PREC_F32);
         if (kv_scale != 1.0f) {
-            out = ggml_scale(ctx, out, 1.0f / kv_scale);
+            out = ggml_ext_scale(ctx, out, 1.0f / kv_scale);
         }
         return out;
     };
@@ -1515,7 +1555,7 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_ext_timestep_embedding(
     int dim,
     int max_period    = 10000,
     float time_factor = 1.0f) {
-    timesteps = ggml_scale(ctx, timesteps, time_factor);
+    timesteps = ggml_ext_scale(ctx, timesteps, time_factor);
     return ggml_timestep_embedding(ctx, timesteps, dim, max_period);
 }
 
diff --git a/llm.hpp b/llm.hpp
index 781774d..7feb8d3 100644
--- a/llm.hpp
+++ b/llm.hpp
@@ -638,7 +638,7 @@ namespace LLM {
             x = ln_q->forward(ctx, x);
             x = ggml_reshape_2d(ctx->ggml_ctx, x, hidden_size, ggml_nelements(x) / hidden_size);
             x = mlp_0->forward(ctx, x);
-            x = ggml_gelu(ctx->ggml_ctx, x);
+            x = ggml_ext_gelu(ctx->ggml_ctx, x);
             x = mlp_2->forward(ctx, x);
             return x;
         }
diff --git a/lora.hpp b/lora.hpp
index 7d83ec5..e5d9906 100644
--- a/lora.hpp
+++ b/lora.hpp
@@ -195,7 +195,7 @@ struct LoraModel : public GGMLRunner {
             scale_value *= multiplier;
 
             auto curr_updown = ggml_ext_merge_lora(ctx, lora_down, lora_up, lora_mid);
-            curr_updown      = ggml_scale_inplace(ctx, curr_updown, scale_value);
+            curr_updown      = ggml_ext_scale(ctx, curr_updown, scale_value, true);
 
             if (updown == nullptr) {
                 updown = curr_updown;
@@ -235,7 +235,7 @@ struct LoraModel : public GGMLRunner {
             float scale_value = 1.0f;
             scale_value *= multiplier;
 
-            curr_updown = ggml_scale_inplace(ctx, curr_updown, scale_value);
+            curr_updown = ggml_ext_scale(ctx, curr_updown, scale_value, true);
 
             if (updown == nullptr) {
                 updown = curr_updown;
@@ -340,7 +340,7 @@ struct LoraModel : public GGMLRunner {
             struct ggml_tensor* updown_1 = ggml_ext_merge_lora(ctx, hada_1_down, hada_1_up, hada_1_mid);
             struct ggml_tensor* updown_2 = ggml_ext_merge_lora(ctx, hada_2_down, hada_2_up, hada_2_mid);
             auto curr_updown             = ggml_mul_inplace(ctx, updown_1, updown_2);
-            curr_updown                  = ggml_scale_inplace(ctx, curr_updown, scale_value);
+            curr_updown                  = ggml_ext_scale(ctx, curr_updown, scale_value, true);
             if (updown == nullptr) {
                 updown = curr_updown;
             } else {
@@ -456,7 +456,7 @@ struct LoraModel : public GGMLRunner {
             scale_value *= multiplier;
 
             auto curr_updown = ggml_ext_kronecker(ctx, lokr_w1, lokr_w2);
-            curr_updown      = ggml_scale_inplace(ctx, curr_updown, scale_value);
+            curr_updown      = ggml_ext_scale(ctx, curr_updown, scale_value, true);
 
             if (updown == nullptr) {
                 updown = curr_updown;
@@ -634,7 +634,7 @@ struct LoraModel : public GGMLRunner {
                                       forward_params.conv2d.scale);
             }
 
-            auto curr_out_diff = ggml_scale_inplace(ctx, lx, scale_value);
+            auto curr_out_diff = ggml_ext_scale(ctx, lx, scale_value, true);
 
             if (out_diff == nullptr) {
                 out_diff = curr_out_diff;
diff --git a/mmdit.hpp b/mmdit.hpp
index cb4be7b..e2636e7 100644
--- a/mmdit.hpp
+++ b/mmdit.hpp
@@ -33,7 +33,7 @@ public:
         auto fc2 = std::dynamic_pointer_cast<Linear>(blocks["fc2"]);
 
         x = fc1->forward(ctx, x);
-        x = ggml_gelu_inplace(ctx->ggml_ctx, x);
+        x = ggml_ext_gelu(ctx->ggml_ctx, x, true);
         x = fc2->forward(ctx, x);
         return x;
     }
diff --git a/pmid.hpp b/pmid.hpp
index 5b70dab..8ce78d3 100644
--- a/pmid.hpp
+++ b/pmid.hpp
@@ -33,7 +33,7 @@ public:
         x = layer_norm->forward(ctx, x);
         // x = ggml_add(ctx, ggml_mul_mat(ctx, fc1_w, x),  fc1_b);
         x = fc1->forward(ctx, x);
-        x = ggml_gelu_inplace(ctx->ggml_ctx, x);
+        x = ggml_ext_gelu(ctx->ggml_ctx, x, true);
         x = fc2->forward(ctx, x);
         // x = ggml_add(ctx, ggml_mul_mat(ctx, fc2_w, x),  fc2_b);
         if (use_residue)
@@ -129,8 +129,8 @@ public:
         k             = reshape_tensor(ctx->ggml_ctx, k, heads);
         v             = reshape_tensor(ctx->ggml_ctx, v, heads);
         scale         = 1.f / sqrt(sqrt((float)dim_head));
-        k             = ggml_scale_inplace(ctx->ggml_ctx, k, scale);
-        q             = ggml_scale_inplace(ctx->ggml_ctx, q, scale);
+        k             = ggml_ext_scale(ctx->ggml_ctx, k, scale, true);
+        q             = ggml_ext_scale(ctx->ggml_ctx, q, scale, true);
         // auto weight = ggml_mul_mat(ctx, q, k);
         auto weight = ggml_mul_mat(ctx->ggml_ctx, k, q);  // NOTE order of mul is opposite to pytorch
 
diff --git a/t5.hpp b/t5.hpp
index 2e3e4b5..fdac347 100644
--- a/t5.hpp
+++ b/t5.hpp
@@ -515,7 +515,7 @@ public:
         auto wi_1 = std::dynamic_pointer_cast<Linear>(blocks["wi_1"]);
         auto wo   = std::dynamic_pointer_cast<Linear>(blocks["wo"]);
 
-        auto hidden_gelu   = ggml_gelu_inplace(ctx->ggml_ctx, wi_0->forward(ctx, x));
+        auto hidden_gelu   = ggml_ext_gelu(ctx->ggml_ctx, wi_0->forward(ctx, x), true);
         auto hidden_linear = wi_1->forward(ctx, x);
         x                  = ggml_mul_inplace(ctx->ggml_ctx, hidden_gelu, hidden_linear);
         x                  = wo->forward(ctx, x);
@@ -608,7 +608,7 @@ public:
             }
         }
 
-        k = ggml_scale_inplace(ctx->ggml_ctx, k, ::sqrtf(static_cast<float>(d_head)));
+        k = ggml_ext_scale(ctx->ggml_ctx, k, ::sqrtf(static_cast<float>(d_head)), true);
 
         x = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, num_heads, mask);  // [N, n_token, d_head * n_head]
 
diff --git a/tae.hpp b/tae.hpp
index a22db19..8315257 100644
--- a/tae.hpp
+++ b/tae.hpp
@@ -161,9 +161,9 @@ public:
         // z: [n, z_channels, h, w]
         // return: [n, out_channels, h*8, w*8]
 
-        auto h = ggml_scale(ctx->ggml_ctx, z, 1.0f / 3.0f);
+        auto h = ggml_ext_scale(ctx->ggml_ctx, z, 1.0f / 3.0f);
         h      = ggml_tanh_inplace(ctx->ggml_ctx, h);
-        h      = ggml_scale(ctx->ggml_ctx, h, 3.0f);
+        h      = ggml_ext_scale(ctx->ggml_ctx, h, 3.0f);
 
         for (int i = 0; i < num_blocks * 3 + 10; i++) {
             if (blocks.find(std::to_string(i)) == blocks.end()) {
@@ -400,10 +400,11 @@ public:
         auto first_conv = std::dynamic_pointer_cast<Conv2d>(blocks["1"]);
 
         // Clamp()
-        auto h = ggml_scale_inplace(ctx->ggml_ctx,
-                                    ggml_tanh_inplace(ctx->ggml_ctx,
-                                                      ggml_scale(ctx->ggml_ctx, z, 1.0f / 3.0f)),
-                                    3.0f);
+        auto h = ggml_ext_scale(ctx->ggml_ctx,
+                                ggml_tanh_inplace(ctx->ggml_ctx,
+                                                  ggml_ext_scale(ctx->ggml_ctx, z, 1.0f / 3.0f)),
+                                3.0f,
+                                true);
 
         h         = first_conv->forward(ctx, h);
         h         = ggml_relu_inplace(ctx->ggml_ctx, h);
diff --git a/unet.hpp b/unet.hpp
index 6e15e1f..2dd79e0 100644
--- a/unet.hpp
+++ b/unet.hpp
@@ -529,7 +529,7 @@ public:
             }
         }
         if (controls.size() > 0) {
-            auto cs = ggml_scale_inplace(ctx->ggml_ctx, controls[controls.size() - 1], control_strength);
+            auto cs = ggml_ext_scale(ctx->ggml_ctx, controls[controls.size() - 1], control_strength, true);
             h       = ggml_add(ctx->ggml_ctx, h, cs);  // middle control
         }
         int control_offset = static_cast<int>(controls.size() - 2);
@@ -542,7 +542,7 @@ public:
                 hs.pop_back();
 
                 if (controls.size() > 0) {
-                    auto cs = ggml_scale_inplace(ctx->ggml_ctx, controls[control_offset], control_strength);
+                    auto cs = ggml_ext_scale(ctx->ggml_ctx, controls[control_offset], control_strength, true);
                     h_skip  = ggml_add(ctx->ggml_ctx, h_skip, cs);  // control net condition
                     control_offset--;
                 }
diff --git a/vae.hpp b/vae.hpp
index 2325002..fdddc8a 100644
--- a/vae.hpp
+++ b/vae.hpp
@@ -253,8 +253,8 @@ public:
 
         float alpha = get_alpha();
         x           = ggml_add(ctx->ggml_ctx,
-                               ggml_scale(ctx->ggml_ctx, x, alpha),
-                               ggml_scale(ctx->ggml_ctx, x_mix, 1.0f - alpha));
+                               ggml_ext_scale(ctx->ggml_ctx, x, alpha),
+                               ggml_ext_scale(ctx->ggml_ctx, x_mix, 1.0f - alpha));
 
         x = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 0, 2, 1, 3));  // b c t (h w) -> b t c (h w)
         x = ggml_reshape_4d(ctx->ggml_ctx, x, W, H, C, T * B);                     // b t c (h w) -> (b t) c h w
diff --git a/wan.hpp b/wan.hpp
index 3ade14b..1f1b2cf 100644
--- a/wan.hpp
+++ b/wan.hpp
@@ -1576,7 +1576,7 @@ namespace WAN {
             y = modulate_add(ctx->ggml_ctx, y, es[3]);
 
             y = ffn_0->forward(ctx, y);
-            y = ggml_gelu_inplace(ctx->ggml_ctx, y);
+            y = ggml_ext_gelu(ctx->ggml_ctx, y, true);
             y = ffn_2->forward(ctx, y);
 
             x = ggml_add(ctx->ggml_ctx, x, modulate_mul(ctx->ggml_ctx, y, es[5]));
@@ -1723,7 +1723,7 @@ namespace WAN {
 
             auto x = proj_0->forward(ctx, image_embeds);
             x      = proj_1->forward(ctx, x);
-            x      = ggml_gelu_inplace(ctx->ggml_ctx, x);
+            x      = ggml_ext_gelu(ctx->ggml_ctx, x, true);
             x      = proj_3->forward(ctx, x);
             x      = proj_4->forward(ctx, x);
 
@@ -1910,7 +1910,7 @@ namespace WAN {
             e0      = ggml_reshape_4d(ctx->ggml_ctx, e0, e0->ne[0] / 6, 6, e0->ne[1], e0->ne[2]);  //  [N, 6, dim] or [N, T, 6, dim]
 
             context = text_embedding_0->forward(ctx, context);
-            context = ggml_gelu(ctx->ggml_ctx, context);
+            context = ggml_ext_gelu(ctx->ggml_ctx, context);
             context = text_embedding_2->forward(ctx, context);  // [N, context_txt_len, dim]
 
             int64_t context_img_len = 0;
@@ -1949,7 +1949,7 @@ namespace WAN {
                     auto result = vace_block->forward(ctx, c, x_orig, e0, pe, context, context_img_len);
                     auto c_skip = result.first;
                     c           = result.second;
-                    c_skip      = ggml_scale(ctx->ggml_ctx, c_skip, vace_strength);
+                    c_skip      = ggml_ext_scale(ctx->ggml_ctx, c_skip, vace_strength);
                     x           = ggml_add(ctx->ggml_ctx, x, c_skip);
                 }
             }
diff --git a/z_image.hpp b/z_image.hpp
index 505fa7e..cee2383 100644
--- a/z_image.hpp
+++ b/z_image.hpp
@@ -517,7 +517,7 @@ namespace ZImage {
             out = ggml_ext_slice(ctx->ggml_ctx, out, 1, 0, H);  // [N, C, H, W + pad_w]
             out = ggml_ext_slice(ctx->ggml_ctx, out, 0, 0, W);  // [N, C, H, W]
 
-            out = ggml_scale(ctx->ggml_ctx, out, -1.f);
+            out = ggml_ext_scale(ctx->ggml_ctx, out, -1.f);
 
             return out;
         }