qwen3 deepstack use layernorm

Author: LETS-BEE

tools/mtmd/clip.cpp

@@ -363,6 +363,7 @@ struct clip_model {
     // qwen3vl deepstack (multi-level feature fusion)
     struct deepstack_merger {
         ggml_tensor * norm_w = nullptr;
+        ggml_tensor * norm_b = nullptr;
         ggml_tensor * fc1_w = nullptr;
         ggml_tensor * fc1_b = nullptr;
         ggml_tensor * fc2_w = nullptr;
@@ -660,7 +661,7 @@ struct clip_graph {
         return gf;
     }
 
-    // Qwen2VL and Qwen2.5VL use M-RoPE
+    // Qwen2VL, Qwen2.5VL and Qwen3VL use M-RoPE
     ggml_cgraph * build_qwen2vl() {
         GGML_ASSERT(model.class_embedding == nullptr);
 
@@ -913,8 +914,7 @@ struct clip_graph {
                 LOG_INF("%s: DeepStack merger %zu weights: norm_w=[%lld], fc1_w=[%lld,%lld], fc2_w=[%lld,%lld]\n",
                         __func__, i, merger.norm_w->ne[0], merger.fc1_w->ne[0], merger.fc1_w->ne[1], merger.fc2_w->ne[0], merger.fc2_w->ne[1]);
 
-                feat = ggml_rms_norm(ctx0, feat, eps);
-                feat = ggml_mul(ctx0, feat, merger.norm_w);
+                feat = build_norm(feat, merger.norm_w, merger.norm_b, norm_t, eps, -1);
 
                 feat = ggml_mul_mat(ctx0, merger.fc1_w, feat);
                 feat = ggml_add(ctx0, feat, merger.fc1_b);
@@ -2851,6 +2851,7 @@ struct clip_model_loader {
                         for (size_t i = 0; i < hparams.deepstack_layers.size(); i++) {
                             auto & merger = model.deepstack_mergers[i];
                             merger.norm_w = get_tensor(string_format("v.deepstack.%d.norm.weight", (int)i), false);
+                            merger.norm_b = get_tensor(string_format("v.deepstack.%d.norm.bias", (int)i), false);
                             merger.fc1_w  = get_tensor(string_format("v.deepstack.%d.fc1.weight", (int)i), false);
                             merger.fc1_b  = get_tensor(string_format("v.deepstack.%d.fc1.bias", (int)i), false);
                             merger.fc2_w  = get_tensor(string_format("v.deepstack.%d.fc2.weight", (int)i), false);