add llm_graph_input_one

Author: ngxson

src/llama-graph.cpp

@@ -350,6 +350,12 @@ void llm_graph_input_mem_hybrid::set_input(const llama_ubatch * ubatch) {
     }
 }
 
+void llm_graph_input_one::set_input(const llama_ubatch *) {
+    GGML_ASSERT(one && ggml_nelements(one) == 1);
+    float f_one = 1.0f;
+    ggml_backend_tensor_set(one, &f_one, 0, sizeof(float));
+}
+
 //
 // llm_graph_context
 //

src/llama-graph.h

@@ -329,6 +329,17 @@ class llm_graph_input_mem_hybrid : public llm_graph_input_i {
     const llama_memory_hybrid_context * mctx;
 };
 
+// TODO: remove this when ggml_scale_add is implemented
+class llm_graph_input_one : public llm_graph_input_i {
+public:
+    llm_graph_input_one() {}
+    virtual ~llm_graph_input_one() = default;
+
+    void set_input(const llama_ubatch *) override;
+
+    ggml_tensor * one = nullptr; // F32
+};
+
 //
 // llm_graph_result
 //

src/llama-model.cpp

@@ -9081,8 +9081,13 @@ struct llm_build_gemma3n_iswa : public llm_graph_context {
         ggml_tensor * cur;
         ggml_tensor * inpL;
 
+        // TODO: remove this when ggml_scale_add is implemented
         one = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1);
-        one = ggml_cos(ctx0, ggml_scale(ctx0, one, 0.0f)); // cos(0.0f) = 1.0f
+        {
+            auto inp = std::make_unique<llm_graph_input_one>();
+            inp->one = one;
+            res->add_input(std::move(inp));
+        }
 
         inpL = build_inp_embd(model.tok_embd);