make github CI happy

ggml-qnn.cpp

@@ -89,6 +89,19 @@
 class qnn_instance;
 
 //TODO: should be removed because this is a workaround method during development stage
+//a minor modification is required during development stage for validate QNN backend on Android phone:
+//
+//modify from
+//
+//static void ggml_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor)
+//
+//to
+//
+//void ggml_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor)
+//
+//in source file ggml.c#L16156
+//
+//this workaround will not be needed when the final QNN backend is complete
 extern "C" void ggml_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor);
 
 #if (defined __ANDROID__) || (defined ANDROID) //Qualcomm's QNN could running on Windows over ARM(aka WoA)

ggml-qnn.h

@@ -46,10 +46,6 @@ GGML_API void           ggml_backend_qnn_get_device_description(int device, char
 GGML_API ggml_backend_buffer_type_t ggml_backend_qnn_buffer_type(size_t dev_num);
 
 
-//temporary API, should be removed in the future
-GGML_API bool           ggml_qnn_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor);
-
-
 #ifdef __cplusplus
 }
 #endif

ggml.c

@@ -16153,8 +16153,7 @@ static void ggml_compute_forward_cross_entropy_loss_back(
 
 /////////////////////////////////
 
-//workaround for Qualcomm QNN backend
-void ggml_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor) {
+static void ggml_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor) {
     GGML_ASSERT(params);
 
     if (tensor->op == GGML_OP_NONE || ggml_is_empty(tensor)) {

llama.cpp

@@ -15403,7 +15403,7 @@ struct llama_context * llama_new_context_with_model(
 #elif defined(GGML_USE_QNN)
         if (model->n_gpu_layers > 0) {
             //the second param is package name of Andorid app, can be got by JNI from Java layer
-            ggml_backend_t backend = ggml_backend_qnn_init(QNN_CPU, "/data/data/com.ggml.llamacpp/");
+            ggml_backend_t backend = ggml_backend_qnn_init(model->main_gpu, "/data/data/com.ggml.llamacpp/");
             if (nullptr == backend) {
                 LLAMA_LOG_ERROR("%s: failed to initialize QNN backend\n", __func__);
                 llama_free(ctx);
@@ -17577,14 +17577,6 @@ void llama_reset_timings(struct llama_context * ctx) {
     ctx->t_p_eval_us = ctx->n_p_eval = 0;
 }
 
-static int llama_has_qnn(void) {
-#ifdef GGML_USE_QNN
-    return 1;
-#else
-    return 0;
-#endif
-}
-
 const char * llama_print_system_info(void) {
     static std::string s;
 
@@ -17606,7 +17598,6 @@ const char * llama_print_system_info(void) {
     s += "SSSE3 = "       + std::to_string(ggml_cpu_has_ssse3())       + " | ";
     s += "VSX = "         + std::to_string(ggml_cpu_has_vsx())         + " | ";
     s += "MATMUL_INT8 = " + std::to_string(ggml_cpu_has_matmul_int8()) + " | ";
-    s += "QNN = "         + std::to_string(llama_has_qnn())            + " | ";
 
     return s.c_str();
 }