modify q3tts loader

otherarch/qwen3tts/audio_tokenizer_decoder.cpp

@@ -317,7 +317,7 @@ bool AudioTokenizerDecoder::load_model(const std::string & model_path) {
 
     if (!load_tensor_data_from_file(model_path, gguf_ctx, model_.ctx,
                                      model_.tensors, model_.buffer, error_msg_,
-                                     GGML_BACKEND_DEVICE_TYPE_IGPU)) {
+                                     qwen3tts_allowgpu)) {
         return false;
     }
 
@@ -328,16 +328,6 @@ bool AudioTokenizerDecoder::load_model(const std::string & model_path) {
     }
 
     normalize_codebooks();
-    // Codebooks are normalized in host memory; sync once to backend tensors.
-    auto upload_if_present = [](struct ggml_tensor * t) {
-        if (t && t->data) {
-            ggml_backend_tensor_set(t, t->data, 0, ggml_nbytes(t));
-        }
-    };
-    upload_if_present(model_.vq_first_codebook);
-    for (int i = 0; i < 15; ++i) {
-        upload_if_present(model_.vq_rest_codebook[i]);
-    }
 
     state_.backend = init_preferred_backend("AudioTokenizerDecoder", &error_msg_, qwen3tts_allowgpu);
     if (!state_.backend) {

otherarch/qwen3tts/audio_tokenizer_encoder.cpp

@@ -249,7 +249,7 @@ bool AudioTokenizerEncoder::load_model(const std::string & model_path) {
     }
 
     if (!load_tensor_data_from_file(model_path, gguf_ctx, model_.ctx,
-                                     model_.tensors, model_.buffer, error_msg_, qwen3tts_allowgpu?GGML_BACKEND_DEVICE_TYPE_GPU:GGML_BACKEND_DEVICE_TYPE_CPU)) {
+                                     model_.tensors, model_.buffer, error_msg_, qwen3tts_allowgpu)) {
         return false;
     }
 

otherarch/qwen3tts/gguf_loader.cpp

@@ -161,12 +161,10 @@ bool load_tensor_data_from_file(
     const std::map<std::string, struct ggml_tensor *> & tensors,
     ggml_backend_buffer_t & buffer,
     std::string & error_msg,
-    enum ggml_backend_dev_type preferred_backend_type
+    bool allowgpu
 ) {
-    ggml_backend_t backend = ggml_backend_init_by_type(preferred_backend_type, nullptr);
-    if (!backend && preferred_backend_type != GGML_BACKEND_DEVICE_TYPE_CPU) {
-        backend = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr);
-    }
+    ggml_backend_t backend = init_preferred_backend("TensorLoader", &error_msg, allowgpu);
+
     if (!backend) {
         error_msg = "Failed to initialize backend for GGUF tensor loader";
         return false;
@@ -176,7 +174,7 @@ bool load_tensor_data_from_file(
     buffer = ggml_backend_alloc_ctx_tensors(model_ctx, backend);
     if (!buffer) {
         error_msg = "Failed to allocate tensor buffer";
-        ggml_backend_free(backend);
+        release_preferred_backend(backend);
         return false;
     }
 
@@ -184,7 +182,7 @@ bool load_tensor_data_from_file(
     FILE * f = fopen(path.c_str(), "rb");
     if (!f) {
         error_msg = "Failed to open file for reading: " + path;
-        ggml_backend_free(backend);
+        release_preferred_backend(backend);
         return false;
     }
 
@@ -209,14 +207,14 @@ bool load_tensor_data_from_file(
         if (fseek(f, data_offset + offset, SEEK_SET) != 0) {
             error_msg = "Failed to seek to tensor data: " + std::string(name);
             fclose(f);
-            ggml_backend_free(backend);
+            release_preferred_backend(backend);
             return false;
         }
 
         if (fread(read_buf.data(), 1, nbytes, f) != nbytes) {
             error_msg = "Failed to read tensor data: " + std::string(name);
             fclose(f);
-            ggml_backend_free(backend);
+            release_preferred_backend(backend);
             return false;
         }
 
@@ -224,7 +222,7 @@ bool load_tensor_data_from_file(
     }
 
     fclose(f);
-    ggml_backend_free(backend);
+    release_preferred_backend(backend);
 
     return true;
 }

otherarch/qwen3tts/gguf_loader.h

@@ -72,7 +72,7 @@ bool load_tensor_data_from_file(
     const std::map<std::string, struct ggml_tensor *> & tensors,
     ggml_backend_buffer_t & buffer,
     std::string & error_msg,
-    enum ggml_backend_dev_type preferred_backend_type = GGML_BACKEND_DEVICE_TYPE_GPU
+    bool allowgpu
 );
 
 // Helper to initialize backend with GPU preference and CPU fallback