refactor: allow specifying output path in lora conversion subprocess

- remove the shutil.move operation
- allow specifying the output path in arguments in convert_lora_to_ggml.py utility
- bump max number of LoRA layers

src/AutoGGUF.py

@@ -1099,21 +1099,10 @@ def restart_task(self, task_item) -> None:
                 task_item.update_status(IN_PROGRESS)
                 break
 
-    def lora_conversion_finished(self, thread, input_path, output_path) -> None:
+    def lora_conversion_finished(self, thread) -> None:
         self.logger.info(LORA_CONVERSION_FINISHED)
         if thread in self.quant_threads:
             self.quant_threads.remove(thread)
-        try:
-            # Only move the file if the output type is GGML
-            if self.lora_output_type_combo.currentText() == "GGML":
-                source_file = os.path.join(input_path, "ggml-adapter-model.bin")
-                if os.path.exists(source_file):
-                    shutil.move(source_file, output_path)
-                    self.logger.info(LORA_FILE_MOVED.format(source_file, output_path))
-                else:
-                    self.logger.warning(LORA_FILE_NOT_FOUND.format(source_file))
-        except Exception as e:
-            self.logger.error(ERROR_MOVING_LORA_FILE.format(str(e)))
 
     def download_finished(self, extract_dir) -> None:
         self.logger.info(DOWNLOAD_FINISHED_EXTRACTED_TO.format(extract_dir))

src/convert_lora_to_ggml.py

@@ -58,7 +58,7 @@ def pyinstaller_include():
 
 if __name__ == "__main__":
     if len(sys.argv) < 2:
-        logger.info(f"Usage: python {sys.argv[0]} <path> [arch]")
+        logger.info(f"Usage: python {sys.argv[0]} <path> <output_path> [arch]")
         logger.info(
             "Path must contain HuggingFace PEFT LoRA files 'adapter_config.json' and 'adapter_model.bin'"
         )
@@ -69,7 +69,7 @@ def pyinstaller_include():
 
     input_json = os.path.join(sys.argv[1], "adapter_config.json")
     input_model = os.path.join(sys.argv[1], "adapter_model.bin")
-    output_path = os.path.join(sys.argv[1], "ggml-adapter-model.bin")
+    output_path = sys.argv[2]
 
     if os.path.exists(input_model):
         model = torch.load(input_model, map_location="cpu")
@@ -80,7 +80,7 @@ def pyinstaller_include():
 
         model = load_file(input_model, device="cpu")
 
-    arch_name = sys.argv[2] if len(sys.argv) == 3 else "llama"
+    arch_name = sys.argv[3] if len(sys.argv) == 4 else "llama"
 
     if arch_name not in gguf.MODEL_ARCH_NAMES.values():
         logger.error(f"Error: unsupported architecture {arch_name}")
@@ -89,7 +89,7 @@ def pyinstaller_include():
     arch = list(gguf.MODEL_ARCH_NAMES.keys())[
         list(gguf.MODEL_ARCH_NAMES.values()).index(arch_name)
     ]
-    name_map = gguf.TensorNameMap(arch, 200)  # 200 layers ought to be enough for anyone
+    name_map = gguf.TensorNameMap(arch, 500)
 
     with open(input_json, "r") as f:
         params = json.load(f)

src/lora_conversion.py

@@ -179,7 +179,12 @@ def convert_lora(self) -> None:
                 raise ValueError(BASE_MODEL_PATH_REQUIRED)
             command.extend(["--base", base_model_path])
         else:  # Use old GGML parameters for GGML
-            command = ["python", "src/convert_lora_to_ggml.py", lora_input_path]
+            command = [
+                "python",
+                "src/convert_lora_to_ggml.py",
+                lora_input_path,
+                lora_output_path,
+            ]
 
         logs_path = self.logs_input.text()
         ensure_directory(logs_path)
@@ -203,11 +208,7 @@ def convert_lora(self) -> None:
         self.task_list.setItemWidget(list_item, task_item)
 
         thread.status_signal.connect(task_item.update_status)
-        thread.finished_signal.connect(
+        thread.finished_signal.connect(lambda: self.lora_conversion_finished(thread))
-            lambda: self.lora_conversion_finished(
-                thread, lora_input_path, lora_output_path
-            )
-        )
         thread.error_signal.connect(
             lambda err: handle_error(self.logger, err, task_item)
         )