fix: parallel_download_decorator compat with transformers >= 5, use dtype instead of torch_dtype

Author: wejoncy

qllm/modeling/base.py

@@ -128,7 +128,12 @@ def cached_file_func_in_thread(task_func, *args, **kwargs):
             return executor.submit(task_func, *args, **kwargs)
         transformers.utils.hub.cached_file = functools.partial(cached_file_func_in_thread, transformers.utils.hub.cached_file)
         result = task_func_shard(*args, **kwargs)
-        result_0 = [future.result() for future in result[0]]
+        result_0 = []
+        for item in result[0]:
+            if isinstance(item, str):
+                result_0.append(item)
+            else:
+                result_0.append(item.result())
     return result_0, result[1]
 
 
@@ -229,7 +234,7 @@ def from_pretrained(
             )
             llm = AutoModelForCausalLM.from_pretrained(
                 pretrained_model_name_or_path,
-                torch_dtype=torch_dtype,
+                dtype=torch_dtype,
                 trust_remote_code=trust_remote_code,
                 attn_implementation=attn_implementation,
                 # device_map="auto",