lmdeploy 支持 dtype 配置

Author: shell-nlp

gpt_server/model_backend/lmdeploy_backend.py

@@ -29,14 +29,21 @@ def __init__(self, model_path) -> None:
         backend = backend_map[os.getenv("backend")]
         enable_prefix_caching = bool(os.getenv("enable_prefix_caching", False))
         max_model_len = os.getenv("max_model_len", None)
+        dtype = os.getenv("dtype", "auto")
         logger.info(f"后端 {backend}")
         if backend == "pytorch":
-            backend_config = PytorchEngineConfig(tp=int(os.getenv("num_gpus", "1")))
+            backend_config = PytorchEngineConfig(
+                tp=int(os.getenv("num_gpus", "1")),
+                dtype=dtype,
+                session_len=int(max_model_len) if max_model_len else None,
+                enable_prefix_caching=enable_prefix_caching,
+            )
         if backend == "turbomind":
             backend_config = TurbomindEngineConfig(
                 tp=int(os.getenv("num_gpus", "1")),
                 enable_prefix_caching=enable_prefix_caching,
                 session_len=int(max_model_len) if max_model_len else None,
+                dtype=dtype,
             )
         pipeline_type, pipeline_class = get_task(model_path)
         logger.info(f"模型架构：{pipeline_type}")

gpt_server/model_backend/vllm_backend.py

@@ -27,9 +27,7 @@ class VllmBackend(ModelBackend):
     def __init__(self, model_path) -> None:
         lora = os.getenv("lora", None)
         enable_prefix_caching = bool(os.getenv("enable_prefix_caching", False))
-
         max_model_len = os.getenv("max_model_len", None)
-
         tensor_parallel_size = int(os.getenv("num_gpus", "1"))
         dtype = os.getenv("dtype", "auto")
         max_loras = 1