set disable custom gather by default (#908)

Author: shihaobai

lightllm/common/basemodel/basemodel.py

@@ -169,7 +169,7 @@ def _init_kv_move_buffer(self):
 
     def _check_mem_size(self):
         self.max_total_token_num = self.mem_manager.size
-        assert self.max_seq_length < self.max_total_token_num
+        assert self.max_seq_length <= self.max_total_token_num
         return
 
     def _init_req_manager(self):

lightllm/distributed/communication_op.py

@@ -79,7 +79,7 @@ def init_custom_gather(self) -> None:
             return
 
         args = get_env_start_args()
-        if args.disable_custom_allgather:
+        if not args.enable_custom_allgather:
             return
 
         cpu_group = create_new_group_for_current_dp("gloo")

lightllm/distributed/custom_all_gather.py

@@ -31,8 +31,11 @@
 from lightllm.common.basemodel.layer_infer.cache_tensor_manager import g_cache_manager
 
 
-if light_ops is not None:
-    light_ops.meta_size()
+try:
+    if light_ops is not None:
+        light_ops.meta_size()
+except:
+    pass
 
 logger = init_logger(__name__)
 

lightllm/server/api_cli.py

@@ -234,7 +234,7 @@ def make_argument_parser() -> argparse.ArgumentParser:
         "--enable_mps", action="store_true", help="Whether to enable nvidia mps for multimodal service."
     )
     parser.add_argument("--disable_custom_allreduce", action="store_true", help="Whether to disable cutom allreduce.")
-    parser.add_argument("--disable_custom_allgather", action="store_true", help="Whether to enable cutom allgather.")
+    parser.add_argument("--enable_custom_allgather", action="store_true", help="Whether to enable cutom allgather.")
     parser.add_argument(
         "--enable_tpsp_mix_mode",
         action="store_true",