Remove cached triton launcher (#656)

python/sglang/srt/server.py

@@ -51,6 +51,7 @@ from sglang.srt.utils import (
     allocate_init_ports,
     assert_pkg_version,
     enable_show_time_cost,
+    set_ulimit,
 )
 from sglang.utils import get_exception_traceback
 
@@ -145,30 +146,6 @@ def _set_global_server_args(server_args: ServerArgs):
     }
 
 
-def _set_ulimit(target_soft_limit=65535):
-    import resource
-
-    resource_type = resource.RLIMIT_NOFILE
-    current_soft, current_hard = resource.getrlimit(resource_type)
-
-    if current_soft >= target_soft_limit:
-        logger.info(
-            f"Current limits are already sufficient: soft={current_soft}, hard={current_hard}"
-        )
-    else:
-        try:
-            resource.setrlimit(resource_type, (target_soft_limit, current_hard))
-            new_soft, new_hard = resource.getrlimit(resource_type)
-            logger.info(
-                f"Successfully set new limits: soft={new_soft}, hard={new_hard}"
-            )
-        except ValueError as e:
-            logger.warn(f"Failed to set new limits: {e}")
-            logger.info(
-                f"Limits remain unchanged: soft={current_soft}, hard={current_hard}"
-            )
-
-
 def launch_server(
     server_args: ServerArgs,
     model_overide_args: Optional[dict] = None,
@@ -186,7 +163,7 @@ def launch_server(
     os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
     os.environ["NCCL_CUMEM_ENABLE"] = "0"
     os.environ["NCCL_NVLS_ENABLE"] = "0"
-    _set_ulimit()
+    set_ulimit()
     if server_args.show_time_cost:
         enable_show_time_cost()
     if server_args.disable_disk_cache: