Clean up server_args, triton cache manager (#8332)

python/sglang/srt/managers/scheduler.py

@@ -653,6 +653,9 @@ class Scheduler(
             )
         )
 
+        embedding_cache_size = int(os.environ.get("SGLANG_VLM_CACHE_SIZE_MB", "100"))
+        init_embedding_cache(embedding_cache_size * 1024 * 1024)
+
     def init_profier(self):
         self.torch_profiler = None
         self.torch_profiler_output_dir: Optional[str] = None
@@ -2895,9 +2898,9 @@ def run_scheduler_process(
         prefix += f" PP{pp_rank}"
 
     # Config the process
-    kill_itself_when_parent_died()
     setproctitle.setproctitle(f"sglang::scheduler{prefix.replace(' ', '_')}")
     faulthandler.enable()
+    kill_itself_when_parent_died()
     parent_process = psutil.Process().parent()
 
     # [For Router] if env var "SGLANG_DP_RANK" exist, set dp_rank to the value of the env var
@@ -2912,10 +2915,6 @@ def run_scheduler_process(
     if get_bool_env_var("SGLANG_SET_CPU_AFFINITY"):
         set_gpu_proc_affinity(server_args.tp_size, server_args.nnodes, gpu_id)
 
-    embedding_cache_size = 100
-    if "SGLANG_VLM_CACHE_SIZE_MB" in os.environ:
-        embedding_cache_size = int(os.environ["SGLANG_VLM_CACHE_SIZE_MB"])
-    init_embedding_cache(embedding_cache_size * 1024 * 1024)
     # Create a scheduler and run the event loop
     try:
         scheduler = Scheduler(server_args, port_args, gpu_id, tp_rank, pp_rank, dp_rank)
@@ -2926,8 +2925,8 @@ def run_scheduler_process(
                 "max_req_input_len": scheduler.max_req_input_len,
             }
         )
-        disaggregation_mode: DisaggregationMode = scheduler.disaggregation_mode
 
+        disaggregation_mode: DisaggregationMode = scheduler.disaggregation_mode
         if disaggregation_mode == DisaggregationMode.NULL:
             if server_args.pp_size > 1:
                 scheduler.event_loop_pp()