Cleanup readme, llava examples, usage examples and nccl init (#1194)

python/sglang/srt/managers/schedule_batch.py

@@ -35,7 +35,7 @@ INIT_INCREMENTAL_DETOKENIZATION_OFFSET = 5
 global_server_args_dict = {
     "disable_flashinfer": False,
     "disable_flashinfer_sampling": False,
-    "attention_reduce_in_fp32": False,
+    "triton_attention_reduce_in_fp32": False,
     "enable_mla": False,
 }
 

python/sglang/srt/managers/tokenizer_manager.py

@@ -606,6 +606,9 @@ class TokenizerManager:
         return background_tasks
 
     def create_handle_loop(self):
+        if not self.to_create_loop:
+            return
+
         self.to_create_loop = False
         loop = asyncio.get_event_loop()
         loop.create_task(self.handle_loop())