Improve streaming, log_level, memory report, weight loading, and benchmark script (#7632)

Co-authored-by: Kan Wu <wukanustc@gmail.com>

python/sglang/srt/warmup.py

@@ -4,6 +4,7 @@ from typing import List
 import numpy as np
 import tqdm
 
+from sglang.srt.disaggregation.utils import FAKE_BOOTSTRAP_HOST
 from sglang.srt.managers.io_struct import GenerateReqInput
 from sglang.srt.managers.tokenizer_manager import TokenizerManager
 
@@ -20,17 +21,21 @@ def warmup(name: str) -> callable:
     return decorator
 
 
-async def execute_warmups(warmup_names: List[str], tokenizer_manager: TokenizerManager):
+async def execute_warmups(
+    disaggregation_mode: str,
+    warmup_names: List[str],
+    tokenizer_manager: TokenizerManager,
+):
     for warmup_name in warmup_names:
         if warmup_name not in _warmup_registry:
             logger.warning(f"Could not find custom warmup {warmup_name}")
             continue
         logger.info(f"Running warmup {warmup_name}")
-        await _warmup_registry[warmup_name](tokenizer_manager)
+        await _warmup_registry[warmup_name](disaggregation_mode, tokenizer_manager)
 
 
 @warmup("voice_chat")
-async def voice_chat(tokenizer_manager: TokenizerManager):
+async def voice_chat(disaggregation_mode: str, tokenizer_manager: TokenizerManager):
     # this warms up the fused_moe triton kernels and caches them
     # if we don't do this we break real time inference for voice chat
     for i in tqdm.trange(1, 512):
@@ -44,4 +49,8 @@ async def voice_chat(tokenizer_manager: TokenizerManager):
                 "min_p": 0.0,
             },
         )
+        if disaggregation_mode != "null":
+            generate_req_input.bootstrap_room = 0
+            generate_req_input.bootstrap_host = FAKE_BOOTSTRAP_HOST
+
         await tokenizer_manager.generate_request(generate_req_input, None).__anext__()