Clean up noisy startup log messages and refactor loader.py (sgl-project#18531)

Author: merrymercy

python/sglang/launch_server.py

@@ -6,6 +6,9 @@
 
 from sglang.srt.server_args import prepare_server_args
 from sglang.srt.utils import kill_process_tree
+from sglang.srt.utils.common import suppress_noisy_warnings
+
+suppress_noisy_warnings()
 
 
 def run_server(server_args):

python/sglang/srt/entrypoints/openai/serving_chat.py

@@ -87,6 +87,8 @@ def _extract_max_dynamic_patch(request: ChatCompletionRequest):
 class OpenAIServingChat(OpenAIServingBase):
     """Handler for /v1/chat/completions requests"""
 
+    _default_sampling_params_logged = False
+
     def __init__(
         self,
         tokenizer_manager: TokenizerManager,
@@ -101,10 +103,14 @@ def __init__(
         self.default_sampling_params = (
             self.tokenizer_manager.model_config.get_default_sampling_params()
         )
-        if self.default_sampling_params:
+        if (
+            self.default_sampling_params
+            and not OpenAIServingChat._default_sampling_params_logged
+        ):
             logger.info(
                 f"Using default chat sampling params from model generation config: {self.default_sampling_params}",
             )
+            OpenAIServingChat._default_sampling_params_logged = True
 
         # Check if the model is a GPT-OSS model
         self.is_gpt_oss = (

python/sglang/srt/layers/moe/utils.py

@@ -10,7 +10,6 @@
     get_attention_dp_size,
     is_dp_attention_enabled,
 )
-from sglang.srt.utils import log_info_on_rank0
 
 if TYPE_CHECKING:
     from sglang.srt.server_args import ServerArgs
@@ -189,10 +188,6 @@ def get_moe_a2a_backend() -> MoeA2ABackend:
 def get_moe_runner_backend() -> MoeRunnerBackend:
     global MOE_RUNNER_BACKEND
     if MOE_RUNNER_BACKEND is None:
-        log_info_on_rank0(
-            logger,
-            "MOE_RUNNER_BACKEND is not initialized, the backend will be automatically selected",
-        )
         MOE_RUNNER_BACKEND = MoeRunnerBackend.AUTO
     return MOE_RUNNER_BACKEND
 

python/sglang/srt/mem_cache/mamba_radix_cache.py

@@ -384,8 +384,6 @@ def __init__(self, params: CacheInitParams):
             assert (
                 self.page_size == 1
             ), f"Page size must be 1 for MambaRadixCache v1, got {self.page_size}"
-        else:
-            logger.info(f"Mamba extra_buffer is enabled.")
 
         if self.token_to_kv_pool_allocator:
             self.device = self.token_to_kv_pool_allocator.device

python/sglang/srt/model_executor/model_runner.py

@@ -1684,8 +1684,6 @@ def init_cublas(self):
 
     def init_attention_backend(self):
         """Init attention kernel backend."""
-        tic = time.perf_counter()
-        logger.info("Init attention backend begin.")
         if self.server_args.enable_pdmux:
             self.attn_backend = self._get_attention_backend(init_new_workspace=True)
             self.decode_attn_backend_group = []
@@ -1696,9 +1694,6 @@ def init_attention_backend(self):
             self.attn_backend = TboAttnBackend.init_new(self._get_attention_backend)
         else:
             self.attn_backend = self._get_attention_backend()
-        logger.info(
-            f"Init attention backend end. elapsed={time.perf_counter() - tic:.2f} s"
-        )
 
     def _get_attention_backend(self, init_new_workspace: bool = False):
         """Init attention kernel backend."""

python/sglang/srt/model_loader/loader.py

@@ -12,6 +12,7 @@
 import logging
 import math
 import os
+import re
 import socket
 import threading
 import time
@@ -304,6 +305,8 @@ class DefaultModelLoader(BaseModelLoader):
     # default number of thread when enable multithread weight loading
     DEFAULT_NUM_THREADS = 8
 
+    _MTP_PATTERN = re.compile(r"model\.mtp\.layers\.(\d+)\.")
+
     @dataclasses.dataclass
     class Source:
         """A source for weights."""
@@ -351,11 +354,11 @@ def __init__(self, load_config: LoadConfig):
 
     def _maybe_download_from_modelscope(
         self, model: str, revision: Optional[str]
-    ) -> Optional[str]:
+    ) -> str:
         """Download model from ModelScope hub if SGLANG_USE_MODELSCOPE is True.
 
-        Returns the path to the downloaded model, or None if the model is not
-        downloaded from ModelScope."""
+        Returns the path to the downloaded model, or the original model path if
+        not downloaded from ModelScope."""
         if get_bool_env_var("SGLANG_USE_MODELSCOPE"):
             # download model from ModelScope hub,
             # lazy import so that modelscope is not required for normal use.
@@ -373,17 +376,16 @@ def _maybe_download_from_modelscope(
             else:
                 model_path = model
             return model_path
-        return None
+        return model
 
     def _prepare_weights(
         self, model_name_or_path: str, revision: Optional[str], fall_back_to_pt: bool
     ) -> Tuple[str, List[str], bool]:
         """Prepare weights for the model.
 
         If the model is not local, it will be downloaded."""
-        model_name_or_path = (
-            self._maybe_download_from_modelscope(model_name_or_path, revision)
-            or model_name_or_path
+        model_name_or_path = self._maybe_download_from_modelscope(
+            model_name_or_path, revision
         )
 
         is_local = os.path.isdir(model_name_or_path)
@@ -474,6 +476,7 @@ def _get_weights_iterator(
     ) -> Generator[Tuple[str, torch.Tensor], None, None]:
         """Get an iterator for the model weights based on the load format."""
         extra_config = self.load_config.model_loader_extra_config
+        use_multithread = extra_config.get("enable_multithread_load", False)
         hf_folder, hf_weights_files, use_safetensors = self._prepare_weights(
             source.model_or_path, source.revision, source.fall_back_to_pt
         )
@@ -504,7 +507,7 @@ def _get_weights_iterator(
                 weights_iterator = fastsafetensors_weights_iterator(
                     hf_weights_files,
                 )
-            elif extra_config.get("enable_multithread_load"):
+            elif use_multithread:
                 weights_iterator = multi_thread_safetensors_weights_iterator(
                     hf_weights_files,
                     max_workers=extra_config.get(
@@ -518,7 +521,7 @@ def _get_weights_iterator(
                 )
 
         else:
-            if extra_config.get("enable_multithread_load"):
+            if use_multithread:
                 weights_iterator = multi_thread_pt_weights_iterator(
                     hf_weights_files,
                     max_workers=extra_config.get(
@@ -529,28 +532,34 @@ def _get_weights_iterator(
                 weights_iterator = pt_weights_iterator(hf_weights_files)
 
         if self.load_config.draft_model_idx is not None:
-            import re
-
-            pattern = r"model.mtp.layers.(\d+)."
-            filtered_weights = []
-            for name, tensor in weights_iterator:
-                group = re.match(pattern, name)
-                if group is not None:
-                    idx = int(group.group(1))
-                    if idx != self.load_config.draft_model_idx:
-                        continue
-                    new_name = name.replace(group.group(), "model.mtp.layers.0.")
-                else:
-                    new_name = name
-                filtered_weights.append((source.prefix + new_name, tensor))
-            return tuple(filtered_weights)
+            return self._filter_mtp_weights(
+                weights_iterator, source.prefix, self.load_config.draft_model_idx
+            )
 
         if self.counter_before_loading_weights == 0.0:
-            logger.info("Beginning to load weights")
             self.counter_before_loading_weights = time.perf_counter()
         # Apply the prefix.
         return ((source.prefix + name, tensor) for (name, tensor) in weights_iterator)
 
+    @classmethod
+    def _filter_mtp_weights(
+        cls, weights_iterator, prefix: str, draft_model_idx: int
+    ) -> Tuple[Tuple[str, torch.Tensor], ...]:
+        """Filter MTP (Multi-Token Prediction) weights to keep only the
+        specified draft model layer and remap it to layer 0."""
+        filtered_weights = []
+        for name, tensor in weights_iterator:
+            match = cls._MTP_PATTERN.match(name)
+            if match is not None:
+                idx = int(match.group(1))
+                if idx != draft_model_idx:
+                    continue
+                new_name = name.replace(match.group(), "model.mtp.layers.0.")
+            else:
+                new_name = name
+            filtered_weights.append((prefix + new_name, tensor))
+        return tuple(filtered_weights)
+
     def _get_all_weights(
         self,
         model_config: ModelConfig,
@@ -670,10 +679,6 @@ def load_model(
             )
 
         self.counter_after_loading_weights = time.perf_counter()
-        logger.info(
-            "Loading weights took %.2f seconds",
-            self.counter_after_loading_weights - self.counter_before_loading_weights,
-        )
         return model.eval()
 
     @staticmethod

python/sglang/srt/utils/common.py

@@ -1057,10 +1057,25 @@ def uniform_sample(l, n):
     return frames
 
 
-def suppress_other_loggers():
+def suppress_noisy_warnings():
+    """Suppress known noisy warnings from third-party libraries."""
     warnings.filterwarnings(
         "ignore", category=UserWarning, message="The given NumPy array is not writable"
     )
+    warnings.filterwarnings(
+        "ignore",
+        message="The cuda.cudart module is deprecated",
+        category=FutureWarning,
+    )
+    warnings.filterwarnings(
+        "ignore",
+        message="The cuda.nvrtc module is deprecated",
+        category=FutureWarning,
+    )
+
+
+def suppress_other_loggers():
+    suppress_noisy_warnings()
 
     try:
         from vllm.logger import logger as vllm_default_logger