Clean up noisy startup warnings from third-party deps (#23669)

python/sglang/srt/multimodal/processors/base_processor.py

@@ -10,7 +10,7 @@ from typing import Any, Dict, Iterator, List, Optional, Tuple, Union
 import numpy as np
 import torch
 from PIL import Image
-from transformers import BaseImageProcessorFast
+from transformers import BaseImageProcessor
 
 from sglang.srt.managers.schedule_batch import (
     Modality,
@@ -428,7 +428,7 @@ class BaseMultimodalProcessor(ABC):
         processor = self._processor
         if (
             hasattr(processor, "image_processor")
-            and isinstance(processor.image_processor, BaseImageProcessorFast)
+            and isinstance(processor.image_processor, BaseImageProcessor)
             and not self.server_args.disable_fast_image_processor
         ):
             if _is_cpu or get_global_server_args().rl_on_policy_target is not None:

python/sglang/srt/multimodal/processors/ernie45_vl.py

@@ -7,7 +7,7 @@ import torch
 import torchvision
 from PIL import Image
 from torchvision.transforms import InterpolationMode
-from transformers import BaseImageProcessorFast
+from transformers import BaseImageProcessor
 
 from sglang.srt.environ import envs
 from sglang.srt.layers.rotary_embedding import MRotaryEmbedding
@@ -302,7 +302,7 @@ class Ernie4_5_VLImageProcessor(SGLangBaseProcessor):
         processor = self._processor
         if (
             hasattr(processor, "image_processor")
-            and isinstance(processor.image_processor, BaseImageProcessorFast)
+            and isinstance(processor.image_processor, BaseImageProcessor)
             and not self.server_args.disable_fast_image_processor
         ):
             if not _is_npu:

python/sglang/srt/multimodal/processors/kimi_k25.py

@@ -233,7 +233,7 @@ class KimiGPUProcessorWrapper:
         self._gpu_norm_tensors = None
 
         # Explicitly expose attributes that base class process_mm_data needs:
-        # - image_processor: checked via isinstance(..., BaseImageProcessorFast)
+        # - image_processor: checked via isinstance(..., BaseImageProcessor)
         # - tokenizer: used for tokenization
         # - media_processor: used by CPU fallback path
         self.image_processor = hf_processor.image_processor

python/sglang/srt/platforms/__init__.py

@@ -90,7 +90,7 @@ def _resolve_platform() -> SRTPlatform:
             logger.exception("Failed to activate platform plugin: %s", name)
 
     if len(activated) == 0:
-        logger.warning("No platform detected. Using base SRTPlatform with defaults.")
+        logger.debug("No platform detected. Using base SRTPlatform with defaults.")
         return SRTPlatform()
 
     if len(activated) == 1:

python/sglang/srt/utils/hf_transformers/common.py

@@ -217,16 +217,13 @@ def get_hf_text_config(config: PretrainedConfig):
 
     # Some models (e.g. DeepSeek-OCR) store sub-configs as plain dicts.
     # Convert to PretrainedConfig early so hasattr() checks and asserts work.
-    parent_dtype = getattr(config, "torch_dtype", None)
+    parent_dtype = getattr(config, "dtype", None)
     for _attr in ("text_config", "llm_config", "language_config", "thinker_config"):
         _sub = getattr(config, _attr, None)
         if isinstance(_sub, dict):
             _converted = PretrainedConfig(**_sub)
-            if (
-                getattr(_converted, "torch_dtype", None) is None
-                and parent_dtype is not None
-            ):
-                _converted.torch_dtype = parent_dtype
+            if getattr(_converted, "dtype", None) is None and parent_dtype is not None:
+                _converted.dtype = parent_dtype
             setattr(config, _attr, _converted)
 
     # Priority: thinker_config > llm_config > language_config > text_config
@@ -236,8 +233,8 @@ def get_hf_text_config(config: PretrainedConfig):
         if hasattr(thinker_config, "text_config"):
             setattr(
                 thinker_config.text_config,
-                "torch_dtype",
-                getattr(thinker_config, "torch_dtype", None),
+                "dtype",
+                getattr(thinker_config, "dtype", None),
             )
             text_config = thinker_config.text_config
         else:

python/sglang/srt/utils/hf_transformers_patches.py

@@ -222,7 +222,19 @@ def _patch_removed_symbols():
     """
     # LlamaFlashAttention2
     try:
-        from transformers.models.llama import modeling_llama
+        import logging
+
+        # Importing modeling_llama triggers a deep import chain:
+        #   modeling_llama -> modeling_utils -> quantizers -> torchao
+        # torchao emits a noisy warning about incompatible torch versions
+        # that is irrelevant here — suppress it during this import.
+        _torchao_logger = logging.getLogger("torchao")
+        _prev_level = _torchao_logger.level
+        _torchao_logger.setLevel(logging.ERROR)
+        try:
+            from transformers.models.llama import modeling_llama
+        finally:
+            _torchao_logger.setLevel(_prev_level)
 
         if not hasattr(modeling_llama, "LlamaFlashAttention2"):
             if hasattr(modeling_llama, "LlamaAttention"):