Clean logging under --weight-loader-prefetch-checkpoints (#33930)

Co-authored-by: Brayden Zhong <brayden@radixark.ai>
Co-authored-by: Mohammad Miadh Angkad <176301910+mmangkad@users.noreply.github.com>

python/sglang/multimodal_gen/runtime/layers/quantization/modelopt_quant.py

@@ -322,10 +322,7 @@ class ModelOptFp4Config(ModelOptQuantConfig):
         super().__init__(exclude_modules, packed_modules_mapping)
         self.is_checkpoint_nvfp4_serialized = is_checkpoint_nvfp4_serialized
         if is_checkpoint_nvfp4_serialized:
-            logger.warning(
-                "Detected nvfp4 checkpoint. Please note that the "
-                "format is experimental and subject to change."
-            )
+            logger.info("Detected nvfp4 checkpoint.")
         self.group_size = group_size
         self.checkpoint_uses_packed_qkv = checkpoint_uses_packed_qkv
         self.swap_weight_nibbles = swap_weight_nibbles

python/sglang/srt/layers/cp/cp_decode_attn_tp.py

@@ -60,7 +60,6 @@ class CpDecodeAttnTpContext:
         else:
             self.decode_tp_rank = None
             self.decode_tp_size = None
-            logger.info("Disable CP decode attention TP")
 
         self.use_decode_attn_tp = False
         self._slice_cache: Dict = {}

python/sglang/srt/layers/moe/fused_moe_triton/layer.py

@@ -83,7 +83,6 @@ from sglang.srt.utils import (
     is_cpu,
     is_hip,
     is_npu,
-    print_info_once,
     round_up,
 )
 from sglang.srt.utils.custom_op import register_custom_op
@@ -474,7 +473,7 @@ class FusedMoE(torch.nn.Module):
         global _deferred_finalize_info_logged
         if not _deferred_finalize_info_logged:
             _deferred_finalize_info_logged = True
-            logging.getLogger(__name__).info(
+            logging.getLogger(__name__).debug(
                 "FlashInfer TRTLLM MoE deferred finalize is "
                 f"{'enabled' if self.supports_deferred_finalize else 'disabled'} "
                 f"(moe_runner_backend={get_exec().moe.moe_runner_backend}, "
@@ -516,10 +515,6 @@ class FusedMoE(torch.nn.Module):
             get_moe_runner_backend().is_flashinfer_trtllm_routed()
             or get_moe_runner_backend().is_flashinfer_trtllm()
         ):
-            if self.moe_runner_config.inplace:
-                print_info_once(
-                    "Setting inplace to False for FlashInfer TRTLLM MoE backend."
-                )
             self.moe_runner_config.inplace = False
 
         self.should_fuse_routed_scaling_factor_in_topk = (

python/sglang/srt/layers/quantization/modelopt_quant.py

@@ -1429,10 +1429,7 @@ class ModelOptFp4Config(ModelOptQuantConfig):
         super().__init__(kv_cache_quant_algo, exclude_modules, packed_modules_mapping)
         self.is_checkpoint_nvfp4_serialized = is_checkpoint_nvfp4_serialized
         if is_checkpoint_nvfp4_serialized:
-            logger.warning(
-                "Detected nvfp4 checkpoint. Please note that the "
-                "format is experimental and subject to change."
-            )
+            logger.info("Detected nvfp4 checkpoint.")
         self.is_awq = is_awq
         self.is_w4a16 = False
         self.group_size = group_size

python/sglang/srt/layers/quantization/petit.py

@@ -45,10 +45,7 @@ class PetitNvFp4Config(QuantizationConfig):
     ) -> None:
         self.is_checkpoint_nvfp4_serialized = is_checkpoint_nvfp4_serialized
         if is_checkpoint_nvfp4_serialized:
-            logger.warning(
-                "Detected nvfp4 checkpoint. Please note that the "
-                "format is experimental and subject to change."
-            )
+            logger.info("Detected nvfp4 checkpoint.")
         self.group_size = group_size
         self.kv_cache_quant_algo = kv_cache_quant_algo
         self.exclude_modules = exclude_modules

python/sglang/srt/model_executor/model_runner_components/cuda_graph_setup.py

@@ -287,6 +287,7 @@ def capture_prefill_graph(
     """Initialize a prefill graph and return its startup resource usage."""
 
     memory_phase = "draft_prefill" if model_runner.is_draft_worker else "prefill"
+    role = "draft" if model_runner.is_draft_worker else "target"
 
     def result(
         runner: Optional[BaseRunner],
@@ -430,8 +431,10 @@ def capture_prefill_graph(
         layer_model = layer_model.model
 
     if not hasattr(layer_model, "layers"):
-        logger.warning(
-            "Disable prefill CUDA graph because the model does not have a 'layers' attribute"
+        log_info_on_rank0(
+            logger,
+            f"Disable {role} prefill CUDA graph because the {role} model does "
+            "not have a 'layers' attribute",
         )
         return result(None)
 
@@ -467,7 +470,6 @@ def capture_prefill_graph(
 
     tic = time.perf_counter()
     before_mem = get_available_gpu_memory(model_runner.device, model_runner.gpu_id)
-    role = "draft" if model_runner.is_draft_worker else "target"
     capture_name = f"{role} prefill"
     logger.info(
         f"Capture {capture_name} CUDA graph begin. "

python/sglang/srt/model_executor/model_runner_components/load_model_utils.py

@@ -132,8 +132,7 @@ def load_kv_cache_scales(*, model, kv_cache_dtype: str) -> None:
         else:
             logger.warning(
                 "Using FP8 KV cache but no scaling factors "
-                "provided. Defaulting to scaling factors of 1.0. "
-                "This may lead to less accurate results!"
+                "provided. Defaulting to scaling factors of 1.0."
             )
 
 

python/sglang/srt/model_executor/runner_backend_utils/breakable_cuda_graph/breakable_cuda_graph.py

@@ -23,6 +23,7 @@ buffers to keep break-point tensors at stable addresses.
 """
 
 import threading
+import warnings
 from contextvars import ContextVar
 from typing import Any, Callable, Optional
 
@@ -394,7 +395,12 @@ class BreakableCUDAGraphCapture:
             forked.clear()
         graph = self._current_graph
         assert graph is not None
-        graph.capture_end()
+        # A segment that enqueued no kernels (back-to-back breaks, or a segment
+        # whose ops all ran eagerly) captures an empty graph, which replays as a
+        # no-op. Torch warns about it on every such capture_end; expected here.
+        with warnings.catch_warnings():
+            warnings.filterwarnings("ignore", message="The CUDA Graph is empty")
+            graph.capture_end()
         self.cuda_graph._append_segment(graph, self._current_graph_needs_instantiate)
         self._current_graph = None
         self._current_graph_needs_instantiate = False

python/sglang/srt/model_loader/loader.py

@@ -640,7 +640,7 @@ class DefaultModelLoader(BaseModelLoader):
                     {"enable_multithread_load", "num_threads"} & extra_config.keys()
                 )
             ):
-                logger.warning(
+                logger.debug(
                     "Checkpoint prefetching is active; falling "
                     "back to single-threaded weight loading to avoid I/O "
                     "oversubscription with the prefetch threads. Set "

python/sglang/srt/model_loader/weight_utils.py

@@ -1021,7 +1021,7 @@ def _prefetch_all_checkpoints(
     succeeded_event = threading.Event()
     errors: List[Tuple[str, Exception]] = []
 
-    logger.info(
+    logger.debug(
         "Rank %d: prefetching %d/%d checkpoint shards into page cache "
         "(background, %d local ranks sharing the work, %d threads per rank)...",
         local_rank,
@@ -1042,7 +1042,7 @@ def _prefetch_all_checkpoints(
             if total_for_rank > 0 and next_log_pct <= 100:
                 pct = 100 * completed / total_for_rank
                 while pct >= next_log_pct and next_log_pct <= 100:
-                    logger.info(
+                    logger.debug(
                         "Rank %d: prefetching checkpoint files: %d%% (%d/%d)",
                         local_rank,
                         next_log_pct,
@@ -1093,7 +1093,7 @@ def _prefetch_all_checkpoints(
         start = time.perf_counter()
         _prefetch_all()
         succeeded_event.set()
-        logger.info(
+        logger.debug(
             "Rank %d: prefetching checkpoint files into page cache finished in %.2fs",
             local_rank,
             time.perf_counter() - start,

python/sglang/srt/models/bailing_moe_nextn.py

@@ -80,7 +80,7 @@ class BailingMoEModelNextN(nn.Module):
         config.for_nextn_model = True
 
         if quant_config is not None and quant_config.get_name() == "modelopt_fp4":
-            logger.warning(
+            logger.debug(
                 "Overriding DeepseekV3ForCausalLMNextN quant config for modelopt_fp4 Deepseek model."
             )
             quant_config = None

python/sglang/srt/models/deepseek_nextn.py

@@ -116,7 +116,7 @@ class DeepseekModelNextN(nn.Module):
             moe_quant_config_override = None
 
         if quant_config is not None and quant_config.get_name() == "modelopt_fp4":
-            logger.warning(
+            logger.debug(
                 "Overriding DeepseekV3ForCausalLMNextN quant config for modelopt_fp4 Deepseek model."
             )
             quant_config = None

python/sglang/srt/models/glm4_moe_lite_nextn.py

@@ -50,7 +50,7 @@ class Glm4MoeLiteModelNextN(nn.Module):
     ) -> None:
         super().__init__()
         if quant_config is not None and quant_config.get_name() == "modelopt_fp4":
-            logger.warning(
+            logger.debug(
                 "Overriding Glm4MoeLiteForCausalLMNextN quant config for modelopt_fp4 "
                 "GLM-4.7-Flash model."
             )

python/sglang/srt/models/glm4_moe_nextn.py

@@ -47,7 +47,7 @@ class Glm4MoeModelNextN(nn.Module):
     ) -> None:
         super().__init__()
         if quant_config is not None and quant_config.get_name() == "modelopt_fp4":
-            logger.warning(
+            logger.debug(
                 "Overriding Glm4MoeForCausalLMNextN quant config for modelopt_fp4 GLM-4.5 / GLM-4.6 / GLM-4.7 model."
             )
             quant_config = None

python/sglang/srt/models/glm_ocr_nextn.py

@@ -49,7 +49,7 @@ class GlmOcrModelNextN(nn.Module):
     ) -> None:
         super().__init__()
         if quant_config is not None and quant_config.get_name() == "modelopt_fp4":
-            logger.warning(
+            logger.debug(
                 "Overriding GlmOcrModelNextN quant config for modelopt_fp4 GLM-OCR model."
             )
             quant_config = None

python/sglang/srt/utils/common.py

@@ -58,6 +58,7 @@ from importlib.metadata import PackageNotFoundError, version
 from importlib.util import find_spec
 from io import BytesIO
 from json import JSONDecodeError
+from multiprocessing import parent_process
 from multiprocessing.reduction import ForkingPickler
 from pathlib import Path
 from typing import (
@@ -2409,6 +2410,14 @@ def configure_logger(server_args, prefix: str = ""):
     for name in ("httpx", "httpcore"):
         logging.getLogger(name).setLevel(logging.WARNING)
 
+    # Server-sent hub warnings (e.g. the unauthenticated-request / HF_TOKEN
+    # hint) are deduplicated per process, so a TP-N launch repeats each one N
+    # times. Keep them only in the launching process -- every worker (scheduler,
+    # detokenizer, DP controller, ...) is spawned via multiprocessing, whether
+    # or not it passes a log prefix -- where they are printed exactly once.
+    if parent_process() is not None:
+        logging.getLogger("huggingface_hub.utils._http").setLevel(logging.ERROR)
+
     if is_flashinfer_available():
         from flashinfer.jit.core import logger as flashinfer_logger
 
@@ -4106,7 +4115,7 @@ def freeze_gc(context: str):
     g0_before, g1_before, g2_before = gc_object_counts()
     gc.freeze()
     g0_after, g1_after, g2_after = gc_object_counts()
-    logger.info(
+    logger.debug(
         f"Freezing GC in {context} process. "
         f"gen0: {g0_before}->{g0_after}, "
         f"gen1: {g1_before}->{g1_after}, "

python/sglang/srt/utils/hf_transformers_patches.py

@@ -53,6 +53,8 @@ def apply_all():
         return
     _applied = True
 
+    _mute_diffusers_torchao_probe()
+
     # v5.4 patches
     _patch_flash_attn_availability()
     _patch_rope_parameters_validation()
@@ -71,6 +73,23 @@ def apply_all():
     logger.debug("transformers compatibility patches applied")
 
 
+def _mute_diffusers_torchao_probe():
+    """Silence diffusers' torchao-Tensor-subclass probe warning.
+
+    diffusers lazily imports its torchao quantizer and warns when the installed
+    torchao has moved the optional Tensor subclasses it probes for. It only
+    affects loading torchao-serialized diffusers checkpoints, which no sglang
+    path does. Set here rather than in ``configure_logger`` because the import
+    can land before logging is configured, and the level sticks whenever the
+    lazy import happens.
+    """
+    import logging
+
+    logging.getLogger("diffusers.quantizers.torchao.torchao_quantizer").setLevel(
+        logging.ERROR
+    )
+
+
 # ---------------------------------------------------------------------------
 # Public API: on-demand helpers (called explicitly by other modules)
 # ---------------------------------------------------------------------------
@@ -203,15 +222,22 @@ def _patch_removed_symbols():
 
         # 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)
+        # torchao emits a noisy warning about incompatible torch versions, and
+        # its register_as_pytree_constant() calls on Enum types make
+        # torch.utils._pytree log a deprecation warning once per Enum and per
+        # rank. Neither is actionable here — suppress both during this import.
+        _muted = [
+            logging.getLogger("torchao"),
+            logging.getLogger("torch.utils._pytree"),
+        ]
+        _prev_levels = [lg.level for lg in _muted]
+        for lg in _muted:
+            lg.setLevel(logging.ERROR)
         try:
             from transformers.models.llama import modeling_llama
         finally:
-            _torchao_logger.setLevel(_prev_level)
+            for lg, level in zip(_muted, _prev_levels):
+                lg.setLevel(level)
 
         if not hasattr(modeling_llama, "LlamaFlashAttention2"):
             if hasattr(modeling_llama, "LlamaAttention"):