Clean up server startup log noise (#26205)

python/sglang/srt/configs/model_config.py

@@ -1450,15 +1450,15 @@ def _get_and_verify_dtype(
     if torch_dtype != config_dtype:
         if torch_dtype == torch.float32:
             # Upcasting to float32 is allowed.
-            logger.info("Upcasting %s to %s.", config_dtype, torch_dtype)
+            logger.debug("Upcasting %s to %s.", config_dtype, torch_dtype)
             pass
         elif config_dtype == torch.float32:
             # Downcasting from float32 to float16 or bfloat16 is allowed.
-            logger.info("Downcasting %s to %s.", config_dtype, torch_dtype)
+            logger.debug("Downcasting %s to %s.", config_dtype, torch_dtype)
             pass
         else:
             # Casting between float16 and bfloat16 is allowed with a warning.
-            logger.warning("Casting %s to %s.", config_dtype, torch_dtype)
+            logger.debug("Casting %s to %s.", config_dtype, torch_dtype)
 
     return torch_dtype
 

python/sglang/srt/layers/attention/flashinfer_backend.py

@@ -243,12 +243,10 @@ class FlashInferAttnBackend(AttentionBackend):
 
         fmha_backend = "auto"
         if is_sm100_supported():
-            # Disable CUTLASS backend when piecewise cuda graph is enabled
-            # due to TMA descriptor initialization issues on B200
             if not model_runner.server_args.disable_piecewise_cuda_graph:
-                logger.warning(
+                logger.info(
                     "CUTLASS backend is disabled when piecewise cuda graph is enabled "
-                    "due to TMA descriptor initialization issues on B200. "
+                    "due to TMA descriptor initialization issues on SM100 GPUs. "
                     "Using auto backend instead for stability."
                 )
             else:

python/sglang/srt/managers/template_detection.py

@@ -334,12 +334,6 @@ def match_rules(
     for rule in rules:
         try:
             if rule.predicate(ctx):
-                logger.info(
-                    "Detected %s '%s' from template rule '%s'.",
-                    label,
-                    rule.value,
-                    rule.name,
-                )
                 return rule.value
         except Exception as e:
             logger.warning(
@@ -367,11 +361,6 @@ def detect_reasoning_pattern(
     )
     for rule in REASONING_MODE_RULES:
         if rule.predicate(ctx):
-            logger.info(
-                "Detected reasoning config '%s' from template rule '%s'.",
-                rule.value,
-                rule.name,
-            )
             return rule.value.always_on, rule.value
 
     return False, None

python/sglang/srt/managers/template_manager.py

@@ -173,14 +173,15 @@ class TemplateManager:
         if tokenizer_manager.tokenizer:
             template = tokenizer_manager.tokenizer.chat_template
             self._run_template_detection(template, tokenizer_manager.tokenizer)
+            parts = []
+            if self._reasoning_config:
+                parts.append(f"reasoning_config={self._reasoning_config}")
             if self._suggested_reasoning_parser:
-                logger.info(
-                    f"Auto-detected reasoning parser: {self._suggested_reasoning_parser}"
-                )
+                parts.append(f"reasoning_parser={self._suggested_reasoning_parser}")
             if self._suggested_tool_call_parser:
-                logger.info(
-                    f"Auto-detected tool-call parser: {self._suggested_tool_call_parser}"
-                )
+                parts.append(f"tool_call_parser={self._suggested_tool_call_parser}")
+            if parts:
+                logger.info(f"Auto-detected template features: {', '.join(parts)}")
 
     def _load_explicit_chat_template(
         self, tokenizer_manager: TokenizerManager, chat_template_arg: str

python/sglang/srt/mem_cache/memory_pool.py

@@ -738,13 +738,13 @@ class KVCache(abc.ABC):
             k_size_GB = k_size / GB
             v_size_GB = v_size / GB
             logger.info(
-                f"KV Cache is allocated. #tokens: {num_tokens}, K size: {k_size_GB:.2f} GB, V size: {v_size_GB:.2f} GB"
+                f"KV Cache is allocated. dtype: {self.dtype}, #tokens: {num_tokens}, K size: {k_size_GB:.2f} GB, V size: {v_size_GB:.2f} GB"
             )
             self.mem_usage = k_size_GB + v_size_GB
         else:
             kv_size_GB = kv_size_bytes / GB
             logger.info(
-                f"KV Cache is allocated. #tokens: {num_tokens}, KV size: {kv_size_GB:.2f} GB"
+                f"KV Cache is allocated. dtype: {self.dtype}, #tokens: {num_tokens}, KV size: {kv_size_GB:.2f} GB"
             )
             self.mem_usage = kv_size_GB
 

python/sglang/srt/model_executor/model_runner.py

@@ -2214,8 +2214,6 @@ class ModelRunner(ModelRunnerKVCacheMixin):
                 f"Unsupported kv_cache_dtype: {self.server_args.kv_cache_dtype}."
             )
 
-        log_info_on_rank0(logger, f"Using KV cache dtype: {self.kv_cache_dtype}")
-
     def init_cublas(self):
         """We need to run a small matmul to init cublas. Otherwise, it will raise some errors later."""
         dtype = torch.float16

python/sglang/srt/models/gpt_oss.py

@@ -219,7 +219,7 @@ class GptOssSparseMoeBlock(nn.Module):
             bias=True,
             quant_config=None,
             prefix=add_prefix("gate", prefix),
-            params_dtype=config.torch_dtype,
+            params_dtype=config.dtype,
         )
 
     def forward(
@@ -468,7 +468,7 @@ class GptOssDecoderLayer(nn.Module):
             prefix=add_prefix("self_attn", prefix),
             sliding_window_size=self.sliding_window_size,
             layer_type=config.layer_types[layer_id],
-            params_dtype=config.torch_dtype,
+            params_dtype=config.dtype,
         )
 
         self.layer_id = layer_id

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

@@ -105,7 +105,7 @@ def _load_tokenizer_by_declared_class(tokenizer_name, *args, **kwargs):
     if tok_cls is None:
         return None
 
-    logger.info(
+    logger.debug(
         "Loading tokenizer for %s directly as %s (bypassing AutoTokenizer)",
         tokenizer_name,
         tok_class_name,
@@ -208,7 +208,7 @@ def _resolve_tokenizers_backend(tokenizer_name, *args, **common_kwargs):
     ``tokenizer_config.json``.  May still return a ``TokenizersBackend``
     if all retries fail (with a warning).
     """
-    logger.warning(
+    logger.debug(
         "Tokenizer loaded as generic TokenizersBackend for %s, "
         "retrying with use_fast=False",
         tokenizer_name,
@@ -239,7 +239,7 @@ def _resolve_tokenizers_backend(tokenizer_name, *args, **common_kwargs):
                 tokenizer_name,
             )
         else:
-            logger.warning(
+            logger.debug(
                 "Tokenizer for %s loaded as generic TokenizersBackend. "
                 "Set --trust-remote-code to load the model-specific tokenizer.",
                 tokenizer_name,