Simplify server startup output (#20885)

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

.gitignore

@@ -224,16 +224,11 @@ work_dirs/
 *.exe
 *.out
 *.app
-
-compile_commands.json
-
 *.iml
 
 # VSCode
 .vscode
 
-1
-
 # Autoenv
 .env.leave
 
@@ -243,6 +238,7 @@ Cargo.lock
 # Generated vision test fixtures (regenerate with: python scripts/generate_vision_golden.py)
 sgl-model-gateway/tests/fixtures/golden/
 
+# Other repos
 lmms-eval
 
 **/.serena/
@@ -262,7 +258,6 @@ inputs/
 # setuptools-scm generated version file
 python/sglang/_version.py
 
-
 # MUSA section
 # Generated source files by torchada
 sgl-kernel/csrc_musa/

python/sglang/cli/main.py

@@ -14,13 +14,11 @@ def main():
 
     # complex sub commands
     subparsers = parser.add_subparsers(dest="subcommand", required=True)
-
     subparsers.add_parser(
         "serve",
-        help="Launch the SGLang server.",
+        help="Launch an SGLang server.",
         add_help=False,
     )
-
     subparsers.add_parser(
         "generate",
         help="Run inference on a multimodal model.",

python/sglang/cli/serve.py

@@ -52,17 +52,13 @@ def serve(args, extra_argv):
         # we can't show the exact help. Instead, we show a general help message and then
         # the help for both possible server types.
         print(
-            "Usage: sglang serve --model-path <model-name-or-path> [additional-arguments]\n"
+            "Usage: sglang serve --model-path <model-name-or-path> [additional-arguments]\n\n"
-        )
+            "This command can launch either a standard language model server or a diffusion model server.\n"
-        print(
+            "The server type is determined by the --model-path.\n"
-            "This command can launch either a standard language model server or a diffusion model server."
-        )
-        print("The server type is determined by the model path.\n")
-        print(
             "Optional override: --model-type {auto,llm,diffusion} "
-            "(default: auto, fallback to LLM on detection failure).\n"
+            "(default: auto, fallback to LLM on detection failure)."
         )
-        print("For specific arguments, please provide a model_path.")
+
         print("\n--- Help for Standard Language Model Server ---")
         from sglang.srt.server_args import prepare_server_args
 
@@ -72,13 +68,22 @@ def serve(args, extra_argv):
             pass  # argparse --help calls sys.exit
 
         print("\n--- Help for Diffusion Model Server ---")
-        from sglang.multimodal_gen.runtime.entrypoints.cli.serve import (
+        try:
-            add_multimodal_gen_serve_args,
+            from sglang.multimodal_gen.runtime.entrypoints.cli.serve import (
-        )
+                add_multimodal_gen_serve_args,
+            )
 
-        parser = argparse.ArgumentParser(description="SGLang Diffusion Model Serving")
+            parser = argparse.ArgumentParser(
-        add_multimodal_gen_serve_args(parser)
+                prog="sglang serve",
-        parser.print_help()
+                description="SGLang Diffusion Model Serving",
+            )
+            add_multimodal_gen_serve_args(parser)
+            parser.print_help()
+        except ImportError:
+            print(
+                "Diffusion model support is not available. "
+                'Install with: pip install "sglang[diffusion]"'
+            )
         return
 
     model_type, dispatch_argv = _extract_model_type_override(extra_argv)

python/sglang/cli/utils.py

@@ -80,8 +80,7 @@ def get_model_path(extra_argv):
             raise Exception(
                 "Usage: sglang serve --model-path <model-name-or-path> [additional-arguments]\n\n"
                 "This command can launch either a standard language model server or a diffusion model server.\n"
-                "The server type is determined by the model path.\n"
+                "The server type is determined by the --model-path.\n"
-                "For specific arguments, please provide a model_path."
             )
         else:
             raise Exception(

python/sglang/launch_server.py

@@ -3,6 +3,7 @@
 import asyncio
 import os
 import sys
+import warnings
 
 from sglang.srt.server_args import prepare_server_args
 from sglang.srt.utils import kill_process_tree
@@ -14,6 +15,7 @@ suppress_noisy_warnings()
 def run_server(server_args):
     """Run the server based on server_args.grpc_mode and server_args.encoder_only."""
     if server_args.encoder_only:
+        # For encoder disaggregation
         if server_args.grpc_mode:
             from sglang.srt.disaggregation.encode_grpc_server import (
                 serve_grpc_encoder,
@@ -46,8 +48,6 @@ def run_server(server_args):
 
 
 if __name__ == "__main__":
-    import warnings
-
     warnings.warn(
         "'python -m sglang.launch_server' is still supported, but "
         "'sglang serve' is the recommended entrypoint.\n"

python/sglang/multimodal_gen/runtime/platforms/__init__.py

@@ -51,7 +51,7 @@ def cuda_platform_plugin() -> str | None:
         if cuda_is_jetson():
             is_cuda = True
     if is_cuda:
-        logger.info("CUDA is available")
+        logger.debug("CUDA is available")
 
     return (
         "sglang.multimodal_gen.runtime.platforms.cuda.CudaPlatform" if is_cuda else None
@@ -67,9 +67,9 @@ def mps_platform_plugin() -> str | None:
 
         if torch.backends.mps.is_available():
             is_mps = True
-            logger.info("MPS (Metal Performance Shaders) is available")
+            logger.debug("MPS (Metal Performance Shaders) is available")
     except Exception as e:
-        logger.info("MPS detection failed: %s", e)
+        logger.debug("MPS detection failed: %s", e)
 
     return "sglang.multimodal_gen.runtime.platforms.mps.MpsPlatform" if is_mps else None
 
@@ -90,11 +90,11 @@ def rocm_platform_plugin() -> str | None:
         try:
             if len(amdsmi.amdsmi_get_processor_handles()) > 0:
                 is_rocm = True
-                logger.info("ROCm platform is available")
+                logger.debug("ROCm platform is available")
         finally:
             amdsmi.amdsmi_shut_down()
     except Exception as e:
-        logger.info("ROCm platform is unavailable: %s", e)
+        logger.debug("ROCm platform is unavailable: %s", e)
 
     return (
         "sglang.multimodal_gen.runtime.platforms.rocm.RocmPlatform" if is_rocm else None
@@ -109,9 +109,9 @@ def npu_platform_plugin() -> str | None:
 
         if torch.npu.is_available():
             is_npu = True
-            logger.info("NPU is available")
+            logger.debug("NPU is available")
     except Exception as e:
-        logger.info("NPU detection failed: %s", e)
+        logger.debug("NPU detection failed: %s", e)
     return (
         "sglang.multimodal_gen.runtime.platforms.npu.NPUPlatformBase"
         if is_npu
@@ -131,7 +131,7 @@ def musa_platform_plugin() -> str | None:
         finally:
             pymtml.mtmlLibraryShutDown()
     except Exception as e:
-        logger.info("MUSA platform is unavailable: %s", e)
+        logger.debug("MUSA platform is unavailable: %s", e)
 
     return (
         "sglang.multimodal_gen.runtime.platforms.musa.MusaPlatform" if is_musa else None

python/sglang/srt/compilation/backend.py

@@ -21,7 +21,7 @@ from sglang.srt.compilation.compiler_interface import EagerAdapter, InductorAdap
 from sglang.srt.compilation.cuda_piecewise_backend import CUDAPiecewiseBackend
 from sglang.srt.compilation.npu_piecewise_backend import NPUPiecewiseBackend
 from sglang.srt.compilation.pass_manager import PostGradPassManager
-from sglang.srt.utils.common import is_npu, rank0_log
+from sglang.srt.utils.common import is_npu
 
 logger = logging.getLogger(__name__)
 
@@ -375,7 +375,6 @@ class SGLangBackend:
         config: CompilationConfig,
         graph_pool: Any,
     ):
-        rank0_log(f"Initializing SGLangBackend")
         assert graph_pool is not None
         self.graph_pool = graph_pool
 
@@ -394,7 +393,6 @@ class SGLangBackend:
         self.inductor_config["post_grad_custom_post_pass"] = self.post_grad_pass_manager
 
     def __call__(self, graph: fx.GraphModule, example_inputs) -> Callable:
-        rank0_log(f"SGLangBackend __call__")
         base_cache_dir = os.path.expanduser(
             os.getenv("SGLANG_CACHE_DIR", "~/.cache/sglang/")
         )
@@ -466,7 +464,5 @@ class SGLangBackend:
                 with open(graph_path, "w") as f:
                     f.write(src)
 
-                rank0_log(f"Computation graph saved to {graph_path}")
-
         self._called = True
         return self.split_gm

python/sglang/srt/compilation/compile.py

@@ -10,7 +10,6 @@ import torch
 
 from sglang.srt.compilation.compilation_config import CompilationConfig
 from sglang.srt.compilation.piecewise_context_manager import is_in_piecewise_cuda_graph
-from sglang.srt.utils.common import rank0_log
 
 logger = logging.getLogger(__name__)
 
@@ -118,7 +117,6 @@ def install_torch_compiled(
     fullgraph: bool = True,
     graph_pool: Any = None,
 ):
-    rank0_log(f"install_torch_compiled")
     unbound_fwd = module.__class__.forward
     if not callable(unbound_fwd):
         raise TypeError("module.__class__.forward must be callable")

python/sglang/srt/layers/attention/mamba/ops/ssu_dispatch.py

@@ -198,7 +198,7 @@ def initialize_mamba_selective_state_update_backend(server_args: ServerArgs) ->
             f"--mamba-backend value."
         )
 
-    logger.info(
+    logger.debug(
         "Mamba selective_state_update backend initialized: %s",
         _mamba_ssu_backend.name,
     )

python/sglang/srt/managers/scheduler.py

@@ -25,6 +25,10 @@ from dataclasses import dataclass
 from http import HTTPStatus
 from typing import Any, Deque, Dict, List, Optional, Tuple, Union
 
+from sglang.srt.utils.common import suppress_noisy_warnings
+
+suppress_noisy_warnings()
+
 import psutil
 import setproctitle
 import torch

python/sglang/srt/observability/scheduler_metrics_mixin.py

@@ -258,9 +258,6 @@ class SchedulerMetricsMixin:
         if self.disaggregation_mode == DisaggregationMode.PREFILL:
             msg += f"#prealloc-req: {len(self.disagg_prefill_bootstrap_queue.queue)}, "
             msg += f"#inflight-req: {len(self.disagg_prefill_inflight_queue)}, "
-            msg += f"input throughput (token/s): {self.last_input_throughput:.2f}, "
-        else:
-            msg += f"input throughput (token/s): {self.last_input_throughput:.2f}, "
 
         if (
             self.server_args.language_only
@@ -275,7 +272,8 @@ class SchedulerMetricsMixin:
             },
         )
 
-        msg += f"{graph_backend[self.device]}: {can_run_cuda_graph}"
+        msg += f"{graph_backend[self.device]}: {can_run_cuda_graph}, "
+        msg += f"input throughput (token/s): {self.last_input_throughput:.2f}"
 
         if self.is_stats_logging_rank:
             logger.info(msg)

python/sglang/srt/server_args.py

@@ -6282,7 +6282,7 @@ def prepare_server_args(argv: List[str]) -> ServerArgs:
     Returns:
         The server arguments.
     """
-    parser = argparse.ArgumentParser()
+    parser = argparse.ArgumentParser(prog="sglang serve")
     ServerArgs.add_cli_args(parser)
 
     # Check for config file and merge arguments if present

python/sglang/srt/utils/common.py

@@ -121,7 +121,7 @@ builtins.FP8_E4M3_MIN = FP8_E4M3_MIN
 # this makes it impossible to see the animation in the progress bar
 # but will avoid messing up with ray or multiprocessing, which wraps
 # each line of output with some prefix.
-BAR_FORMAT = "{desc}: {percentage:3.0f}% Completed | {n_fmt}/{total_fmt} [{elapsed}<{remaining}, {rate_fmt}]\n"  # noqa: E501
+BAR_FORMAT = "{desc}: {percentage:3.0f}% Completed | {n_fmt}/{total_fmt} [{elapsed}<{remaining}, {rate_fmt}]"
 
 
 @lru_cache(maxsize=1)