@tduncan / Nvidia-Optimization-Stack-VLLM / commits / 90329913e9

[UX] Reduce startup log noise (#50590)

Signed-off-by: mgoin <mgoin64@gmail.com> Co-authored-by: OpenAI Codex <noreply@openai.com>

Michael Goin committed Jul 31, 2026 at 16:04 UTC 90329913e945cb3ed515bd8e1712094cecb76a0c

9 files changed +36 -50

vllm/config/vllm.py

+11 -16

@@ -1142,11 +1142,6 @@ class VllmConfig:

1142 else:

1143 self.scheduler_config.async_scheduling = True

1144

1145 - logger.info_once(

1146 - "Asynchronous scheduling is %s.",

1147 - "enabled" if self.scheduler_config.async_scheduling else "disabled",

1148 - )

1149 -

1145 if self.parallel_config.disable_nccl_for_dp_synchronization is None:

1146 if self.scheduler_config.async_scheduling:

1147 if self.parallel_config.data_parallel_size > 1 and (

@@ -1196,7 +1191,7 @@ class VllmConfig:

1191 )

1192

1193 if self.model_config is not None and self.model_config.enforce_eager:

1199 - logger.warning(

1194 + logger.warning_once(

1195 "Enforce eager set, disabling torch.compile and CUDAGraphs. "

1196 "This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none"

1197 )

@@ -1204,7 +1199,7 @@ class VllmConfig:

1199 self.compilation_config.cudagraph_mode = CUDAGraphMode.NONE

1200

1201 if os.environ.get("TORCH_COMPILE_DISABLE") == "1":

1207 - logger.warning(

1202 + logger.warning_once(

1203 "TORCH_COMPILE_DISABLE is set, disabling torch.compile. "

1204 "This is equivalent to setting -cc.mode=none"

1205 )

@@ -1249,7 +1244,7 @@ class VllmConfig:

1244 self.compilation_config.mode is not None

1245 and self.compilation_config.mode != CompilationMode.VLLM_COMPILE

1246 ):

1252 - logger.warning(

1247 + logger.warning_once(

1248 "Inductor compilation was disabled by user settings, "

1249 "optimizations settings that are only active during "

1250 "inductor compilation will be ignored."

@@ -1317,7 +1312,7 @@ class VllmConfig:

1312 and self.compilation_config.mode != CompilationMode.VLLM_COMPILE

1313 and not envs.VLLM_USE_BREAKABLE_CUDAGRAPH

1314 ):

1320 - logger.info(

1315 + logger.info_once(

1316 "Cudagraph mode %s is not compatible with compilation mode %s."

1317 "Overriding to NONE.",

1318 self.compilation_config.cudagraph_mode,

@@ -1331,7 +1326,7 @@ class VllmConfig:

1326 pass_config.enable_sp = True

1327 if pass_config.enable_sp:

1328 if self.parallel_config.tensor_parallel_size == 1:

1334 - logger.warning("Sequence Parallelism requires TP>1, disabling")

1329 + logger.warning_once("Sequence Parallelism requires TP>1, disabling")

1330 pass_config.enable_sp = False

1331 pass_config.fuse_gemm_comms = False

1332 else:

@@ -1349,7 +1344,7 @@ class VllmConfig:

1344 )

1345

1346 if pass_config.sp_min_token_num is None:

1352 - logger.warning(

1347 + logger.warning_once(

1348 "Model hidden_size too small for the SP "

1349 "threshold heuristic, disabling. To force SP, "

1350 "set pass_config.sp_min_token_num manually."

@@ -1428,7 +1423,7 @@ class VllmConfig:

1423

1424 # disable cudagraph when enforce eager execution

1425 if self.model_config is not None and self.model_config.enforce_eager:

1431 - logger.info("Cudagraph is disabled under eager mode")

1426 + logger.info_once("Cudagraph is disabled under eager mode")

1427 self.compilation_config.cudagraph_mode = CUDAGraphMode.NONE

1428 # override related settings when enforce eager

1429 self.compilation_config.max_cudagraph_capture_size = 0

@@ -1463,7 +1458,7 @@ class VllmConfig:

1458 and self.model_config.architecture == "WhisperForConditionalGeneration"

1459 and os.environ.get("VLLM_WORKER_MULTIPROC_METHOD") != "spawn"

1460 ):

1466 - logger.warning(

1461 + logger.warning_once(

1462 "Whisper is known to have issues with "

1463 "forked workers. If startup is hanging, "

1464 "try setting 'VLLM_WORKER_MULTIPROC_METHOD' "

@@ -1475,7 +1470,7 @@ class VllmConfig:

1470 and self.kv_events_config.enable_kv_cache_events

1471 and not self.cache_config.enable_prefix_caching

1472 ):

1478 - logger.warning(

1473 + logger.warning_once(

1474 "KV cache events are on, but prefix caching is not enabled. "

1475 "Use --enable-prefix-caching to enable."

1476 )

@@ -1484,7 +1479,7 @@ class VllmConfig:

1479 and self.kv_events_config.publisher != "null"

1480 and not self.kv_events_config.enable_kv_cache_events

1481 ):

1487 - logger.warning(

1482 + logger.warning_once(

1483 "KV cache events are disabled, "

1484 "but the scheduler is configured to publish them. "

1485 "Modify KVEventsConfig.enable_kv_cache_events "

@@ -1522,7 +1517,7 @@ class VllmConfig:

1517 # the pass will operate on higher-level IR to avoid the issue.

1518 # TODO: https://github.com/vllm-project/vllm/issues/27894

1519 if self.compilation_config.mode != CompilationMode.VLLM_COMPILE:

1525 - logger.warning(

1520 + logger.warning_once(

1521 "Sequence parallelism is enabled, but running in wrong "

1522 "vllm compile mode: %s.",

1523 self.compilation_config.mode,

vllm/distributed/device_communicators/custom_all_reduce.py

+3 -3

@@ -107,7 +107,7 @@ class CustomAllreduce:

107 if not custom_ar:

108 # disable because of missing custom allreduce library

109 # e.g. in a non-GPU environment

110 - logger.info(

110 + logger.info_once(

111 "Custom allreduce is disabled because "

112 "of missing custom allreduce library"

113 )

@@ -130,7 +130,7 @@ class CustomAllreduce:

130 return

131

132 if world_size not in CustomAllreduce._SUPPORTED_WORLD_SIZES:

133 - logger.warning(

133 + logger.warning_once(

134 "Custom allreduce is disabled due to an unsupported world"

135 " size: %d. Supported world sizes: %s. To silence this "

136 "warning, specify disable_custom_all_reduce=True explicitly.",

@@ -328,7 +328,7 @@ class CustomAllreduce:

328

329 def register_graph_buffers(self):

330 handle, offset = ops.get_graph_buffer_ipc_meta(self._ptr)

331 - logger.info("Registering %d cuda graph addresses", len(offset))

331 + logger.debug("Registering %d cuda graph addresses", len(offset))

332 # We cannot directly use `dist.all_gather_object` here

333 # because it is incompatible with `gloo` backend under inference mode.

334 # see https://github.com/pytorch/pytorch/issues/126032 for details.

vllm/distributed/device_communicators/flashinfer_all_reduce.py

+2 -2

@@ -121,7 +121,7 @@ def _resolve_fi_ar_backend() -> tuple[str, bool]:

121 """

122 backend = envs.VLLM_FLASHINFER_ALLREDUCE_BACKEND

123 if backend != "auto":

124 - logger.info_once(f"Using flashinfer allreduce backend: {backend}")

124 + logger.debug_once("Using flashinfer allreduce backend: %s", backend)

125 return backend, False

126

127 # Default to mnnvl for both single- and multi-node setups. The mnnvl

@@ -134,7 +134,7 @@ def _resolve_fi_ar_backend() -> tuple[str, bool]:

134 backend = "mnnvl"

135 allow_trtllm_fallback = get_node_count() == 1

136

137 - logger.info_once(f"Auto-selected flashinfer allreduce backend: {backend}")

137 + logger.debug_once("Auto-selected flashinfer allreduce backend: %s", backend)

138 return backend, allow_trtllm_fallback

139

140

vllm/model_executor/warmup/cutedsl_warmup.py

+5 -14

@@ -8,7 +8,6 @@

9 from __future__ import annotations

11 -import time

11 import weakref

12 from collections.abc import Callable, Hashable, Iterable

13 from dataclasses import dataclass

@@ -97,25 +96,17 @@ def _compile_cutedsl_warmup_units(

96 def cutedsl_warmup() -> None:

97 """Run CuTeDSL compile providers before serving."""

98 if not current_platform.is_cuda():

100 - logger.info("Skipping CuTeDSL warmup on non-CUDA platform.")

99 + logger.debug("Skipping CuTeDSL warmup on non-CUDA platform.")

100 return

101

102 compile_units = _collect_unique_compile_units(_iter_cutedsl_warmup_compile_units())

103 if not compile_units:

105 - logger.info("Skipping CuTeDSL warmup because no compile units were requested.")

104 + logger.debug("Skipping CuTeDSL warmup because no compile units were requested.")

105 return

106

108 - unit_names = list(dict.fromkeys(unit.name for unit in compile_units))

109 - logger.info(

107 + logger.info_once(

108 "Warming up CuTeDSL compile_units=%d names=%s.",

109 len(compile_units),

112 - unit_names,

113 - )

114 -

115 - start_time = time.perf_counter()

116 - compiled_count = _compile_cutedsl_warmup_units(compile_units)

117 - logger.info(

118 - "CuTeDSL warmup compiled %d units in %.2f s.",

119 - compiled_count,

120 - time.perf_counter() - start_time,

110 + tuple(dict.fromkeys(unit.name for unit in compile_units)),

111 )

112 + _compile_cutedsl_warmup_units(compile_units)

vllm/model_executor/warmup/kernel_warmup.py

+11 -11

@@ -84,12 +84,12 @@ def _warmup_ll_bf16_router_gemm(model: torch.nn.Module) -> None:

85 shapes = _ll_bf16_router_shapes_from_model(model)

86 if not shapes:

87 - logger.info(

87 + logger.debug_once(

88 "Skipping ll_bf16 router GEMM warmup: no bf16 GateLinear shapes found."

89 )

90 return

92 - logger.info("Warming up ll_bf16 router GEMM kernels for shapes: %s.", shapes)

92 + logger.info_once("Warming up ll_bf16 router GEMM kernels for shapes: %s.", shapes)

93 ll_bf16_gemm_kernel.warmup(

94 shapes=shapes,

95 m_values=_LL_BF16_WARMUP_M_RANGE,

@@ -163,7 +163,7 @@ def kernel_warmup(worker: "Worker", *, process_local_only: bool = False):

163 )

164 # FlashInfer autotune for Hopper (SM 9.0) and Blackwell (SM 10.0) GPUs

165 if enable_flashinfer_autotune is False:

166 - logger.info("Skipping FlashInfer autotune because it is disabled.")

166 + logger.info_once("Skipping FlashInfer autotune because it is disabled.")

167 elif has_flashinfer() and current_platform.has_device_capability(90):

168 flashinfer_autotune(worker.model_runner)

169

@@ -188,7 +188,7 @@ def kernel_warmup(worker: "Worker", *, process_local_only: bool = False):

188 for group in groups

189 )

190 ):

191 - logger.info("Warming up FlashInfer attention.")

191 + logger.info_once("Warming up FlashInfer attention.")

192 # Warmup with mixed batch containing both prefill and decode tokens

193 # This is to warm up both prefill and decode attention kernels

194 worker.model_runner._dummy_run(

@@ -238,9 +238,9 @@ def flashinfer_autotune(runner: "GPUModelRunner") -> None:

238 autotune_kwargs: dict = {}

239 skip_ops = _flashinfer_autotune_skip_ops(runner)

240 if skip_ops:

241 - logger.info(

241 + logger.info_once(

242 "Skipping FlashInfer autotuning for ops %s",

243 - sorted(skip_ops),

243 + tuple(sorted(skip_ops)),

244 )

245 autotune_kwargs["skip_ops"] = skip_ops

246

@@ -265,7 +265,7 @@ def flashinfer_autotune(runner: "GPUModelRunner") -> None:

265

266 cache_path = resolve_flashinfer_autotune_file(runner)

267 if is_leader:

268 - logger.info("Using FlashInfer autotune cache file: %s", cache_path)

268 + logger.info_once("Using FlashInfer autotune cache file: %s", cache_path)

269

270 # We skip EPLB here since we don't want to record dummy metrics.

271 # When autotuning with number of tokens m, flashinfer will autotune

@@ -296,9 +296,9 @@ def flashinfer_autotune(runner: "GPUModelRunner") -> None:

296 tune_results = world.broadcast_object(tune_results, src=0)

297

298 if tune_results is None:

299 - logger.warning(

300 - "No FlashInfer autotune cache entries found."

301 - "Falling back to default tactics."

299 + logger.warning_once(

300 + "No FlashInfer autotune cache entries found; "

301 + "falling back to default tactics."

302 )

303 else:

304 write_flashinfer_autotune_cache(cache_path, tune_results)

@@ -306,7 +306,7 @@ def flashinfer_autotune(runner: "GPUModelRunner") -> None:

306 from flashinfer.autotuner import AutoTuner

307

308 AutoTuner.get().load_configs(str(cache_path))

309 - logger.info(

309 + logger.info_once(

310 "FlashInfer autotune cache loaded on rank %d from %s.",

311 world.rank_in_group,

312 cache_path,

vllm/utils/system_utils.py

+1 -1

@@ -284,7 +284,7 @@ def kill_process_tree(pid: int):

284 # Adapted from: https://github.com/sgl-project/sglang/blob/v0.4.1/python/sglang/srt/utils.py#L630

285 def set_ulimit(target_soft_limit: int = 65535):

286 if sys.platform.startswith("win"):

287 - logger.info("Windows detected, skipping ulimit adjustment.")

287 + logger.debug("Windows detected, skipping ulimit adjustment.")

288 return

289

290 import resource

vllm/v1/attention/backends/utils.py

+1 -1

@@ -87,7 +87,7 @@ def get_kv_cache_layout():

87 cache_layout: Literal["NHD", "HND"] | None = None

88 if _KV_CACHE_LAYOUT_OVERRIDE is not None:

89 cache_layout = _KV_CACHE_LAYOUT_OVERRIDE

90 - logger.info_once(

90 + logger.debug_once(

91 "`_KV_CACHE_LAYOUT_OVERRIDE` variable detected. "

92 "Setting KV cache layout to %s.",

93 cache_layout,

vllm/v1/attention/selector.py

+1 -1

@@ -199,7 +199,7 @@ def _cached_get_attn_backend(

199 from vllm.v1.attention.backends.utils import set_kv_cache_layout

200

201 set_kv_cache_layout(required_layout)

202 - logger.info(

202 + logger.info_once(

203 "Using %s KV cache layout for %s backend.",

204 required_layout,

205 backend.get_name(),

vllm/v1/worker/gpu/model_runner.py

+1 -1

@@ -305,7 +305,7 @@ class GPUModelRunner(LoRAModelRunnerMixin):

305 eplb_models_added = False

306 with DeviceMemoryProfiler() as m:

307 model_loader = get_model_loader(self.vllm_config.load_config)

308 - logger.info("Loading model from scratch...")

308 + logger.info_once("Loading model from scratch...")

309

310 self.model = model_loader.load_model(

311 vllm_config=self.vllm_config, model_config=self.vllm_config.model_config