[UX] Reduce startup log noise (#50590)
Signed-off-by: mgoin <mgoin64@gmail.com> Co-authored-by: OpenAI Codex <noreply@openai.com>
This commit is contained in:
co-authored by
OpenAI Codex
parent
e8b358b4c0
commit
90329913e9
+11
-16
@@ -1142,11 +1142,6 @@ class VllmConfig:
|
||||
else:
|
||||
self.scheduler_config.async_scheduling = True
|
||||
|
||||
logger.info_once(
|
||||
"Asynchronous scheduling is %s.",
|
||||
"enabled" if self.scheduler_config.async_scheduling else "disabled",
|
||||
)
|
||||
|
||||
if self.parallel_config.disable_nccl_for_dp_synchronization is None:
|
||||
if self.scheduler_config.async_scheduling:
|
||||
if self.parallel_config.data_parallel_size > 1 and (
|
||||
@@ -1196,7 +1191,7 @@ class VllmConfig:
|
||||
)
|
||||
|
||||
if self.model_config is not None and self.model_config.enforce_eager:
|
||||
logger.warning(
|
||||
logger.warning_once(
|
||||
"Enforce eager set, disabling torch.compile and CUDAGraphs. "
|
||||
"This is equivalent to setting -cc.mode=none -cc.cudagraph_mode=none"
|
||||
)
|
||||
@@ -1204,7 +1199,7 @@ class VllmConfig:
|
||||
self.compilation_config.cudagraph_mode = CUDAGraphMode.NONE
|
||||
|
||||
if os.environ.get("TORCH_COMPILE_DISABLE") == "1":
|
||||
logger.warning(
|
||||
logger.warning_once(
|
||||
"TORCH_COMPILE_DISABLE is set, disabling torch.compile. "
|
||||
"This is equivalent to setting -cc.mode=none"
|
||||
)
|
||||
@@ -1249,7 +1244,7 @@ class VllmConfig:
|
||||
self.compilation_config.mode is not None
|
||||
and self.compilation_config.mode != CompilationMode.VLLM_COMPILE
|
||||
):
|
||||
logger.warning(
|
||||
logger.warning_once(
|
||||
"Inductor compilation was disabled by user settings, "
|
||||
"optimizations settings that are only active during "
|
||||
"inductor compilation will be ignored."
|
||||
@@ -1317,7 +1312,7 @@ class VllmConfig:
|
||||
and self.compilation_config.mode != CompilationMode.VLLM_COMPILE
|
||||
and not envs.VLLM_USE_BREAKABLE_CUDAGRAPH
|
||||
):
|
||||
logger.info(
|
||||
logger.info_once(
|
||||
"Cudagraph mode %s is not compatible with compilation mode %s."
|
||||
"Overriding to NONE.",
|
||||
self.compilation_config.cudagraph_mode,
|
||||
@@ -1331,7 +1326,7 @@ class VllmConfig:
|
||||
pass_config.enable_sp = True
|
||||
if pass_config.enable_sp:
|
||||
if self.parallel_config.tensor_parallel_size == 1:
|
||||
logger.warning("Sequence Parallelism requires TP>1, disabling")
|
||||
logger.warning_once("Sequence Parallelism requires TP>1, disabling")
|
||||
pass_config.enable_sp = False
|
||||
pass_config.fuse_gemm_comms = False
|
||||
else:
|
||||
@@ -1349,7 +1344,7 @@ class VllmConfig:
|
||||
)
|
||||
|
||||
if pass_config.sp_min_token_num is None:
|
||||
logger.warning(
|
||||
logger.warning_once(
|
||||
"Model hidden_size too small for the SP "
|
||||
"threshold heuristic, disabling. To force SP, "
|
||||
"set pass_config.sp_min_token_num manually."
|
||||
@@ -1428,7 +1423,7 @@ class VllmConfig:
|
||||
|
||||
# disable cudagraph when enforce eager execution
|
||||
if self.model_config is not None and self.model_config.enforce_eager:
|
||||
logger.info("Cudagraph is disabled under eager mode")
|
||||
logger.info_once("Cudagraph is disabled under eager mode")
|
||||
self.compilation_config.cudagraph_mode = CUDAGraphMode.NONE
|
||||
# override related settings when enforce eager
|
||||
self.compilation_config.max_cudagraph_capture_size = 0
|
||||
@@ -1463,7 +1458,7 @@ class VllmConfig:
|
||||
and self.model_config.architecture == "WhisperForConditionalGeneration"
|
||||
and os.environ.get("VLLM_WORKER_MULTIPROC_METHOD") != "spawn"
|
||||
):
|
||||
logger.warning(
|
||||
logger.warning_once(
|
||||
"Whisper is known to have issues with "
|
||||
"forked workers. If startup is hanging, "
|
||||
"try setting 'VLLM_WORKER_MULTIPROC_METHOD' "
|
||||
@@ -1475,7 +1470,7 @@ class VllmConfig:
|
||||
and self.kv_events_config.enable_kv_cache_events
|
||||
and not self.cache_config.enable_prefix_caching
|
||||
):
|
||||
logger.warning(
|
||||
logger.warning_once(
|
||||
"KV cache events are on, but prefix caching is not enabled. "
|
||||
"Use --enable-prefix-caching to enable."
|
||||
)
|
||||
@@ -1484,7 +1479,7 @@ class VllmConfig:
|
||||
and self.kv_events_config.publisher != "null"
|
||||
and not self.kv_events_config.enable_kv_cache_events
|
||||
):
|
||||
logger.warning(
|
||||
logger.warning_once(
|
||||
"KV cache events are disabled, "
|
||||
"but the scheduler is configured to publish them. "
|
||||
"Modify KVEventsConfig.enable_kv_cache_events "
|
||||
@@ -1522,7 +1517,7 @@ class VllmConfig:
|
||||
# the pass will operate on higher-level IR to avoid the issue.
|
||||
# TODO: https://github.com/vllm-project/vllm/issues/27894
|
||||
if self.compilation_config.mode != CompilationMode.VLLM_COMPILE:
|
||||
logger.warning(
|
||||
logger.warning_once(
|
||||
"Sequence parallelism is enabled, but running in wrong "
|
||||
"vllm compile mode: %s.",
|
||||
self.compilation_config.mode,
|
||||
|
||||
@@ -107,7 +107,7 @@ class CustomAllreduce:
|
||||
if not custom_ar:
|
||||
# disable because of missing custom allreduce library
|
||||
# e.g. in a non-GPU environment
|
||||
logger.info(
|
||||
logger.info_once(
|
||||
"Custom allreduce is disabled because "
|
||||
"of missing custom allreduce library"
|
||||
)
|
||||
@@ -130,7 +130,7 @@ class CustomAllreduce:
|
||||
return
|
||||
|
||||
if world_size not in CustomAllreduce._SUPPORTED_WORLD_SIZES:
|
||||
logger.warning(
|
||||
logger.warning_once(
|
||||
"Custom allreduce is disabled due to an unsupported world"
|
||||
" size: %d. Supported world sizes: %s. To silence this "
|
||||
"warning, specify disable_custom_all_reduce=True explicitly.",
|
||||
@@ -328,7 +328,7 @@ class CustomAllreduce:
|
||||
|
||||
def register_graph_buffers(self):
|
||||
handle, offset = ops.get_graph_buffer_ipc_meta(self._ptr)
|
||||
logger.info("Registering %d cuda graph addresses", len(offset))
|
||||
logger.debug("Registering %d cuda graph addresses", len(offset))
|
||||
# We cannot directly use `dist.all_gather_object` here
|
||||
# because it is incompatible with `gloo` backend under inference mode.
|
||||
# see https://github.com/pytorch/pytorch/issues/126032 for details.
|
||||
|
||||
@@ -121,7 +121,7 @@ def _resolve_fi_ar_backend() -> tuple[str, bool]:
|
||||
"""
|
||||
backend = envs.VLLM_FLASHINFER_ALLREDUCE_BACKEND
|
||||
if backend != "auto":
|
||||
logger.info_once(f"Using flashinfer allreduce backend: {backend}")
|
||||
logger.debug_once("Using flashinfer allreduce backend: %s", backend)
|
||||
return backend, False
|
||||
|
||||
# Default to mnnvl for both single- and multi-node setups. The mnnvl
|
||||
@@ -134,7 +134,7 @@ def _resolve_fi_ar_backend() -> tuple[str, bool]:
|
||||
backend = "mnnvl"
|
||||
allow_trtllm_fallback = get_node_count() == 1
|
||||
|
||||
logger.info_once(f"Auto-selected flashinfer allreduce backend: {backend}")
|
||||
logger.debug_once("Auto-selected flashinfer allreduce backend: %s", backend)
|
||||
return backend, allow_trtllm_fallback
|
||||
|
||||
|
||||
|
||||
@@ -8,7 +8,6 @@
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import time
|
||||
import weakref
|
||||
from collections.abc import Callable, Hashable, Iterable
|
||||
from dataclasses import dataclass
|
||||
@@ -97,25 +96,17 @@ def _compile_cutedsl_warmup_units(
|
||||
def cutedsl_warmup() -> None:
|
||||
"""Run CuTeDSL compile providers before serving."""
|
||||
if not current_platform.is_cuda():
|
||||
logger.info("Skipping CuTeDSL warmup on non-CUDA platform.")
|
||||
logger.debug("Skipping CuTeDSL warmup on non-CUDA platform.")
|
||||
return
|
||||
|
||||
compile_units = _collect_unique_compile_units(_iter_cutedsl_warmup_compile_units())
|
||||
if not compile_units:
|
||||
logger.info("Skipping CuTeDSL warmup because no compile units were requested.")
|
||||
logger.debug("Skipping CuTeDSL warmup because no compile units were requested.")
|
||||
return
|
||||
|
||||
unit_names = list(dict.fromkeys(unit.name for unit in compile_units))
|
||||
logger.info(
|
||||
logger.info_once(
|
||||
"Warming up CuTeDSL compile_units=%d names=%s.",
|
||||
len(compile_units),
|
||||
unit_names,
|
||||
)
|
||||
|
||||
start_time = time.perf_counter()
|
||||
compiled_count = _compile_cutedsl_warmup_units(compile_units)
|
||||
logger.info(
|
||||
"CuTeDSL warmup compiled %d units in %.2f s.",
|
||||
compiled_count,
|
||||
time.perf_counter() - start_time,
|
||||
tuple(dict.fromkeys(unit.name for unit in compile_units)),
|
||||
)
|
||||
_compile_cutedsl_warmup_units(compile_units)
|
||||
|
||||
@@ -84,12 +84,12 @@ def _warmup_ll_bf16_router_gemm(model: torch.nn.Module) -> None:
|
||||
|
||||
shapes = _ll_bf16_router_shapes_from_model(model)
|
||||
if not shapes:
|
||||
logger.info(
|
||||
logger.debug_once(
|
||||
"Skipping ll_bf16 router GEMM warmup: no bf16 GateLinear shapes found."
|
||||
)
|
||||
return
|
||||
|
||||
logger.info("Warming up ll_bf16 router GEMM kernels for shapes: %s.", shapes)
|
||||
logger.info_once("Warming up ll_bf16 router GEMM kernels for shapes: %s.", shapes)
|
||||
ll_bf16_gemm_kernel.warmup(
|
||||
shapes=shapes,
|
||||
m_values=_LL_BF16_WARMUP_M_RANGE,
|
||||
@@ -163,7 +163,7 @@ def kernel_warmup(worker: "Worker", *, process_local_only: bool = False):
|
||||
)
|
||||
# FlashInfer autotune for Hopper (SM 9.0) and Blackwell (SM 10.0) GPUs
|
||||
if enable_flashinfer_autotune is False:
|
||||
logger.info("Skipping FlashInfer autotune because it is disabled.")
|
||||
logger.info_once("Skipping FlashInfer autotune because it is disabled.")
|
||||
elif has_flashinfer() and current_platform.has_device_capability(90):
|
||||
flashinfer_autotune(worker.model_runner)
|
||||
|
||||
@@ -188,7 +188,7 @@ def kernel_warmup(worker: "Worker", *, process_local_only: bool = False):
|
||||
for group in groups
|
||||
)
|
||||
):
|
||||
logger.info("Warming up FlashInfer attention.")
|
||||
logger.info_once("Warming up FlashInfer attention.")
|
||||
# Warmup with mixed batch containing both prefill and decode tokens
|
||||
# This is to warm up both prefill and decode attention kernels
|
||||
worker.model_runner._dummy_run(
|
||||
@@ -238,9 +238,9 @@ def flashinfer_autotune(runner: "GPUModelRunner") -> None:
|
||||
autotune_kwargs: dict = {}
|
||||
skip_ops = _flashinfer_autotune_skip_ops(runner)
|
||||
if skip_ops:
|
||||
logger.info(
|
||||
logger.info_once(
|
||||
"Skipping FlashInfer autotuning for ops %s",
|
||||
sorted(skip_ops),
|
||||
tuple(sorted(skip_ops)),
|
||||
)
|
||||
autotune_kwargs["skip_ops"] = skip_ops
|
||||
|
||||
@@ -265,7 +265,7 @@ def flashinfer_autotune(runner: "GPUModelRunner") -> None:
|
||||
|
||||
cache_path = resolve_flashinfer_autotune_file(runner)
|
||||
if is_leader:
|
||||
logger.info("Using FlashInfer autotune cache file: %s", cache_path)
|
||||
logger.info_once("Using FlashInfer autotune cache file: %s", cache_path)
|
||||
|
||||
# We skip EPLB here since we don't want to record dummy metrics.
|
||||
# When autotuning with number of tokens m, flashinfer will autotune
|
||||
@@ -296,9 +296,9 @@ def flashinfer_autotune(runner: "GPUModelRunner") -> None:
|
||||
tune_results = world.broadcast_object(tune_results, src=0)
|
||||
|
||||
if tune_results is None:
|
||||
logger.warning(
|
||||
"No FlashInfer autotune cache entries found."
|
||||
"Falling back to default tactics."
|
||||
logger.warning_once(
|
||||
"No FlashInfer autotune cache entries found; "
|
||||
"falling back to default tactics."
|
||||
)
|
||||
else:
|
||||
write_flashinfer_autotune_cache(cache_path, tune_results)
|
||||
@@ -306,7 +306,7 @@ def flashinfer_autotune(runner: "GPUModelRunner") -> None:
|
||||
from flashinfer.autotuner import AutoTuner
|
||||
|
||||
AutoTuner.get().load_configs(str(cache_path))
|
||||
logger.info(
|
||||
logger.info_once(
|
||||
"FlashInfer autotune cache loaded on rank %d from %s.",
|
||||
world.rank_in_group,
|
||||
cache_path,
|
||||
|
||||
@@ -284,7 +284,7 @@ def kill_process_tree(pid: int):
|
||||
# Adapted from: https://github.com/sgl-project/sglang/blob/v0.4.1/python/sglang/srt/utils.py#L630
|
||||
def set_ulimit(target_soft_limit: int = 65535):
|
||||
if sys.platform.startswith("win"):
|
||||
logger.info("Windows detected, skipping ulimit adjustment.")
|
||||
logger.debug("Windows detected, skipping ulimit adjustment.")
|
||||
return
|
||||
|
||||
import resource
|
||||
|
||||
@@ -87,7 +87,7 @@ def get_kv_cache_layout():
|
||||
cache_layout: Literal["NHD", "HND"] | None = None
|
||||
if _KV_CACHE_LAYOUT_OVERRIDE is not None:
|
||||
cache_layout = _KV_CACHE_LAYOUT_OVERRIDE
|
||||
logger.info_once(
|
||||
logger.debug_once(
|
||||
"`_KV_CACHE_LAYOUT_OVERRIDE` variable detected. "
|
||||
"Setting KV cache layout to %s.",
|
||||
cache_layout,
|
||||
|
||||
@@ -199,7 +199,7 @@ def _cached_get_attn_backend(
|
||||
from vllm.v1.attention.backends.utils import set_kv_cache_layout
|
||||
|
||||
set_kv_cache_layout(required_layout)
|
||||
logger.info(
|
||||
logger.info_once(
|
||||
"Using %s KV cache layout for %s backend.",
|
||||
required_layout,
|
||||
backend.get_name(),
|
||||
|
||||
@@ -305,7 +305,7 @@ class GPUModelRunner(LoRAModelRunnerMixin):
|
||||
eplb_models_added = False
|
||||
with DeviceMemoryProfiler() as m:
|
||||
model_loader = get_model_loader(self.vllm_config.load_config)
|
||||
logger.info("Loading model from scratch...")
|
||||
logger.info_once("Loading model from scratch...")
|
||||
|
||||
self.model = model_loader.load_model(
|
||||
vllm_config=self.vllm_config, model_config=self.vllm_config.model_config
|
||||
|
||||
Reference in New Issue
Block a user