A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 6, 12:00 UTC
[Log] Reduce duplicate log (#37313)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Wentao Ye committed
Mar 18, 2026 at 10:57 UTC
c373b5c00d1a6f0830099ce5c4b5276e70bc6388
8 files changed
+20
-10
vllm/compilation/backends.py
+3
-1
@@ -371,13 +371,15 @@ class CompilerManager:
371
logger.info_once(
372
"Cache the graph of compile range %s for later use",
373
str(compile_range),
374
+ scope="local",
375
)
375
- logger.debug(
376
+ logger.debug_once(
377
"Store the %s-th graph for compile range%s from %s via handle %s",
378
graph_index,
379
str(compile_range),
380
self.compiler.name,
381
handle,
382
+ scope="local",
383
)
384
385
# after compiling the last graph, record the end time
vllm/config/scheduler.py
+2
-1
@@ -228,9 +228,10 @@ class SchedulerConfig:
228
self.encoder_cache_size = self.max_num_batched_tokens
229
230
if self.enable_chunked_prefill:
231
- logger.info(
231
+ logger.info_once(
232
"Chunked prefill is enabled with max_num_batched_tokens=%d.",
233
self.max_num_batched_tokens,
234
+ scope="local",
235
)
236
237
if self.max_num_partial_prefills > 1:
vllm/model_executor/layers/attention/mm_encoder_attention.py
+3
-1
@@ -227,7 +227,9 @@ class MMEncoderAttention(CustomOp):
227
if self.attn_backend == AttentionBackendEnum.FLASHINFER:
228
_get_flashinfer_workspace_buffer()
229
230
- logger.info_once(f"Using {self.attn_backend} for MMEncoderAttention.")
230
+ logger.info_once(
231
+ f"Using {self.attn_backend} for MMEncoderAttention.", scope="local"
232
+ )
233
234
@classmethod
235
def enabled(cls) -> bool:
vllm/model_executor/models/qwen3_next.py
+4
-3
@@ -192,14 +192,15 @@ class ChunkGatedDeltaRule(CustomOp):
192
use_flashinfer = supports_flashinfer
193
194
if use_flashinfer:
195
- logger.info_once("Using FlashInfer GDN prefill kernel")
195
+ logger.info_once("Using FlashInfer GDN prefill kernel", scope="local")
196
logger.info_once(
197
"FlashInfer GDN prefill kernel is JIT-compiled; first run may "
198
"take a while to compile. Set `--gdn-prefill-backend triton` to "
199
- "avoid JIT compile time."
199
+ "avoid JIT compile time.",
200
+ scope="local",
201
)
202
else:
202
- logger.info_once("Using Triton/FLA GDN prefill kernel")
203
+ logger.info_once("Using Triton/FLA GDN prefill kernel", scope="local")
204
205
self._forward_method = (
206
self.forward_cuda if use_flashinfer else self.forward_native
vllm/platforms/cuda.py
+2
-1
@@ -387,7 +387,8 @@ class CudaPlatformBase(Platform):
387
)
388
if is_backend_supported:
389
logger.info_once(
390
- f"Using backend {vit_attn_backend} for vit attention"
390
+ f"Using backend {vit_attn_backend} for vit attention",
391
+ scope="local",
392
)
393
return vit_attn_backend
394
except ImportError:
vllm/v1/executor/multiproc_executor.py
+2
-1
@@ -998,12 +998,13 @@ def set_multiprocessing_worker_envs():
998
"OMP_NUM_THREADS" not in os.environ
999
and (current_parallelism := torch.get_num_threads()) > default_omp_num_threads
1000
):
1001
- logger.warning(
1001
+ logger.warning_once(
1002
"Reducing Torch parallelism from %d threads to %d to avoid "
1003
"unnecessary CPU contention. Set OMP_NUM_THREADS in the "
1004
"external environment to tune this value as needed.",
1005
current_parallelism,
1006
default_omp_num_threads,
1007
+ scope="local",
1008
)
1009
os.environ["OMP_NUM_THREADS"] = str(default_omp_num_threads)
1010
torch.set_num_threads(default_omp_num_threads)
vllm/v1/worker/dp_utils.py
+2
-1
@@ -28,7 +28,8 @@ def _get_device_and_group(parallel_config: ParallelConfig):
28
# this optimization if we run into this case.
29
if parallel_config.disable_nccl_for_dp_synchronization:
30
logger.info_once(
31
- "Using CPU all reduce to synchronize DP padding between ranks."
31
+ "Using CPU all reduce to synchronize DP padding between ranks.",
32
+ scope="local",
33
)
34
device = "cpu"
35
group = get_dp_group().cpu_group
vllm/v1/worker/gpu_model_runner.py
+2
-1
@@ -5510,13 +5510,14 @@ class GPUModelRunner(
5510
dummy_modality
5511
]
5512
5513
- logger.info(
5513
+ logger.info_once(
5514
"Encoder cache will be initialized with a "
5515
"budget of %s tokens, and profiled with "
5516
"%s %s items of the maximum feature size.",
5517
encoder_budget,
5518
max_mm_items_per_batch,
5519
dummy_modality,
5520
+ scope="local",
5521
)
5522
5523
# Create dummy batch of multimodal inputs.