obscura/vllm
2
0
Fork 0
mirror of https://github.com/vllm-project/vllm.git synced 2026-08-17 19:20:14 +00:00
Code Issues Packages Projects Releases Wiki Activity

[Refactor] Clean up log once scope="local" (#40540)

Browse Source 
Signed-off-by: yewentao256 <[email protected]>
Signed-off-by: Wentao Ye <[email protected]>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
This commit is contained in:
Wentao Ye
2026-04-22 16:42:43 -04:00
committed by GitHub
co-authored by gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
parent cfa49213d7
commit 8f87eb4622
44 changed files with 56 additions and 148 deletions
Download Patch File Download Diff File Expand all files Collapse all files
vllm/compilation/backends.py
+3 -13
View File
@@ -292,7 +292,6 @@ class CompilerManager:
"from the cache, took %.3f s",
str(compile_range),
elapsed,
scope="local",
)
return compiled_graph
@@ -377,7 +376,6 @@ class CompilerManager:
logger.info_once(
"Cache the graph of compile range %s for later use",
str(compile_range),
scope="local",
)
logger.debug_once(
"Store the %s-th graph for compile range%s from %s via handle %s",
@@ -385,7 +383,6 @@ class CompilerManager:
str(compile_range),
self.compiler.name,
handle,
scope="local",
)
# after compiling the last graph, record the end time
@@ -399,7 +396,6 @@ class CompilerManager:
"Compiling a graph for compile range %s takes %.2f s",
str(compile_range),
elapsed,
scope="local",
)
return compiled_graph
@@ -1072,12 +1068,11 @@ class VllmBackend:
disable_cache = disable_cache or is_ngram_gpu_enabled
if disable_cache:
logger.info_once("vLLM's torch.compile cache is disabled.", scope="local")
logger.info_once("vLLM's torch.compile cache is disabled.")
else:
logger.info_once(
"Using cache directory: %s for vLLM's torch.compile",
local_cache_dir,
scope="local",
)
self.compiler_manager.initialize_cache(
@@ -1134,9 +1129,7 @@ class VllmBackend:
from .monitor import torch_compile_start_time
dynamo_time = time.perf_counter() - torch_compile_start_time
logger.info_once(
"Dynamo bytecode transform time: %.2f s", dynamo_time, scope="local"
)
logger.info_once("Dynamo bytecode transform time: %.2f s", dynamo_time)
if self.is_encoder:
self.compilation_config.encoder_compilation_time += dynamo_time
else:
@@ -1215,7 +1208,6 @@ class VllmBackend:
logger.info_once(
"Saved compiler manager cache in %.2f seconds.",
elapsed,
scope="local",
)
from torch._guards import detect_fake_mode
@@ -1254,9 +1246,7 @@ class VllmBackend:
with open(graph_path, "w") as f:
f.write(src)
logger.debug_once(
"Computation graph saved to %s", graph_path, scope="local"
)
logger.debug_once("Computation graph saved to %s", graph_path)
self._called = True
graph_to_serialize = (
vllm/compilation/decorators.py
-1
View File
@@ -665,7 +665,6 @@ def _support_torch_compile(
logger.info_once(
"saved AOT compiled function to %s",
self._aot_compilation_path,
scope="local",
)
except Exception as e:
logger.warning(
vllm/compilation/monitor.py
+1 -2
View File
@@ -45,7 +45,7 @@ def monitor_torch_compile(
else:
total_compile_time = time.perf_counter() - torch_compile_start_time
if compilation_config.mode == CompilationMode.VLLM_COMPILE:
logger.info_once(message, total_compile_time, scope="local")
logger.info_once(message, total_compile_time)
finally:
if depyf_cm is not None:
try:
@@ -76,7 +76,6 @@ def monitor_profiling_run() -> Generator[None, None, None]:
logger.info_once(
"Initial profiling/warmup run took %.2f s",
elapsed,
scope="local",
)
vllm/config/scheduler.py
-1
View File
@@ -239,7 +239,6 @@ class SchedulerConfig:
logger.info_once(
"Chunked prefill is enabled with max_num_batched_tokens=%d.",
self.max_num_batched_tokens,
scope="local",
)
if self.max_num_partial_prefills > 1:
vllm/config/vllm.py
+1 -10
View File
@@ -716,9 +716,7 @@ class VllmConfig:
self.instance_id = f"{time.time_ns()}"
if self.performance_mode != "balanced":
logger.info_once(
"Performance mode set to '%s'.", self.performance_mode, scope="local"
)
logger.info_once("Performance mode set to '%s'.", self.performance_mode)
self.try_verify_and_update_config()
@@ -818,7 +816,6 @@ class VllmConfig:
"Async scheduling not supported with %s-based "
"speculative decoding and will be disabled.",
self.speculative_config.method,
scope="local",
)
self.scheduler_config.async_scheduling = False
elif (
@@ -828,7 +825,6 @@ class VllmConfig:
logger.warning_once(
"Async scheduling is not compatible with "
"disable_padded_drafter_batch=True and will be disabled.",
scope="local",
)
self.scheduler_config.async_scheduling = False
elif not executor_supports_async_sched:
@@ -836,7 +832,6 @@ class VllmConfig:
"Async scheduling will be disabled because it is not supported "
"with the `%s` distributed executor backend. ",
executor_backend,
scope="local",
)
self.scheduler_config.async_scheduling = False
else:
@@ -855,7 +850,6 @@ class VllmConfig:
logger.info_once(
"Disabling NCCL for DP synchronization "
"when using async scheduling.",
scope="local",
)
self.parallel_config.disable_nccl_for_dp_synchronization = True
else:
@@ -870,7 +864,6 @@ class VllmConfig:
logger.warning_once(
"Disabling cascade attention (not yet compatible with "
"async speculative decoding).",
scope="local",
)
self.model_config.disable_cascade_attn = True
@@ -1231,7 +1224,6 @@ class VllmConfig:
self.model_config.disable_cascade_attn = True
logger.warning_once(
"Disabling cascade attention when VLLM_BATCH_INVARIANT is enabled.",
scope="local",
)
if self.parallel_config.use_ubatching:
@@ -1418,7 +1410,6 @@ class VllmConfig:
" performance. Consider increasing max_num_batched_tokens to"
" accommodate the additional draft token slots, or decrease"
" num_speculative_tokens or max_num_seqs.",
scope="local",
)
max_num_scheduled_tokens = self.scheduler_config.max_num_scheduled_tokens
vllm/distributed/device_communicators/pynccl.py
+1 -3
View File
@@ -108,9 +108,7 @@ class PyNcclCommunicator:
if self.rank == 0:
# get the unique id from NCCL
self.unique_id = self.nccl.ncclGetUniqueId()
logger.info_once(
"vLLM is using nccl==%s", self.nccl.ncclGetVersion(), scope="local"
)
logger.info_once("vLLM is using nccl==%s", self.nccl.ncclGetVersion())
else:
# construct an empty unique id
self.unique_id = ncclUniqueId()
vllm/engine/arg_utils.py
-3
View File
@@ -2254,7 +2254,6 @@ class EngineArgs:
"This model does not officially support disabling chunked prefill. "
"Disabling this manually may cause the engine to crash "
"or produce incorrect outputs.",
scope="local",
)
elif (
model_config.runner_type == "pooling"
@@ -2265,7 +2264,6 @@ class EngineArgs:
"This model does not officially support chunked prefill. "
"Enabling this manually may cause the engine to crash "
"or produce incorrect outputs.",
scope="local",
)
if self.enable_prefix_caching is None:
@@ -2284,7 +2282,6 @@ class EngineArgs:
"This model does not officially support prefix caching. "
"Enabling this manually may cause the engine to crash "
"or produce incorrect outputs.",
scope="local",
)
# Disable chunked prefill and prefix caching for:
vllm/lora/model_manager.py
-1
View File
@@ -387,7 +387,6 @@ class LoRAModelManager:
"LoRA is not supported for non-gated MoE gate module."
" %s will be ignored.",
module_name,
scope="local",
)
continue
vllm/model_executor/layers/attention/attention.py
-1
View File
@@ -332,7 +332,6 @@ class Attention(nn.Module, AttentionLayerBase):
logger.warning_once(
"Disabling prefix caching for FLASHINFER/TRITON_MLA "
"with batch invariance, as it is not yet supported.",
scope="local",
)
cache_config.enable_prefix_caching = False
vllm/model_executor/layers/attention/mla_attention.py
+5 -12
View File
@@ -427,7 +427,6 @@ class MLAAttention(nn.Module, AttentionLayerBase):
logger.warning_once(
"Disabling prefix caching for TRITON_MLA / FLASHINFER "
"with batch invariance, as it is not yet supported.",
scope="local",
)
cache_config.enable_prefix_caching = False
@@ -1523,9 +1522,7 @@ class MLACommonMetadataBuilder(AttentionMetadataBuilder[M]):
if use_fp8:
fp8_dtype = current_platform.fp8_dtype()
logger.info_once(
"FP8 prefill attention enabled: query data type is FP8", scope="local"
)
logger.info_once("FP8 prefill attention enabled: query data type is FP8")
return fp8_dtype
elif vllm_config.attention_config.use_prefill_query_quantization:
logger.info_once(
@@ -1533,7 +1530,6 @@ class MLACommonMetadataBuilder(AttentionMetadataBuilder[M]):
" use_prefill_query_quantization is enabled. Please"
" ensure that --kv-cache-dtype is set to fp8 and your prefill"
" backend is compatible with FP8 attention.",
scope="local",
)
return model_dtype
elif (
@@ -1547,7 +1543,6 @@ class MLACommonMetadataBuilder(AttentionMetadataBuilder[M]):
"prefill latency. To enable, add: "
'--attention-config \'{"use_prefill_query_quantization"'
": true}'",
scope="local",
)
return model_dtype
@@ -2225,21 +2220,19 @@ class MLACommonImpl(MLAAttentionImpl[M], Generic[M]):
)
if use_trtllm_ragged_deepseek_prefill():
logger.info_once(
"Using TRT-LLM ragged DeepSeek prefill for MLA", scope="local"
)
logger.info_once("Using TRT-LLM ragged DeepSeek prefill for MLA")
self._run_prefill_context_chunk = (
self._run_prefill_context_chunk_trtllm_ragged
)
self._run_prefill_new_tokens = self._run_prefill_new_tokens_trtllm_ragged
self._pad_v = False
elif use_flashinfer_prefill():
logger.info_once("Using FlashInfer prefill for MLA", scope="local")
logger.info_once("Using FlashInfer prefill for MLA")
self._run_prefill_context_chunk = self._run_prefill_context_chunk_fi
self._run_prefill_new_tokens = self._run_prefill_new_tokens_fi
self._pad_v = False
elif use_cudnn_prefill():
logger.info_once("Using CUDNN prefill for MLA", scope="local")
logger.info_once("Using CUDNN prefill for MLA")
self._run_prefill_context_chunk = self._run_prefill_context_chunk_cudnn
self._run_prefill_new_tokens = self._run_prefill_new_tokens_cudnn
self._pad_v = False
@@ -2250,7 +2243,7 @@ class MLACommonImpl(MLAAttentionImpl[M], Generic[M]):
"available. Please install flash_attn or use "
"--attention-backend ROCM_AITER_MLA."
)
logger.info_once("Using FlashAttention prefill for MLA", scope="local")
logger.info_once("Using FlashAttention prefill for MLA")
self._run_prefill_context_chunk = self._run_prefill_context_chunk_fa
self._run_prefill_new_tokens = self._run_prefill_new_tokens_fa
vllm/model_executor/layers/attention/mm_encoder_attention.py
+1 -3
View File
@@ -227,9 +227,7 @@ class MMEncoderAttention(CustomOp):
if self.attn_backend == AttentionBackendEnum.FLASHINFER:
_get_flashinfer_workspace_buffer()
logger.info_once(
f"Using {self.attn_backend} for MMEncoderAttention.", scope="local"
)
logger.info_once(f"Using {self.attn_backend} for MMEncoderAttention.")
@classmethod
def enabled(cls) -> bool:
vllm/model_executor/layers/batch_invariant.py
+1 -1
View File
@@ -1020,7 +1020,7 @@ def override_envs_for_invariance(
"You are using a non-decode-invariant form of batch invariance. "
"This will not be invariant between prefill and decode."
)
logger.warning_once(warning, scope="local")
logger.warning_once(warning)
os.environ["VLLM_ALLREDUCE_USE_SYMM_MEM"] = "0"
os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":4096:8"
vllm/model_executor/layers/fused_moe/experts/batched_deep_gemm_moe.py
-1
View File
@@ -369,7 +369,6 @@ class BatchedDeepGemmExperts(mk.FusedMoEExpertsModular):
logger.warning_once(
"DPMetadata unavailable. Defaulting expected_m to "
f"{max_tokens_per_expert}.",
scope="local",
)
return max_tokens_per_expert
vllm/model_executor/layers/fused_moe/fused_moe.py
-1
View File
@@ -1091,7 +1091,6 @@ def get_moe_configs(
"Using default MoE config. Performance might be sub-optimal! "
"Config file not found at %s",
", ".join(config_file_paths),
scope="local",
)
return None
vllm/model_executor/layers/fused_moe/nixl_ep_prepare_finalize.py
-1
View File
@@ -123,7 +123,6 @@ class NixlEPPrepareAndFinalize(mk.FusedMoEPrepareAndFinalizeModular):
"NixlEPPrepareAndFinalize is setup to dispatch raw/unquantized "
f"activations despite ({fused_experts.__class__.__name__}) being able "
"to support quantized activations.",
scope="local",
)
def num_dispatchers(self) -> int:
vllm/model_executor/layers/fused_moe/oracle/fp8.py
+6 -8
View File
@@ -266,7 +266,7 @@ def select_fp8_moe_backend(
k_cls, config, weight_key, activation_key, activation_format
)
if supported:
logger.info_once(_make_log_backend(backend), scope="local")
logger.info_once(_make_log_backend(backend))
return backend, k_cls
raise ValueError(_make_log_unsupported(backend, reason))
@@ -337,12 +337,10 @@ def select_fp8_moe_backend(
)
if supported:
logger.info_once(_make_log_backend(backend), scope="local")
logger.info_once(_make_log_backend(backend))
return backend, k_cls
else:
logger.debug_once(
_make_log_unsupported(backend, reason), scope="local"
)
logger.debug_once(_make_log_unsupported(backend, reason))
raise NotImplementedError(
"Found VLLM_USE_FLASHINFER_MOE_FP8=1, but no "
@@ -396,10 +394,10 @@ def select_fp8_moe_backend(
activation_format,
)
if supported:
logger.info_once(_make_log_backend(backend), scope="local")
logger.info_once(_make_log_backend(backend))
return backend, k_cls
else:
logger.debug_once(_make_log_unsupported(backend, reason), scope="local")
logger.debug_once(_make_log_unsupported(backend, reason))
# TODO(rob): per discussion with TPU team, we need a way to register
# MoE backends by OOT plugins, rather than having an explicit list
@@ -580,7 +578,7 @@ def make_fp8_moe_kernel(
)
assert prepare_finalize is not None
logger.info_once("Using %s", prepare_finalize.__class__.__name__, scope="local")
logger.info_once("Using %s", prepare_finalize.__class__.__name__)
# Create Experts.
if prepare_finalize.activation_format == mk.FusedMoEActivationFormat.BatchedExperts:
vllm/model_executor/layers/fused_moe/oracle/int8.py
+4 -4
View File
@@ -117,7 +117,7 @@ def select_int8_moe_backend(
k_cls, config, weight_key, activation_key, activation_format
)
if supported:
logger.info_once(_make_log_backend(backend), scope="local")
logger.info_once(_make_log_backend(backend))
return backend, k_cls
raise ValueError(_make_log_unsupported(backend, reason))
@@ -138,10 +138,10 @@ def select_int8_moe_backend(
activation_format,
)
if supported:
logger.info_once(_make_log_backend(backend), scope="local")
logger.info_once(_make_log_backend(backend))
return backend, k_cls
else:
logger.debug_once(_make_log_unsupported(backend, reason), scope="local")
logger.debug_once(_make_log_unsupported(backend, reason))
raise NotImplementedError(
"No Int8 MoE backend supports the deployment configuration."
@@ -193,7 +193,7 @@ def make_int8_moe_kernel(
)
assert prepare_finalize is not None
logger.info_once("Using %s", prepare_finalize.__class__.__name__, scope="local")
logger.info_once("Using %s", prepare_finalize.__class__.__name__)
# Create Experts.
if prepare_finalize.activation_format == mk.FusedMoEActivationFormat.BatchedExperts:
vllm/model_executor/layers/fused_moe/oracle/mxfp4.py
+4 -4
View File
@@ -269,7 +269,7 @@ def select_gpt_oss_mxfp4_moe_backend(
k_cls, config, weight_key, activation_key, activation_format
)
if supported:
logger.info_once(_make_log_backend(backend), scope="local")
logger.info_once(_make_log_backend(backend))
return backend, k_cls
raise ValueError(_make_log_unsupported(backend, reason))
@@ -363,10 +363,10 @@ def select_gpt_oss_mxfp4_moe_backend(
k_cls, config, kMxfp4Static, activation_key, activation_format
)
if supported:
logger.info_once(_make_log_backend(backend), scope="local")
logger.info_once(_make_log_backend(backend))
return backend, k_cls
else:
logger.debug_once(_make_log_unsupported(backend, reason), scope="local")
logger.debug_once(_make_log_unsupported(backend, reason))
if current_platform.is_xpu():
backend = Mxfp4MoeBackend.XPU
@@ -861,7 +861,7 @@ def make_mxfp4_moe_kernel(
)
assert prepare_finalize is not None
logger.info_once("Using %s", prepare_finalize.__class__.__name__, scope="local")
logger.info_once("Using %s", prepare_finalize.__class__.__name__)
# Create Experts.
if prepare_finalize.activation_format == mk.FusedMoEActivationFormat.BatchedExperts:
vllm/model_executor/layers/fused_moe/oracle/nvfp4.py
+4 -6
View File
@@ -252,12 +252,10 @@ def select_nvfp4_moe_backend(
activation_format,
)
if supported:
logger.info_once(_make_log_backend(backend), scope="local")
logger.info_once(_make_log_backend(backend))
return backend, k_cls
else:
logger.debug_once(
_make_log_unsupported(backend, reason), scope="local"
)
logger.debug_once(_make_log_unsupported(backend, reason))
raise NotImplementedError(
"Found VLLM_USE_FLASHINFER_MOE_FP4=1, but no "
@@ -282,10 +280,10 @@ def select_nvfp4_moe_backend(
)
if supported:
logger.info_once(_make_log_backend(backend), scope="local")
logger.info_once(_make_log_backend(backend))
return backend, k_cls
else:
logger.debug_once(_make_log_unsupported(backend, reason), scope="local")
logger.debug_once(_make_log_unsupported(backend, reason))
raise NotImplementedError(
"No NvFp4 MoE backend supports the deployment configuration."
vllm/model_executor/layers/fused_moe/oracle/unquantized.py
+6 -8
View File
@@ -210,7 +210,7 @@ def select_unquantized_moe_backend(
k_cls, config, None, None, activation_format
)
if supported:
logger.info_once(_make_log_backend(backend), scope="local")
logger.info_once(_make_log_backend(backend))
return backend, k_cls
raise ValueError(_make_log_unsupported(backend, reason))
@@ -271,12 +271,10 @@ def select_unquantized_moe_backend(
k_cls, moe_config, None, None, activation_format
)
if supported:
logger.info_once(_make_log_backend(backend), scope="local")
logger.info_once(_make_log_backend(backend))
return backend, k_cls
else:
logger.debug_once(
_make_log_unsupported(backend, reason), scope="local"
)
logger.debug_once(_make_log_unsupported(backend, reason))
raise NotImplementedError(
"Found VLLM_USE_FLASHINFER_MOE_FP16=1, but no "
@@ -298,10 +296,10 @@ def select_unquantized_moe_backend(
k_cls, moe_config, None, None, activation_format
)
if supported:
logger.info_once(_make_log_backend(backend), scope="local")
logger.info_once(_make_log_backend(backend))
return backend, k_cls
logger.debug_once(_make_log_unsupported(backend, reason), scope="local")
logger.debug_once(_make_log_unsupported(backend, reason))
raise NotImplementedError(
"No Unquantized MoE backend supports the deployment configuration."
@@ -355,7 +353,7 @@ def make_unquantized_moe_kernel(
)
assert prepare_finalize is not None
logger.info_once("Using %s", prepare_finalize.__class__.__name__, scope="local")
logger.info_once("Using %s", prepare_finalize.__class__.__name__)
# Create Experts
if prepare_finalize.activation_format == mk.FusedMoEActivationFormat.BatchedExperts:
vllm/model_executor/layers/fused_moe/prepare_finalize/deepep_ll.py
-1
View File
@@ -135,7 +135,6 @@ class DeepEPLLPrepareAndFinalize(mk.FusedMoEPrepareAndFinalizeModular):
"DeepEPLLPrepareAndFinalize is setup to dispatch raw/unquantized "
f"activations despite ({fused_experts.__class__.__name__}) being able "
"to support quantized activations.",
scope="local",
)
def num_dispatchers(self) -> int:
vllm/model_executor/layers/fused_moe/runner/shared_experts.py
+2 -4
View File
@@ -69,16 +69,14 @@ class SharedExperts:
# TODO: Remove this after more extensive testings with TP/DP
# and other execution modes
if envs.VLLM_DISABLE_SHARED_EXPERTS_STREAM:
logger.debug_once("Disabling MoE shared_experts cuda stream", scope="local")
logger.debug_once("Disabling MoE shared_experts cuda stream")
self._stream = None
else:
# TODO(rob): enable shared expert overlap with non-cuda-alike.
# aux_stream() returns None on non-cuda-alike platforms.
self._stream = aux_stream()
if self._stream is not None:
logger.debug_once(
"Enabled separate cuda stream for MoE shared_experts", scope="local"
)
logger.debug_once("Enabled separate cuda stream for MoE shared_experts")
@property
def _disable_shared_experts_overlap(self) -> bool:
vllm/model_executor/layers/mamba/gdn_linear_attn.py
+2 -3
View File
@@ -143,15 +143,14 @@ class ChunkGatedDeltaRule(CustomOp):
use_flashinfer = supports_flashinfer
if use_flashinfer:
logger.info_once("Using FlashInfer GDN prefill kernel", scope="local")
logger.info_once("Using FlashInfer GDN prefill kernel")
logger.info_once(
"FlashInfer GDN prefill kernel is JIT-compiled; first run may "
"take a while to compile. Set `--gdn-prefill-backend triton` to "
"avoid JIT compile time.",
scope="local",
)
else:
logger.info_once("Using Triton/FLA GDN prefill kernel", scope="local")
logger.info_once("Using Triton/FLA GDN prefill kernel")
self._forward_method = (
self.forward_cuda if use_flashinfer else self.forward_native
vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe/compressed_tensors_moe_w4a4_mxfp4.py
+2 -2
View File
@@ -44,10 +44,10 @@ class CompressedTensorsW4A4Mxfp4MoEMethod(CompressedTensorsMoEMethod):
self.use_cutlass_mxfp4 = CutlassExpertsMxfp4._supports_current_device()
self.experts_cls: type[mk.FusedMoEExperts]
if self.use_cutlass_mxfp4:
logger.info_once("Using CutlassExpertsMxfp4 for MXFP4 MoE", scope="local")
logger.info_once("Using CutlassExpertsMxfp4 for MXFP4 MoE")
self.experts_cls = CutlassExpertsMxfp4
else:
logger.info_once("Using MarlinExperts for MXFP4 MoE", scope="local")
logger.info_once("Using MarlinExperts for MXFP4 MoE")
self.experts_cls = MarlinExperts
def create_weights(
vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe/compressed_tensors_moe_wna16_marlin.py
-1
View File
@@ -87,7 +87,6 @@ class CompressedTensorsWNA16MarlinMoEMethod(CompressedTensorsMoEMethod):
logger.info_once(
f"Using {self.kernel_backend} backend for WNA16 MoE "
f"(group_size={self.group_size}, num_bits={self.num_bits})",
scope="local",
)
def get_weight_shape(
vllm/model_executor/layers/quantization/mxfp4.py
-2
View File
@@ -83,7 +83,6 @@ class Mxfp4Config(QuantizationConfig):
logger.debug_once(
"MXFP4 linear layer is not implemented - falling back to "
"UnquantizedLinearMethod.",
scope="local",
)
return UnquantizedLinearMethod()
elif isinstance(layer, FusedMoE):
@@ -92,7 +91,6 @@ class Mxfp4Config(QuantizationConfig):
logger.debug_once(
"MXFP4 attention layer is not implemented. "
"Skipping quantization for this layer.",
scope="local",
)
return None
vllm/model_executor/layers/quantization/utils/flashinfer_utils.py
-4
View File
@@ -111,7 +111,6 @@ def get_flashinfer_moe_backend() -> FlashinferMoeBackend:
logger.info_once(
"Flashinfer TRTLLM MOE backend is only supported on "
"SM100 and later, using CUTLASS backend instead",
scope="local",
)
return FlashinferMoeBackend.CUTLASS
return backend_map[flashinfer_moe_backend]
@@ -239,7 +238,6 @@ def align_fp4_moe_weights_for_fi(
"Padding intermediate size from %d to %d for up/down projection weights.",
intermediate,
padded_intermediate,
scope="local",
)
up_mult = 2 if is_act_and_mul else 1
@@ -285,7 +283,6 @@ def align_trtllm_fp4_moe_hidden_dim_for_fi(
"performance degradation.",
hidden_size,
padded_hidden_size,
scope="local",
)
padded_w13 = w13.new_zeros((num_experts, gate_up_dim, padded_hidden_size // 2))
@@ -331,7 +328,6 @@ def align_fp8_moe_weights_for_fi(
"Padding intermediate size from %d to %d for up/down projection weights.",
intermediate,
padded_intermediate,
scope="local",
)
up_mult = 2 if is_act_and_mul else 1
vllm/model_executor/model_loader/base_loader.py
-1
View File
@@ -70,7 +70,6 @@ class BaseModelLoader(ABC):
logger.debug_once(
"Peak GPU memory after loading weights: %s GiB",
format_gib(peak_memory),
scope="local",
)
# Process weights into kernel format. Note that when using online
vllm/model_executor/model_loader/default_loader.py
-1
View File
@@ -384,7 +384,6 @@ class DefaultModelLoader(BaseModelLoader):
logger.info_once(
"Loading weights took %.2f seconds",
self.counter_after_loading_weights - self.counter_before_loading_weights,
scope="local",
)
# We only enable strict check for non-quantized models
# that have loaded weights tracking currently.
vllm/model_executor/model_loader/sharded_state_loader.py
-1
View File
@@ -157,7 +157,6 @@ class ShardedStateLoader(BaseModelLoader):
logger.info_once(
"Loading weights took %.2f seconds",
counter_after_loading_weights - counter_before_loading_weights,
scope="local",
)
if state_dict:
raise ValueError(f"Missing keys {tuple(state_dict)} in loaded state!")
vllm/model_executor/offloader/base.py
+2 -4
View File
@@ -118,11 +118,9 @@ def set_offloader(instance: BaseOffloader) -> None:
global _instance
_instance = instance
if isinstance(instance, NoopOffloader):
logger.debug_once(
"Offloader set to NoopOffloader (no offloading).", scope="local"
)
logger.debug_once("Offloader set to NoopOffloader (no offloading).")
else:
logger.info_once("Offloader set to %s", type(instance).__name__, scope="local")
logger.info_once("Offloader set to %s", type(instance).__name__)
def create_offloader(offload_config: "OffloadConfig") -> BaseOffloader:
vllm/platforms/cuda.py
-2
View File
@@ -369,7 +369,6 @@ class CudaPlatformBase(Platform):
"Using %s attention backend out of potential backends: %s.",
selected_backend.name,
"[" + ", ".join(f"'{b[0].name}'" for b in valid_backends_priorities) + "]",
scope="local",
)
return selected_backend.get_path()
@@ -423,7 +422,6 @@ class CudaPlatformBase(Platform):
if is_backend_supported:
logger.info_once(
f"Using backend {vit_attn_backend} for vit attention",
scope="local",
)
return vit_attn_backend
except ImportError:
vllm/profiler/wrapper.py
+4 -8
View File
@@ -63,7 +63,7 @@ class WorkerProfiler(ABC):
"""Call _stop with error handling but no safeguards."""
try:
self._stop()
logger.info_once("Profiler stopped successfully.", scope="local")
logger.info_once("Profiler stopped successfully.")
except Exception as e:
logger.warning("Failed to stop profiler: %s", e)
self._running = False # Always mark as not running, assume stop worked
@@ -93,7 +93,7 @@ class WorkerProfiler(ABC):
and self._delay_iters > 0
and self._active_iteration_count == self._delay_iters
):
logger.info_once("Starting profiler after delay...", scope="local")
logger.info_once("Starting profiler after delay...")
self._call_start()
# Call profiler step for schedule-based profiling
@@ -109,9 +109,7 @@ class WorkerProfiler(ABC):
# Automatically stop the profiler after max iters
# will be marked as not running, but leave as active so that stop
# can clean up properly
logger.info_once(
"Max profiling iterations reached. Stopping profiler...", scope="local"
)
logger.info_once("Max profiling iterations reached. Stopping profiler...")
self._call_stop()
return
@@ -141,7 +139,7 @@ class WorkerProfiler(ABC):
def shutdown(self) -> None:
"""Ensure profiler is stopped when shutting down."""
logger.info_once("Shutting down profiler", scope="local")
logger.info_once("Shutting down profiler")
if self._running:
self.stop()
@@ -176,7 +174,6 @@ class TorchProfilerWrapper(WorkerProfiler):
logger.info_once(
"Torch profiling enabled. Traces will be saved to: %s",
torch_profiler_trace_dir,
scope="local",
)
logger.debug(
"Profiler config: record_shapes=%s,"
@@ -216,7 +213,6 @@ class TorchProfilerWrapper(WorkerProfiler):
profiler_config.wait_iterations,
profiler_config.warmup_iterations,
profiler_config.active_iterations,
scope="local",
)
self.profiler = torch.profiler.profile(
vllm/utils/deep_gemm.py
+3 -5
View File
@@ -106,16 +106,14 @@ def is_deep_gemm_e8m0_used() -> bool:
_lazy_init()
if _fp8_gemm_nt_impl is None:
logger.info_once(
"DeepGEMM E8M0 disabled: _fp8_gemm_nt_impl not found", scope="local"
)
logger.info_once("DeepGEMM E8M0 disabled: _fp8_gemm_nt_impl not found")
return False
if envs.VLLM_USE_DEEP_GEMM_E8M0:
logger.info_once("DeepGEMM E8M0 enabled on current platform.", scope="local")
logger.info_once("DeepGEMM E8M0 enabled on current platform.")
return True
logger.info_once("DeepGEMM E8M0 disabled on current configuration.", scope="local")
logger.info_once("DeepGEMM E8M0 disabled on current configuration.")
return False
vllm/utils/import_utils.py
-2
View File
@@ -66,14 +66,12 @@ def import_triton_kernels():
logger.debug_once(
f"Loading module triton_kernels from {triton_kernels.__file__}.",
scope="local",
)
elif _has_module("vllm.third_party.triton_kernels"):
import vllm.third_party.triton_kernels as triton_kernels
logger.debug_once(
f"Loading module triton_kernels from {triton_kernels.__file__}.",
scope="local",
)
sys.modules["triton_kernels"] = triton_kernels
else:
vllm/v1/attention/backends/fa_utils.py
-1
View File
@@ -118,7 +118,6 @@ def get_flash_attn_version(
logger.warning_once(
"Cannot use FA version 4 with batch invariance, "
"defaulting to FA version 2.",
scope="local",
)
fa_version = 2
vllm/v1/attention/backends/flash_attn.py
-1
View File
@@ -637,7 +637,6 @@ class FlashAttentionImpl(AttentionImpl):
logger.info_once(
"Using FlashAttention version %s",
self.vllm_flash_attn_version,
scope="local",
)
# Cache the batch invariant result for use in forward passes
self.batch_invariant_enabled = envs.VLLM_BATCH_INVARIANT
vllm/v1/core/kv_cache_utils.py
+1 -5
View File
@@ -1334,7 +1334,7 @@ def _report_kv_cache_config(
dcp_size,
)
num_tokens_str = f"{num_tokens:,}"
logger.info_once("GPU KV cache size: %s tokens", num_tokens_str, scope="local")
logger.info_once("GPU KV cache size: %s tokens", num_tokens_str)
max_model_len_str = f"{vllm_config.model_config.max_model_len:,}"
max_concurrency = get_max_concurrency_for_kv_cache_config(
vllm_config, kv_cache_config
@@ -1343,7 +1343,6 @@ def _report_kv_cache_config(
"Maximum concurrency for %s tokens per request: %.2fx",
max_model_len_str,
max_concurrency,
scope="local",
)
@@ -1445,7 +1444,6 @@ def _auto_fit_max_model_len(
"Auto-fit max_model_len: attention-free model, "
"using derived max_model_len=%d",
original_max,
scope="local",
)
return
@@ -1472,7 +1470,6 @@ def _auto_fit_max_model_len(
"Auto-fit max_model_len: full model context length %d fits in "
"available GPU memory",
original_max,
scope="local",
)
else:
# Need to reduce max_model_len to fit in memory
@@ -1483,7 +1480,6 @@ def _auto_fit_max_model_len(
original_max,
auto_fit_max,
format_gib(limiting_worker_mem),
scope="local",
)
vllm/v1/engine/core.py
-3
View File
@@ -293,7 +293,6 @@ class EngineCore:
compile_time + encoder_compile_time,
compile_time,
encoder_compile_time,
scope="local",
)
elif compile_time > 0:
logger.info_once(
@@ -301,13 +300,11 @@ class EngineCore:
"%.2f s (compilation: %.2f s)",
elapsed,
compile_time,
scope="local",
)
else:
logger.info_once(
"init engine (profile, create kv cache, warmup model) took %.2f s",
elapsed,
scope="local",
)
return scheduler_kv_cache_config
vllm/v1/executor/multiproc_executor.py
-1
View File
@@ -1032,7 +1032,6 @@ def set_multiprocessing_worker_envs():
"external environment to tune this value as needed.",
current_parallelism,
default_omp_num_threads,
scope="local",
)
os.environ["OMP_NUM_THREADS"] = str(default_omp_num_threads)
torch.set_num_threads(default_omp_num_threads)
vllm/v1/worker/dp_utils.py
-1
View File
@@ -29,7 +29,6 @@ def _get_device_and_group(parallel_config: ParallelConfig):
if parallel_config.disable_nccl_for_dp_synchronization:
logger.info_once(
"Using CPU all reduce to synchronize DP padding between ranks.",
scope="local",
)
device = "cpu"
group = get_dp_group().cpu_group
vllm/v1/worker/gpu/eplb_utils.py
+1 -3
View File
@@ -92,9 +92,7 @@ class EPLBController:
if not is_mixture_of_experts(model):
return False
logger.info_once(
"EPLB is enabled for model %s.", model_config.model, scope="local"
)
logger.info_once("EPLB is enabled for model %s.", model_config.model)
assert self.state is not None
self.state.add_model(model, model_config)
self._has_registered_models = True
vllm/v1/worker/gpu_model_runner.py
+1 -6
View File
@@ -4855,7 +4855,6 @@ class GPUModelRunner(
"Model loading took %s GiB memory and %.6f seconds",
format_gib(self.model_memory_usage),
time_after_load - time_before_load,
scope="local",
)
if not load_dummy_weights:
prepare_communication_buffer_for_model(self.model)
@@ -4989,7 +4988,7 @@ class GPUModelRunner(
)
# begin loading weights
logger.info_once("Reloading weights inplace...", scope="local")
logger.info_once("Reloading weights inplace...")
if is_checkpoint_format:
# load weights from checkpoint/ original model format
initialize_layerwise_reload(model)
@@ -5001,7 +5000,6 @@ class GPUModelRunner(
logger.warning_once(
"Reloading with `is_checkpoint_format=True` requires that "
"weights be in kernel format and already sharded",
scope="local",
)
loaded_weights = set()
for name, loaded_weight in weights_iterator:
@@ -5015,7 +5013,6 @@ class GPUModelRunner(
logger.info_once(
"Reloading and processing weights took %.2f seconds",
diff_seconds,
scope="local",
)
if self.model_config.quantization is None and loaded_weights is not None:
weights_not_loaded = weights_to_load - loaded_weights
@@ -5802,7 +5799,6 @@ class GPUModelRunner(
encoder_budget,
max_mm_items_per_batch,
dummy_modality,
scope="local",
)
# Create dummy batch of multimodal inputs.
@@ -6099,7 +6095,6 @@ class GPUModelRunner(
"Graph capturing finished in %.0f secs, took %.2f GiB",
elapsed_time,
cuda_graph_size / (1 << 30),
scope="local",
)
return cuda_graph_size
vllm/v1/worker/gpu_worker.py
+1 -2
View File
@@ -269,7 +269,7 @@ class Worker(WorkerBase):
)
if self.use_v2_model_runner:
logger.info_once("Using V2 Model Runner", scope="local")
logger.info_once("Using V2 Model Runner")
# Set random seed.
set_random_seed(self.model_config.seed)
@@ -440,7 +440,6 @@ class Worker(WorkerBase):
logger.info_once(
"Available KV cache memory: %s GiB",
format_gib(self.available_kv_cache_memory_bytes),
scope="local",
)
if cudagraph_memory_estimate > 0: