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

Signed-off-by: yewentao256 <zhyanwentao@126.com>
Signed-off-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>
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
@@ -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
@@ -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
@@ -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:
@@ -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"
@@ -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
@@ -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
@@ -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:
@@ -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:
@@ -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:
@@ -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:
@@ -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."
@@ -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:
@@ -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:
@@ -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:
@@ -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
@@ -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(
@@ -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(
@@ -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
@@ -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
@@ -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
@@ -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.
@@ -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!")
+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: