[Misc] Move print_*_once from utils to logger (#11298)

Signed-off-by: DarkLight1337 <[email protected]>
Signed-off-by: Maxime Fournioux <[email protected]>
Co-authored-by: Maxime Fournioux <[email protected]>

Author: DarkLight1337

.github/workflows/lint-and-deploy.yaml

@@ -64,6 +64,7 @@ jobs:
         run: |
           export AWS_ACCESS_KEY_ID=minioadmin
           export AWS_SECRET_ACCESS_KEY=minioadmin
+          sleep 30 && kubectl -n ns-vllm logs -f "$(kubectl -n ns-vllm get pods | awk '/deployment/ {print $1;exit}')" &
           helm install --wait --wait-for-jobs --timeout 5m0s --debug --create-namespace --namespace=ns-vllm test-vllm examples/online_serving/chart-helm -f examples/online_serving/chart-helm/values.yaml --set secrets.s3endpoint=http://minio:9000 --set secrets.s3bucketname=testbucket --set secrets.s3accesskeyid=$AWS_ACCESS_KEY_ID --set secrets.s3accesskey=$AWS_SECRET_ACCESS_KEY --set resources.requests.cpu=1 --set resources.requests.memory=4Gi --set resources.limits.cpu=2 --set resources.limits.memory=5Gi --set image.env[0].name=VLLM_CPU_KVCACHE_SPACE --set image.env[1].name=VLLM_LOGGING_LEVEL --set-string image.env[0].value="1" --set-string image.env[1].value="DEBUG" --set-string extraInit.s3modelpath="opt-125m/" --set-string 'resources.limits.nvidia\.com/gpu=0' --set-string 'resources.requests.nvidia\.com/gpu=0' --set-string image.repository="vllm-cpu-env"
     
       - name: curl test

vllm/attention/backends/torch_sdpa.py

@@ -13,9 +13,12 @@
 from vllm.attention.backends.utils import CommonAttentionState
 from vllm.attention.ops.ipex_attn import PagedAttention
 from vllm.attention.ops.paged_attn import PagedAttentionMetadata
-from vllm.utils import make_tensor_with_pad, print_warning_once
+from vllm.logger import init_logger
+from vllm.utils import make_tensor_with_pad
 from vllm.worker.cpu_model_runner import ModelInputForCPUBuilder
 
+logger = init_logger(__name__)
+
 
 class TorchSDPABackend(AttentionBackend):
 
@@ -396,8 +399,8 @@ def __init__(
             raise ValueError(
                 "Torch SPDA does not support block-sparse attention.")
         if logits_soft_cap is not None:
-            print_warning_once("Torch SPDA does not support logits soft cap. "
-                               "Outputs may be slightly off.")
+            logger.warning_once("Torch SPDA does not support logits soft cap. "
+                                "Outputs may be slightly off.")
         self.num_heads = num_heads
         self.head_size = head_size
         self.scale = float(scale)

vllm/attention/backends/xformers.py

@@ -17,7 +17,9 @@
     is_all_cross_attn_metadata_set, is_all_encoder_attn_metadata_set)
 from vllm.attention.ops.paged_attn import (PagedAttention,
                                            PagedAttentionMetadata)
-from vllm.utils import print_warning_once
+from vllm.logger import init_logger
+
+logger = init_logger(__name__)
 
 
 class XFormersBackend(AttentionBackend):
@@ -385,8 +387,8 @@ def __init__(
             raise ValueError(
                 "XFormers does not support block-sparse attention.")
         if logits_soft_cap is not None:
-            print_warning_once("XFormers does not support logits soft cap. "
-                               "Outputs may be slightly off.")
+            logger.warning_once("XFormers does not support logits soft cap. "
+                                "Outputs may be slightly off.")
         self.num_heads = num_heads
         self.head_size = head_size
         self.scale = float(scale)

vllm/config.py

@@ -32,8 +32,7 @@
 from vllm.transformers_utils.s3_utils import S3Model
 from vllm.transformers_utils.utils import is_s3
 from vllm.utils import (GiB_bytes, LayerBlockType, cuda_device_count_stateless,
-                        get_cpu_memory, print_warning_once, random_uuid,
-                        resolve_obj_by_qualname)
+                        get_cpu_memory, random_uuid, resolve_obj_by_qualname)
 
 if TYPE_CHECKING:
     from ray.util.placement_group import PlacementGroup
@@ -314,7 +313,7 @@ def __init__(self,
                 sliding_window_len_min = get_min_sliding_window(
                     self.hf_text_config.sliding_window)
 
-                print_warning_once(
+                logger.warning_once(
                     f"{self.hf_text_config.model_type} has interleaved "
                     "attention, which is currently not supported by the "
                     "XFORMERS backend. Disabling sliding window and capping "
@@ -2758,7 +2757,7 @@ def uuid(self):
 
         def model_post_init(self, __context: Any) -> None:
             if not self.enable_reshape and self.enable_fusion:
-                print_warning_once(
+                logger.warning_once(
                     "Fusion enabled but reshape elimination disabled."
                     "RMSNorm + quant (fp8) fusion might not work")
 
@@ -3151,7 +3150,7 @@ def __post_init__(self):
             self.scheduler_config.chunked_prefill_enabled and \
             self.model_config.dtype == torch.float32 and \
             current_platform.get_device_capability() == (7, 5):
-            print_warning_once(
+            logger.warning_once(
                 "Turing devices tensor cores do not support float32 matmul. "
                 "To workaround this limitation, vLLM will set 'ieee' input "
                 "precision for chunked prefill triton kernels.")

vllm/entrypoints/chat_utils.py

@@ -35,7 +35,6 @@
 from vllm.multimodal import MultiModalDataDict
 from vllm.multimodal.utils import MediaConnector
 from vllm.transformers_utils.tokenizer import AnyTokenizer, MistralTokenizer
-from vllm.utils import print_warning_once
 
 logger = init_logger(__name__)
 
@@ -985,14 +984,14 @@ def apply_mistral_chat_template(
     **kwargs: Any,
 ) -> List[int]:
     if chat_template is not None:
-        print_warning_once(
+        logger.warning_once(
             "'chat_template' cannot be overridden for mistral tokenizer.")
     if "add_generation_prompt" in kwargs:
-        print_warning_once(
+        logger.warning_once(
             "'add_generation_prompt' is not supported for mistral tokenizer, "
             "so it will be ignored.")
     if "continue_final_message" in kwargs:
-        print_warning_once(
+        logger.warning_once(
             "'continue_final_message' is not supported for mistral tokenizer, "
             "so it will be ignored.")
 

vllm/inputs/preprocess.py

@@ -10,7 +10,6 @@
 from vllm.multimodal.inputs import MultiModalDataDict, MultiModalInputsV2
 from vllm.prompt_adapter.request import PromptAdapterRequest
 from vllm.transformers_utils.tokenizer_group import BaseTokenizerGroup
-from vllm.utils import print_info_once, print_warning_once
 
 from .data import (DecoderOnlyInputs, EncoderDecoderInputs, ProcessorInputs,
                    PromptType, SingletonInputs, SingletonPrompt, token_inputs)
@@ -68,21 +67,24 @@ def get_decoder_start_token_id(self) -> Optional[int]:
         '''
 
         if not self.model_config.is_encoder_decoder:
-            print_warning_once("Using None for decoder start token id because "
-                               "this is not an encoder/decoder model.")
+            logger.warning_once(
+                "Using None for decoder start token id because "
+                "this is not an encoder/decoder model.")
             return None
 
         if (self.model_config is None or self.model_config.hf_config is None):
-            print_warning_once("Using None for decoder start token id because "
-                               "model config is not available.")
+            logger.warning_once(
+                "Using None for decoder start token id because "
+                "model config is not available.")
             return None
 
         dec_start_token_id = getattr(self.model_config.hf_config,
                                      'decoder_start_token_id', None)
         if dec_start_token_id is None:
-            print_warning_once("Falling back on <BOS> for decoder start token "
-                               "id because decoder start token id is not "
-                               "available.")
+            logger.warning_once(
+                "Falling back on <BOS> for decoder start token "
+                "id because decoder start token id is not "
+                "available.")
             dec_start_token_id = self.get_bos_token_id()
 
         return dec_start_token_id
@@ -231,7 +233,7 @@ def _can_process_multimodal(self) -> bool:
         # updated to use the new multi-modal processor
         can_process_multimodal = self.mm_registry.has_processor(model_config)
         if not can_process_multimodal:
-            print_info_once(
+            logger.info_once(
                 "Your model uses the legacy input pipeline instead of the new "
                 "multi-modal processor. Please note that the legacy pipeline "
                 "will be removed in a future release. For more details, see: "

vllm/inputs/registry.py

@@ -12,7 +12,7 @@
 from vllm.transformers_utils.processor import cached_get_processor
 from vllm.transformers_utils.tokenizer import AnyTokenizer
 from vllm.utils import (ClassRegistry, get_allowed_kwarg_only_overrides,
-                        print_warning_once, resolve_mm_processor_kwargs)
+                        resolve_mm_processor_kwargs)
 
 from .data import ProcessorInputs, SingletonInputs
 from .parse import is_encoder_decoder_inputs
@@ -352,7 +352,7 @@ def dummy_data_for_profiling(
         num_tokens = dummy_data.seq_data.prompt_token_ids
         if len(num_tokens) < seq_len:
             if is_encoder_data:
-                print_warning_once(
+                logger.warning_once(
                     f"Expected at least {seq_len} dummy encoder tokens for "
                     f"profiling, but found {len(num_tokens)} tokens instead.")
             else:

vllm/logger.py

@@ -4,11 +4,12 @@
 import logging
 import os
 import sys
-from functools import partial
+from functools import lru_cache, partial
 from logging import Logger
 from logging.config import dictConfig
 from os import path
-from typing import Dict, Optional
+from types import MethodType
+from typing import Any, Optional, cast
 
 import vllm.envs as envs
 
@@ -49,8 +50,44 @@
 }
 
 
+@lru_cache
+def _print_info_once(logger: Logger, msg: str) -> None:
+    # Set the stacklevel to 2 to print the original caller's line info
+    logger.info(msg, stacklevel=2)
+
+
+@lru_cache
+def _print_warning_once(logger: Logger, msg: str) -> None:
+    # Set the stacklevel to 2 to print the original caller's line info
+    logger.warning(msg, stacklevel=2)
+
+
+class _VllmLogger(Logger):
+    """
+    Note:
+        This class is just to provide type information.
+        We actually patch the methods directly on the :class:`logging.Logger`
+        instance to avoid conflicting with other libraries such as
+        `intel_extension_for_pytorch.utils._logger`.
+    """
+
+    def info_once(self, msg: str) -> None:
+        """
+        As :meth:`info`, but subsequent calls with the same message
+        are silently dropped.
+        """
+        _print_info_once(self, msg)
+
+    def warning_once(self, msg: str) -> None:
+        """
+        As :meth:`warning`, but subsequent calls with the same message
+        are silently dropped.
+        """
+        _print_warning_once(self, msg)
+
+
 def _configure_vllm_root_logger() -> None:
-    logging_config: Dict = {}
+    logging_config = dict[str, Any]()
 
     if not VLLM_CONFIGURE_LOGGING and VLLM_LOGGING_CONFIG_PATH:
         raise RuntimeError(
@@ -84,12 +121,22 @@ def _configure_vllm_root_logger() -> None:
         dictConfig(logging_config)
 
 
-def init_logger(name: str) -> Logger:
+def init_logger(name: str) -> _VllmLogger:
     """The main purpose of this function is to ensure that loggers are
     retrieved in such a way that we can be sure the root vllm logger has
     already been configured."""
 
-    return logging.getLogger(name)
+    logger = logging.getLogger(name)
+
+    methods_to_patch = {
+        "info_once": _print_info_once,
+        "warning_once": _print_warning_once,
+    }
+
+    for method_name, method in methods_to_patch.items():
+        setattr(logger, method_name, MethodType(method, logger))
+
+    return cast(_VllmLogger, logger)
 
 
 # The root logger is initialized when the module is imported.

vllm/lora/peft_helper.py

@@ -4,7 +4,9 @@
 from dataclasses import MISSING, dataclass, field, fields
 from typing import Literal, Optional, Union
 
-from vllm.utils import print_info_once
+from vllm.logger import init_logger
+
+logger = init_logger(__name__)
 
 
 @dataclass
@@ -42,7 +44,7 @@ def _validate_features(self):
     def __post_init__(self):
         self._validate_features()
         if self.use_rslora:
-            print_info_once("Loading LoRA weights trained with rsLoRA.")
+            logger.info_once("Loading LoRA weights trained with rsLoRA.")
             self.vllm_lora_scaling_factor = self.lora_alpha / math.sqrt(self.r)
         else:
             self.vllm_lora_scaling_factor = self.lora_alpha / self.r

vllm/lora/punica_wrapper/punica_selector.py

@@ -1,19 +1,21 @@
+from vllm.logger import init_logger
 from vllm.platforms import current_platform
-from vllm.utils import print_info_once
 
 from .punica_base import PunicaWrapperBase
 
+logger = init_logger(__name__)
+
 
 def get_punica_wrapper(*args, **kwargs) -> PunicaWrapperBase:
     if current_platform.is_cuda_alike():
         # Lazy import to avoid ImportError
         from vllm.lora.punica_wrapper.punica_gpu import PunicaWrapperGPU
-        print_info_once("Using PunicaWrapperGPU.")
+        logger.info_once("Using PunicaWrapperGPU.")
         return PunicaWrapperGPU(*args, **kwargs)
     elif current_platform.is_hpu():
         # Lazy import to avoid ImportError
         from vllm.lora.punica_wrapper.punica_hpu import PunicaWrapperHPU
-        print_info_once("Using PunicaWrapperHPU.")
+        logger.info_once("Using PunicaWrapperHPU.")
         return PunicaWrapperHPU(*args, **kwargs)
     else:
         raise NotImplementedError

vllm/model_executor/custom_op.py

@@ -5,7 +5,6 @@
 from vllm.config import get_current_vllm_config
 from vllm.logger import init_logger
 from vllm.platforms import current_platform
-from vllm.utils import print_warning_once
 
 logger = init_logger(__name__)
 
@@ -91,7 +90,7 @@ def enabled(cls) -> bool:
         compilation_config = get_current_vllm_config().compilation_config
         custom_ops = compilation_config.custom_ops
         if not hasattr(cls, "name"):
-            print_warning_once(
+            logger.warning_once(
                 f"Custom op {cls.__name__} was not registered, "
                 f"which means it won't appear in the op registry. "
                 f"It will be enabled/disabled based on the global settings.")

vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py

@@ -8,6 +8,7 @@
 
 import vllm.model_executor.layers.fused_moe  # noqa
 from vllm import _custom_ops as ops
+from vllm.logger import init_logger
 from vllm.model_executor.layers.fused_moe import (FusedMoE, FusedMoEMethodBase,
                                                   FusedMoeWeightScaleSupported)
 from vllm.model_executor.layers.quantization.compressed_tensors.schemes import (
@@ -16,7 +17,8 @@
     all_close_1d, normalize_e4m3fn_to_e4m3fnuz, per_tensor_dequantize)
 from vllm.model_executor.utils import set_weight_attrs
 from vllm.platforms import current_platform
-from vllm.utils import print_warning_once
+
+logger = init_logger(__name__)
 
 
 class GPTQMarlinState(Enum):
@@ -142,10 +144,10 @@ def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
                     "activation scales are None.")
             if (not all_close_1d(layer.w13_input_scale)
                     or not all_close_1d(layer.w2_input_scale)):
-                print_warning_once(
+                logger.warning_once(
                     "Found input_scales that are not equal for "
                     "fp8 MoE layer. Using the maximum across experts "
-                    "for each layer. ")
+                    "for each layer.")
             layer.w13_input_scale = torch.nn.Parameter(
                 layer.w13_input_scale.max(), requires_grad=False)
             layer.w2_input_scale = torch.nn.Parameter(

vllm/model_executor/layers/quantization/fp8.py

@@ -28,7 +28,6 @@
                                            PerTensorScaleParameter)
 from vllm.model_executor.utils import set_weight_attrs
 from vllm.platforms import current_platform
-from vllm.utils import print_warning_once
 
 ACTIVATION_SCHEMES = ["static", "dynamic"]
 
@@ -539,10 +538,10 @@ def process_weights_after_loading(self, layer: Module) -> None:
                         "activation scales are None.")
                 if (not all_close_1d(layer.w13_input_scale)
                         or not all_close_1d(layer.w2_input_scale)):
-                    print_warning_once(
+                    logger.warning_once(
                         "Found input_scales that are not equal for "
                         "fp8 MoE layer. Using the maximum across experts "
-                        "for each layer. ")
+                        "for each layer.")
                 layer.w13_input_scale = torch.nn.Parameter(
                     layer.w13_input_scale.max(), requires_grad=False)
                 layer.w2_input_scale = torch.nn.Parameter(