[Fix] Support passing args to logger (#17425)

Signed-off-by: Aaron Pham <contact@aarnphm.xyz>

vllm/config.py

@@ -518,11 +518,11 @@ class ModelConfig:
                     self.hf_text_config.sliding_window)
 
                 logger.warning_once(
-                    f"{self.hf_text_config.model_type} has interleaved "
+                    "%s has interleaved attention, which is currently not supported by the %s backend. Disabling sliding window and capping the max length to the sliding window size (%d).",  # noqa: E501
-                    "attention, which is currently not supported by the "
+                    self.hf_text_config.model_type,
-                    f"{backend} backend. Disabling sliding window and capping "
+                    backend,
-                    "the max length to the sliding window size "
+                    sliding_window_len_min,
-                    f"({sliding_window_len_min}).")
+                )
                 self.disable_sliding_window = True
             else:
                 # for a model with interleaved attention,

vllm/logger.py

@@ -5,6 +5,7 @@ import json
 import logging
 import os
 import sys
+from collections.abc import Hashable
 from functools import lru_cache, partial
 from logging import Logger
 from logging.config import dictConfig
@@ -52,15 +53,15 @@ DEFAULT_LOGGING_CONFIG = {
 
 
 @lru_cache
-def _print_info_once(logger: Logger, msg: str) -> None:
+def _print_info_once(logger: Logger, msg: str, *args: Hashable) -> None:
     # Set the stacklevel to 2 to print the original caller's line info
-    logger.info(msg, stacklevel=2)
+    logger.info(msg, *args, stacklevel=2)
 
 
 @lru_cache
-def _print_warning_once(logger: Logger, msg: str) -> None:
+def _print_warning_once(logger: Logger, msg: str, *args: Hashable) -> None:
     # Set the stacklevel to 2 to print the original caller's line info
-    logger.warning(msg, stacklevel=2)
+    logger.warning(msg, *args, stacklevel=2)
 
 
 class _VllmLogger(Logger):
@@ -72,19 +73,19 @@ class _VllmLogger(Logger):
         `intel_extension_for_pytorch.utils._logger`.
     """
 
-    def info_once(self, msg: str) -> None:
+    def info_once(self, msg: str, *args: Hashable) -> None:
         """
         As :meth:`info`, but subsequent calls with the same message
         are silently dropped.
         """
-        _print_info_once(self, msg)
+        _print_info_once(self, msg, *args)
 
-    def warning_once(self, msg: str) -> None:
+    def warning_once(self, msg: str, *args: Hashable) -> None:
         """
         As :meth:`warning`, but subsequent calls with the same message
         are silently dropped.
         """
-        _print_warning_once(self, msg)
+        _print_warning_once(self, msg, *args)
 
 
 def _configure_vllm_root_logger() -> None:

vllm/lora/punica_wrapper/punica_selector.py

@@ -15,6 +15,5 @@ def get_punica_wrapper(*args, **kwargs) -> PunicaWrapperBase:
     punica_wrapper = punica_wrapper_cls(*args, **kwargs)
     assert punica_wrapper is not None, \
         "the punica_wrapper_qualname(" + punica_wrapper_qualname + ") is wrong."
-    logger.info_once("Using " + punica_wrapper_qualname.rsplit(".", 1)[1] +
+    logger.info_once("Using %s.", punica_wrapper_qualname.rsplit(".", 1)[1])
-                     ".")
     return punica_wrapper

vllm/model_executor/custom_op.py

@@ -107,9 +107,9 @@ class CustomOp(nn.Module):
         custom_ops = compilation_config.custom_ops
         if not hasattr(cls, "name"):
             logger.warning_once(
-                f"Custom op {cls.__name__} was not registered, "
+                "Custom op %s was not registered, which means it won't appear in the op registry. It will be enabled/disabled based on the global settings.",  # noqa: E501
-                f"which means it won't appear in the op registry. "
+                cls.__name__,
-                f"It will be enabled/disabled based on the global settings.")
+            )
             return CustomOp.default_on()
 
         enabled = f"+{cls.name}" in custom_ops

vllm/model_executor/guided_decoding/xgrammar_decoding.py

@@ -191,9 +191,9 @@ class GrammarConfig:
 
             if model_with_warn is not None and any_whitespace:
                 logger.info_once(
-                    f"{model_with_warn} model detected, consider setting "
+                    "%s model detected, consider setting `disable_any_whitespace` to prevent runaway generation of whitespaces.",  # noqa: E501
-                    "`disable_any_whitespace` to prevent runaway generation "
+                    model_with_warn,
-                    "of whitespaces.")
+                )
             # Validate the schema and raise ValueError here if it is invalid.
             # This is to avoid exceptions in model execution, which will crash
             # the engine worker process.

vllm/model_executor/layers/quantization/awq_marlin.py

@@ -130,8 +130,9 @@ class AWQMarlinConfig(QuantizationConfig):
             # Check if the layer is supported by AWQMarlin.
             if not check_marlin_supports_layer(layer, self.group_size):
                 logger.warning_once(
-                    f"Layer '{prefix}' is not supported by AWQMarlin. "
+                    "Layer '%s' is not supported by AWQMarlin. Falling back to unoptimized AWQ kernels.",  # noqa: E501
-                    "Falling back to unoptimized AWQ kernels.")
+                    prefix,
+                )
                 return AWQConfig.from_config(
                     self.full_config).get_quant_method(layer, prefix)
             return AWQMarlinLinearMethod(self)

vllm/model_executor/model_loader/weight_utils.py

@@ -716,10 +716,10 @@ def maybe_remap_kv_scale_name(name: str, params_dict: dict) -> Optional[str]:
         remapped_name = name.replace(".kv_scale", ".attn.k_scale")
         if remapped_name not in params_dict:
             logger.warning_once(
-                f"Found kv_scale in the checkpoint (e.g. {name}), "
+                "Found kv_scale in the checkpoint (e.g. %s), but not found the expected name in the model (e.g. %s). kv_scale is not loaded.",  #  noqa: E501
-                "but not found the expected name in the model "
+                name,
-                f"(e.g. {remapped_name}). kv_scale is "
+                remapped_name,
-                "not loaded.")
+            )
             return None
         return remapped_name
 
@@ -738,10 +738,12 @@ def maybe_remap_kv_scale_name(name: str, params_dict: dict) -> Optional[str]:
                 remapped_name = name.replace(scale_name, f".attn{scale_name}")
             if remapped_name not in params_dict:
                 logger.warning_once(
-                    f"Found {scale_name} in the checkpoint (e.g. {name}), "
+                    "Found %s in the checkpoint (e.g. %s), but not found the expected name in the model (e.g. %s). %s is not loaded.",  # noqa: E501
-                    "but not found the expected name in the model "
+                    scale_name,
-                    f"(e.g. {remapped_name}). {scale_name} is "
+                    name,
-                    "not loaded.")
+                    remapped_name,
+                    scale_name,
+                )
                 return None
             return remapped_name
 

vllm/model_executor/models/chameleon.py

@@ -1111,10 +1111,10 @@ class ChameleonForConditionalGeneration(nn.Module, SupportsMultiModal,
                             ".kv_scale", ".attn.kv_scale")
                         if remapped_kv_scale_name not in params_dict:
                             logger.warning_once(
-                                "Found kv scale in the checkpoint (e.g. "
+                                "Found kv scale in the checkpoint (e.g. %s), but not found the expected name in the model (e.g. %s). kv-scale is not loaded.",  # noqa: E501
-                                f"{name}), but not found the expected name in "
+                                name,
-                                f"the model (e.g. {remapped_kv_scale_name}). "
+                                remapped_kv_scale_name,
-                                "kv-scale is not loaded.")
+                            )
                             continue
                         else:
                             name = remapped_kv_scale_name

vllm/model_executor/models/olmoe.py

@@ -385,11 +385,10 @@ class OlmoeModel(nn.Module):
                             ".kv_scale", ".attn.kv_scale")
                         if remapped_kv_scale_name not in params_dict:
                             logger.warning_once(
-                                "Found kv scale in the checkpoint "
+                                "Found kv scale in the checkpoint (e.g. %s), but not found the expected name in the model (e.g. %s). kv-scale is not loaded.",  # noqa: E501
-                                f"(e.g. {name}), but not found the expected "
+                                name,
-                                f"name in the model "
+                                remapped_kv_scale_name,
-                                f"(e.g. {remapped_kv_scale_name}). "
+                            )
-                                "kv-scale is not loaded.")
                             continue
                         else:
                             name = remapped_kv_scale_name

vllm/model_executor/models/qwen2_moe.py

@@ -462,11 +462,10 @@ class Qwen2MoeModel(nn.Module):
                             ".kv_scale", ".attn.kv_scale")
                         if remapped_kv_scale_name not in params_dict:
                             logger.warning_once(
-                                "Found kv scale in the checkpoint "
+                                "Found kv_scale in the checkpoint (e.g. %s), but not found the expected name in the model (e.g. %s). kv_scale is not loaded.",  #  noqa: E501
-                                f"(e.g. {name}), but not found the expected "
+                                name,
-                                f"name in the model "
+                                remapped_kv_scale_name,
-                                f"(e.g. {remapped_kv_scale_name}). "
+                            )
-                                "kv-scale is not loaded.")
                             continue
                         else:
                             name = remapped_kv_scale_name

vllm/model_executor/models/qwen3_moe.py

@@ -459,11 +459,10 @@ class Qwen3MoeModel(nn.Module):
                             ".kv_scale", ".attn.kv_scale")
                         if remapped_kv_scale_name not in params_dict:
                             logger.warning_once(
-                                "Found kv scale in the checkpoint "
+                                "Found kv scale in the checkpoint (e.g. %s), but not found the expected name in the model (e.g. %s). kv-scale is not loaded.",  # noqa: E501
-                                f"(e.g. {name}), but not found the expected "
+                                name,
-                                f"name in the model "
+                                remapped_kv_scale_name,
-                                f"(e.g. {remapped_kv_scale_name}). "
+                            )
-                                "kv-scale is not loaded.")
                             continue
                         else:
                             name = remapped_kv_scale_name

vllm/multimodal/profiling.py

@@ -215,17 +215,14 @@ class MultiModalProfiler(Generic[_I]):
         elif total_len > seq_len and not envs.VLLM_USE_V1:
             # `max_num_batched_tokens` is defined by `SchedulerConfig`
             logger.warning_once(
-                "The encoder sequence length used for profiling ("
+                "The encoder sequence length used for profiling (max_num_batched_tokens / max_num_seqs = %d) "  # noqa: E501
-                f"max_num_batched_tokens / max_num_seqs = {seq_len}) "
+                "is too short to hold the multi-modal embeddings in the worst case (%d tokens in total, out of which %s are reserved for multi-modal embeddings). "  # noqa: E501
-                " is too short "
+                "This may cause certain multi-modal inputs to fail during inference, even when the input text is short. "  # noqa: E501
-                "to hold the multi-modal embeddings in the worst case "
+                "To avoid this, you should increase `max_model_len`, reduce `max_num_seqs`, and/or reduce `mm_counts`.",  # noqa: E501
-                f"({total_len} tokens in total, out of which "
+                seq_len,
-                f"{self._get_mm_num_tokens(mm_inputs)} are reserved for "
+                total_len,
-                "multi-modal embeddings). This may cause certain "
+                str(self._get_mm_num_tokens(mm_inputs)),
-                "multi-modal inputs to fail during inference, even when "
+            )
-                "the input text is short. To avoid this, you should "
-                "increase `max_model_len`, reduce `max_num_seqs`, "
-                "and/or reduce `mm_counts`.")
 
         return DummyEncoderData(encoder_prompt_token_ids)
 
@@ -243,17 +240,14 @@ class MultiModalProfiler(Generic[_I]):
         if total_len > seq_len and not envs.VLLM_USE_V1:
             # `max_num_batched_tokens` is defined by `SchedulerConfig`
             logger.warning_once(
-                "The sequence length used for profiling ("
+                "The sequence length used for profiling (max_num_batched_tokens / max_num_seqs = %d) "  # noqa: E501
-                f"max_num_batched_tokens / max_num_seqs = {seq_len}) "
+                "is too short to hold the multi-modal embeddings in the worst case (%d tokens in total, out of which %s are reserved for multi-modal embeddings). "  # noqa: E501
-                "is too short "
+                "This may cause certain multi-modal inputs to fail during inference, even when the input text is short. "  # noqa: E501
-                "to hold the multi-modal embeddings in the worst case "
+                "To avoid this, you should increase `max_model_len`, reduce `max_num_seqs`, and/or reduce `mm_counts`.",  # noqa: E501
-                f"({total_len} tokens in total, out of which "
+                seq_len,
-                f"{self._get_mm_num_tokens(mm_inputs)} are reserved for "
+                total_len,
-                "multi-modal embeddings). This may cause certain "
+                str(self._get_mm_num_tokens(mm_inputs)),
-                "multi-modal inputs to fail during inference, even when "
+            )
-                "the input text is short. To avoid this, you should "
-                "increase `max_model_len`, reduce `max_num_seqs`, "
-                "and/or reduce `mm_counts`.")
 
         if total_len < seq_len:
             prompt_token_ids.extend([0] * (seq_len - total_len))

vllm/multimodal/registry.py

@@ -316,7 +316,9 @@ class MultiModalRegistry:
         token_ids = dummy_data.prompt_token_ids
         if len(token_ids) < seq_len:
             logger.warning_once(
-                f"Expected at least {seq_len} dummy encoder tokens for "
+                "Expected at least %d dummy encoder tokens for profiling, but found %d tokens instead.",  # noqa: E501
-                f"profiling, but found {len(token_ids)} tokens instead.")
+                seq_len,
+                len(token_ids),
+            )
 
         return dummy_data