da4e7687b5 · tduncan/Nvidia-Optimization-Stack-VLLM

vllm/config.py

+6 -6

@@ -278,7 +278,7 @@ class ModelConfig:

278 max_model_len: int = None # type: ignore

279 """Model context length (prompt and output). If unspecified, will be

280 automatically derived from the model config.

281 -

281 +

282 When passing via `--max-model-len`, supports k/m/g/K/M/G in human-readable

283 format. Examples:\n

284 - 1k -> 1000\n

@@ -518,11 +518,11 @@ class ModelConfig:

518 self.hf_text_config.sliding_window)

519

520 logger.warning_once(

521 - f"{self.hf_text_config.model_type} has interleaved "

522 - "attention, which is currently not supported by the "

523 - f"{backend} backend. Disabling sliding window and capping "

524 - "the max length to the sliding window size "

525 - f"({sliding_window_len_min}).")

521 + "%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

522 + self.hf_text_config.model_type,

523 + backend,

524 + sliding_window_len_min,

525 + )

526 self.disable_sliding_window = True

527 else:

528 # for a model with interleaved attention,

vllm/logger.py

+9 -8

@@ -5,6 +5,7 @@ import json

5 import logging

6 import os

7 import sys

8 +from collections.abc import Hashable

9 from functools import lru_cache, partial

10 from logging import Logger

11 from logging.config import dictConfig

@@ -52,15 +53,15 @@ DEFAULT_LOGGING_CONFIG = {

53

54

55 @lru_cache

55 -def _print_info_once(logger: Logger, msg: str) -> None:

56 +def _print_info_once(logger: Logger, msg: str, *args: Hashable) -> None:

57 # Set the stacklevel to 2 to print the original caller's line info

57 - logger.info(msg, stacklevel=2)

58 + logger.info(msg, *args, stacklevel=2)

59

60

61 @lru_cache

61 -def _print_warning_once(logger: Logger, msg: str) -> None:

62 +def _print_warning_once(logger: Logger, msg: str, *args: Hashable) -> None:

63 # Set the stacklevel to 2 to print the original caller's line info

63 - logger.warning(msg, stacklevel=2)

64 + logger.warning(msg, *args, stacklevel=2)

65

66

67 class _VllmLogger(Logger):

@@ -72,19 +73,19 @@ class _VllmLogger(Logger):

73 `intel_extension_for_pytorch.utils._logger`.

74 """

75

75 - def info_once(self, msg: str) -> None:

76 + def info_once(self, msg: str, *args: Hashable) -> None:

77 """

78 As :meth:`info`, but subsequent calls with the same message

79 are silently dropped.

80 """

80 - _print_info_once(self, msg)

81 + _print_info_once(self, msg, *args)

82

82 - def warning_once(self, msg: str) -> None:

83 + def warning_once(self, msg: str, *args: Hashable) -> None:

84 """

85 As :meth:`warning`, but subsequent calls with the same message

86 are silently dropped.

87 """

87 - _print_warning_once(self, msg)

88 + _print_warning_once(self, msg, *args)

89

90

91 def _configure_vllm_root_logger() -> None:

vllm/lora/punica_wrapper/punica_selector.py

+1 -2

@@ -15,6 +15,5 @@ def get_punica_wrapper(*args, **kwargs) -> PunicaWrapperBase:

15 punica_wrapper = punica_wrapper_cls(*args, **kwargs)

16 assert punica_wrapper is not None, \

17 "the punica_wrapper_qualname(" + punica_wrapper_qualname + ") is wrong."

18 - logger.info_once("Using " + punica_wrapper_qualname.rsplit(".", 1)[1] +

19 - ".")

18 + logger.info_once("Using %s.", punica_wrapper_qualname.rsplit(".", 1)[1])

19 return punica_wrapper

vllm/model_executor/custom_op.py

+3 -3

@@ -107,9 +107,9 @@ class CustomOp(nn.Module):

107 custom_ops = compilation_config.custom_ops

108 if not hasattr(cls, "name"):

109 logger.warning_once(

110 - f"Custom op {cls.__name__} was not registered, "

111 - f"which means it won't appear in the op registry. "

112 - f"It will be enabled/disabled based on the global settings.")

110 + "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

111 + cls.__name__,

112 + )

113 return CustomOp.default_on()

114

115 enabled = f"+{cls.name}" in custom_ops

vllm/model_executor/guided_decoding/xgrammar_decoding.py

+3 -3

@@ -191,9 +191,9 @@ class GrammarConfig:

191

192 if model_with_warn is not None and any_whitespace:

193 logger.info_once(

194 - f"{model_with_warn} model detected, consider setting "

195 - "`disable_any_whitespace` to prevent runaway generation "

196 - "of whitespaces.")

194 + "%s model detected, consider setting `disable_any_whitespace` to prevent runaway generation of whitespaces.", # noqa: E501

195 + model_with_warn,

196 + )

197 # Validate the schema and raise ValueError here if it is invalid.

198 # This is to avoid exceptions in model execution, which will crash

199 # the engine worker process.

vllm/model_executor/layers/quantization/awq_marlin.py

+3 -2

@@ -130,8 +130,9 @@ class AWQMarlinConfig(QuantizationConfig):

130 # Check if the layer is supported by AWQMarlin.

131 if not check_marlin_supports_layer(layer, self.group_size):

132 logger.warning_once(

133 - f"Layer '{prefix}' is not supported by AWQMarlin. "

134 - "Falling back to unoptimized AWQ kernels.")

133 + "Layer '%s' is not supported by AWQMarlin. Falling back to unoptimized AWQ kernels.", # noqa: E501

134 + prefix,

135 + )

136 return AWQConfig.from_config(

137 self.full_config).get_quant_method(layer, prefix)

138 return AWQMarlinLinearMethod(self)

vllm/model_executor/model_loader/weight_utils.py

+11 -9

@@ -464,7 +464,7 @@ def fastsafetensors_weights_iterator(

464 hf_weights_files: List[str],

465 use_tqdm_on_load: bool,

466 ) -> Generator[Tuple[str, torch.Tensor], None, None]:

467 - """Iterate over the weights in the model safetensor files

467 + """Iterate over the weights in the model safetensor files

468 using fastsafetensor library."""

469 if torch.distributed.is_initialized():

470 pg = torch.distributed.group.WORLD

@@ -716,10 +716,10 @@ def maybe_remap_kv_scale_name(name: str, params_dict: dict) -> Optional[str]:

716 remapped_name = name.replace(".kv_scale", ".attn.k_scale")

717 if remapped_name not in params_dict:

718 logger.warning_once(

719 - f"Found kv_scale in the checkpoint (e.g. {name}), "

720 - "but not found the expected name in the model "

721 - f"(e.g. {remapped_name}). kv_scale is "

722 - "not loaded.")

719 + "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

720 + name,

721 + remapped_name,

722 + )

723 return None

724 return remapped_name

725

@@ -738,10 +738,12 @@ def maybe_remap_kv_scale_name(name: str, params_dict: dict) -> Optional[str]:

738 remapped_name = name.replace(scale_name, f".attn{scale_name}")

739 if remapped_name not in params_dict:

740 logger.warning_once(

741 - f"Found {scale_name} in the checkpoint (e.g. {name}), "

742 - "but not found the expected name in the model "

743 - f"(e.g. {remapped_name}). {scale_name} is "

744 - "not loaded.")

741 + "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

742 + scale_name,

743 + name,

744 + remapped_name,

745 + scale_name,

746 + )

747 return None

748 return remapped_name

749

vllm/model_executor/models/chameleon.py

+4 -4

@@ -1111,10 +1111,10 @@ class ChameleonForConditionalGeneration(nn.Module, SupportsMultiModal,

1111 ".kv_scale", ".attn.kv_scale")

1112 if remapped_kv_scale_name not in params_dict:

1113 logger.warning_once(

1114 - "Found kv scale in the checkpoint (e.g. "

1115 - f"{name}), but not found the expected name in "

1116 - f"the model (e.g. {remapped_kv_scale_name}). "

1117 - "kv-scale is not loaded.")

1114 + "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

1115 + name,

1116 + remapped_kv_scale_name,

1117 + )

1118 continue

1119 else:

1120 name = remapped_kv_scale_name

vllm/model_executor/models/olmoe.py

+4 -5

@@ -385,11 +385,10 @@ class OlmoeModel(nn.Module):

385 ".kv_scale", ".attn.kv_scale")

386 if remapped_kv_scale_name not in params_dict:

387 logger.warning_once(

388 - "Found kv scale in the checkpoint "

389 - f"(e.g. {name}), but not found the expected "

390 - f"name in the model "

391 - f"(e.g. {remapped_kv_scale_name}). "

392 - "kv-scale is not loaded.")

388 + "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

389 + name,

390 + remapped_kv_scale_name,

391 + )

392 continue

393 else:

394 name = remapped_kv_scale_name

vllm/model_executor/models/qwen2_moe.py

+4 -5

@@ -462,11 +462,10 @@ class Qwen2MoeModel(nn.Module):

462 ".kv_scale", ".attn.kv_scale")

463 if remapped_kv_scale_name not in params_dict:

464 logger.warning_once(

465 - "Found kv scale in the checkpoint "

466 - f"(e.g. {name}), but not found the expected "

467 - f"name in the model "

468 - f"(e.g. {remapped_kv_scale_name}). "

469 - "kv-scale is not loaded.")

465 + "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

466 + name,

467 + remapped_kv_scale_name,

468 + )

469 continue

470 else:

471 name = remapped_kv_scale_name

vllm/model_executor/models/qwen3_moe.py

+4 -5

@@ -459,11 +459,10 @@ class Qwen3MoeModel(nn.Module):

459 ".kv_scale", ".attn.kv_scale")

460 if remapped_kv_scale_name not in params_dict:

461 logger.warning_once(

462 - "Found kv scale in the checkpoint "

463 - f"(e.g. {name}), but not found the expected "

464 - f"name in the model "

465 - f"(e.g. {remapped_kv_scale_name}). "

466 - "kv-scale is not loaded.")

462 + "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

463 + name,

464 + remapped_kv_scale_name,

465 + )

466 continue

467 else:

468 name = remapped_kv_scale_name

vllm/multimodal/profiling.py

+16 -22

@@ -215,17 +215,14 @@ class MultiModalProfiler(Generic[_I]):

215 elif total_len > seq_len and not envs.VLLM_USE_V1:

216 # `max_num_batched_tokens` is defined by `SchedulerConfig`

217 logger.warning_once(

218 - "The encoder sequence length used for profiling ("

219 - f"max_num_batched_tokens / max_num_seqs = {seq_len}) "

220 - " is too short "

221 - "to hold the multi-modal embeddings in the worst case "

222 - f"({total_len} tokens in total, out of which "

223 - f"{self._get_mm_num_tokens(mm_inputs)} are reserved for "

224 - "multi-modal embeddings). This may cause certain "

225 - "multi-modal inputs to fail during inference, even when "

226 - "the input text is short. To avoid this, you should "

227 - "increase `max_model_len`, reduce `max_num_seqs`, "

228 - "and/or reduce `mm_counts`.")

218 + "The encoder sequence length used for profiling (max_num_batched_tokens / max_num_seqs = %d) " # noqa: E501

219 + "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

220 + "This may cause certain multi-modal inputs to fail during inference, even when the input text is short. " # noqa: E501

221 + "To avoid this, you should increase `max_model_len`, reduce `max_num_seqs`, and/or reduce `mm_counts`.", # noqa: E501

222 + seq_len,

223 + total_len,

224 + str(self._get_mm_num_tokens(mm_inputs)),

225 + )

226

227 return DummyEncoderData(encoder_prompt_token_ids)

228

@@ -243,17 +240,14 @@ class MultiModalProfiler(Generic[_I]):

240 if total_len > seq_len and not envs.VLLM_USE_V1:

241 # `max_num_batched_tokens` is defined by `SchedulerConfig`

242 logger.warning_once(

246 - "The sequence length used for profiling ("

247 - f"max_num_batched_tokens / max_num_seqs = {seq_len}) "

248 - "is too short "

249 - "to hold the multi-modal embeddings in the worst case "

250 - f"({total_len} tokens in total, out of which "

251 - f"{self._get_mm_num_tokens(mm_inputs)} are reserved for "

252 - "multi-modal embeddings). This may cause certain "

253 - "multi-modal inputs to fail during inference, even when "

254 - "the input text is short. To avoid this, you should "

255 - "increase `max_model_len`, reduce `max_num_seqs`, "

256 - "and/or reduce `mm_counts`.")

243 + "The sequence length used for profiling (max_num_batched_tokens / max_num_seqs = %d) " # noqa: E501

244 + "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

245 + "This may cause certain multi-modal inputs to fail during inference, even when the input text is short. " # noqa: E501

246 + "To avoid this, you should increase `max_model_len`, reduce `max_num_seqs`, and/or reduce `mm_counts`.", # noqa: E501

247 + seq_len,

248 + total_len,

249 + str(self._get_mm_num_tokens(mm_inputs)),

250 + )

251

252 if total_len < seq_len:

253 prompt_token_ids.extend([0] * (seq_len - total_len))

vllm/multimodal/registry.py

+7 -5

@@ -100,7 +100,7 @@ class MultiModalRegistry:

100 model_config: "ModelConfig",

101 ) -> Mapping[str, int]:

102 """

103 - Get the maximum number of tokens per data item from each modality based

103 + Get the maximum number of tokens per data item from each modality based

104 on underlying model configuration.

105 """

106 if not model_config.is_multimodal_model:

@@ -126,11 +126,11 @@ class MultiModalRegistry:

126 ) -> Mapping[str, int]:

127 """

128 Get the maximum number of tokens per data item from each modality based

129 - on underlying model configuration, excluding modalities that user

129 + on underlying model configuration, excluding modalities that user

130 explicitly disabled via `limit_mm_per_prompt`.

131

132 Note:

133 - This is currently directly used only in V1 for profiling the memory

133 + This is currently directly used only in V1 for profiling the memory

134 usage of a model.

135 """

136 mm_limits = self.get_mm_limits_per_prompt(model_config)

@@ -316,7 +316,9 @@ class MultiModalRegistry:

316 token_ids = dummy_data.prompt_token_ids

317 if len(token_ids) < seq_len:

318 logger.warning_once(

319 - f"Expected at least {seq_len} dummy encoder tokens for "

320 - f"profiling, but found {len(token_ids)} tokens instead.")

319 + "Expected at least %d dummy encoder tokens for profiling, but found %d tokens instead.", # noqa: E501

320 + seq_len,

321 + len(token_ids),

322 + )

323

324 return dummy_data

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