@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / b28d2104de
Fork 0

A high-throughput and memory-efficient inference and serving engine for LLMs

Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git

Synced Nov 16, 15:45 UTC

[Misc] Change dummy profiling and BOS fallback warns to log once (#8820)

Michael Goin committed Sep 26, 2024 at 19:18 UTC b28d2104dea6ba80c0f1f6c4596b5703d7ef923d
2 files changed +12 -10
vllm/inputs/preprocess.py
+8 -6
@@ -8,6 +8,7 @@ from vllm.logger import init_logger
8 from vllm.lora.request import LoRARequest
9 from vllm.prompt_adapter.request import PromptAdapterRequest
10 from vllm.transformers_utils.tokenizer_group import BaseTokenizerGroup
11 +from vllm.utils import print_warning_once
12
13 from .data import (EncoderDecoderLLMInputs, LLMInputs, PromptInputs,
14 SingletonPromptInputs)
@@ -71,20 +72,21 @@ class InputPreprocessor:
72 '''
73
74 if not self.is_encoder_decoder_model():
74 - logger.warning("Using None for decoder start token id because "
75 - "this is not an encoder/decoder model.")
75 + print_warning_once("Using None for decoder start token id because "
76 + "this is not an encoder/decoder model.")
77 return None
78
79 if (self.model_config is None or self.model_config.hf_config is None):
79 - logger.warning("Using None for decoder start token id because "
80 - "model config is not available.")
80 + print_warning_once("Using None for decoder start token id because "
81 + "model config is not available.")
82 return None
83
84 dec_start_token_id = getattr(self.model_config.hf_config,
85 'decoder_start_token_id', None)
86 if dec_start_token_id is None:
86 - logger.warning("Falling back on <BOS> for decoder start token id "
87 - "because decoder start token id is not available.")
87 + print_warning_once("Falling back on <BOS> for decoder start token "
88 + "id because decoder start token id is not "
89 + "available.")
90 dec_start_token_id = self.get_bos_token_id()
91
92 return dec_start_token_id
vllm/inputs/registry.py
+4 -4
@@ -9,7 +9,7 @@ from transformers import PretrainedConfig
9 from typing_extensions import TypeVar
10
11 from vllm.logger import init_logger
12 -from vllm.utils import get_allowed_kwarg_only_overrides
12 +from vllm.utils import get_allowed_kwarg_only_overrides, print_warning_once
13
14 from .data import LLMInputs
15
@@ -235,9 +235,9 @@ class InputRegistry:
235 num_tokens = seq_data.prompt_token_ids
236 if len(num_tokens) < seq_len:
237 if is_encoder_data:
238 - logger.warning(
239 - "Expected at least %d dummy encoder tokens for profiling, "
240 - "but found %d tokens instead.", seq_len, len(num_tokens))
238 + print_warning_once(
239 + f"Expected at least {seq_len} dummy encoder tokens for "
240 + f"profiling, but found {len(num_tokens)} tokens instead.")
241 else:
242 raise AssertionError(
243 f"Expected at least {seq_len} dummy tokens for profiling, "