@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / b0c1f6202d

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 15, 13:30 UTC

[Misc] Remove OLMo2 config copy (#17066)

Signed-off-by: Isotr0py <2037008807@qq.com>

Isotr0py committed Apr 24, 2025 at 21:14 UTC b0c1f6202d2558c33bdcf252b0e7d47f9b5d19e3

4 files changed +4 -176

vllm/model_executor/models/olmo2.py

+1 -1

@@ -28,6 +28,7 @@ from typing import Iterable, Optional, Tuple, Union

29 import torch

30 from torch import nn

31 +from transformers import Olmo2Config

33 from vllm.attention import Attention

34 from vllm.config import VllmConfig

@@ -51,7 +52,6 @@ from vllm.model_executor.models.utils import (

52 make_layers, maybe_prefix)

53 from vllm.model_executor.sampling_metadata import SamplingMetadata

54 from vllm.sequence import IntermediateTensors

54 -from vllm.transformers_utils.configs.olmo2 import Olmo2Config

57 class Olmo2Attention(nn.Module):

vllm/transformers_utils/config.py

+3 -5

@@ -36,10 +36,9 @@ from vllm.transformers_utils.configs import (ChatGLMConfig, Cohere2Config,

36 KimiVLConfig, MedusaConfig,

37 MllamaConfig, MLPSpeculatorConfig,

38 MPTConfig, NemotronConfig,

39 - NVLM_D_Config, Olmo2Config,

40 - RWConfig, SkyworkR1VChatConfig,

41 - SolarConfig, Telechat2Config,

42 - UltravoxConfig)

39 + NVLM_D_Config, RWConfig,

40 + SkyworkR1VChatConfig, SolarConfig,

41 + Telechat2Config, UltravoxConfig)

42 # yapf: enable

43 from vllm.transformers_utils.utils import check_gguf_file

44 from vllm.utils import resolve_obj_by_qualname

@@ -76,7 +75,6 @@ _CONFIG_REGISTRY: Dict[str, Type[PretrainedConfig]] = {

75 "internvl_chat": InternVLChatConfig,

76 "nemotron": NemotronConfig,

77 "NVLM_D": NVLM_D_Config,

79 - "olmo2": Olmo2Config,

78 "solar": SolarConfig,

79 "skywork_chat": SkyworkR1VChatConfig,

80 "telechat": Telechat2Config,

vllm/transformers_utils/configs/__init__.py

-2

@@ -21,7 +21,6 @@ from vllm.transformers_utils.configs.moonvit import MoonViTConfig

21 from vllm.transformers_utils.configs.mpt import MPTConfig

22 from vllm.transformers_utils.configs.nemotron import NemotronConfig

23 from vllm.transformers_utils.configs.nvlm_d import NVLM_D_Config

24 -from vllm.transformers_utils.configs.olmo2 import Olmo2Config

24 from vllm.transformers_utils.configs.skyworkr1v import SkyworkR1VChatConfig

25 from vllm.transformers_utils.configs.solar import SolarConfig

26 from vllm.transformers_utils.configs.telechat2 import Telechat2Config

@@ -46,7 +45,6 @@ __all__ = [

45 "KimiVLConfig",

46 "NemotronConfig",

47 "NVLM_D_Config",

49 - "Olmo2Config",

48 "SkyworkR1VChatConfig",

49 "SolarConfig",

50 "Telechat2Config",

vllm/transformers_utils/configs/olmo2.py deleted

-168

@@ -1,168 +0,0 @@

1 -# SPDX-License-Identifier: Apache-2.0

2 -

3 -# yapf: disable

4 -# ruff: noqa: E501

5 -# coding=utf-8

6 -# Copied from

7 -# https://github.com/huggingface/transformers/blob/main/src/transformers/models/olmo2/configuration_olmo2.py

8 -"""OLMo 2 configuration."""

9 -

10 -from transformers.configuration_utils import PretrainedConfig

11 -from transformers.utils import logging

12 -

13 -logger = logging.get_logger(__name__)

14 -

15 -

16 -class Olmo2Config(PretrainedConfig):

17 - r"""

18 - This is the configuration class to store the configuration of a [`Olmo2Model`]. It is used to instantiate an OLMo2

19 - model according to the specified arguments, defining the model architecture. Instantiating a configuration with the

20 - defaults will yield a similar configuration to that of the [allenai/Olmo2-7B-1124-hf](https://huggingface.co/allenai/Olmo2-7B-1124-hf).

21 -

22 - Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the

23 - documentation from [`PretrainedConfig`] for more information.

24 -

25 -

26 - Args:

27 - vocab_size (`int`, *optional*, defaults to 50304):

28 - Vocabulary size of the Olmo2 model. Defines the number of different tokens that can be represented by the

29 - `inputs_ids` passed when calling [`Olmo2Model`]

30 - hidden_size (`int`, *optional*, defaults to 4096):

31 - Dimension of the hidden representations.

32 - intermediate_size (`int`, *optional*, defaults to 11008):

33 - Dimension of the MLP representations.

34 - num_hidden_layers (`int`, *optional*, defaults to 32):

35 - Number of hidden layers in the Transformer decoder.

36 - num_attention_heads (`int`, *optional*, defaults to 32):

37 - Number of attention heads for each attention layer in the Transformer decoder.

38 - num_key_value_heads (`int`, *optional*):

39 - This is the number of key_value heads that should be used to implement Grouped Query Attention. If

40 - `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if

41 - `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When

42 - converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed

43 - by meanpooling all the original heads within that group. For more details checkout [this

44 - paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to

45 - `num_attention_heads`.

46 - hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):

47 - The non-linear activation function (function or string) in the decoder.

48 - max_position_embeddings (`int`, *optional*, defaults to 2048):

49 - The maximum sequence length that this model might ever be used with.

50 - initializer_range (`float`, *optional*, defaults to 0.02):

51 - The standard deviation of the truncated_normal_initializer for initializing all weight matrices.

52 - use_cache (`bool`, *optional*, defaults to `True`):

53 - Whether or not the model should return the last key/values attentions (not used by all models). Only

54 - relevant if `config.is_decoder=True`.

55 - pad_token_id (`int`, *optional*, defaults to 1):

56 - Padding token id.

57 - bos_token_id (`int`, *optional*):

58 - Beginning of stream token id.

59 - eos_token_id (`int`, *optional*, defaults to 50279):

60 - End of stream token id.

61 - tie_word_embeddings (`bool`, *optional*, defaults to `False`):

62 - Whether to tie weight embeddings

63 - rope_theta (`float`, *optional*, defaults to 10000.0):

64 - The base period of the RoPE embeddings.

65 - rope_scaling (`Dict`, *optional*):

66 - Dictionary containing the scaling configuration for the RoPE embeddings. Currently supports two scaling

67 - strategies: linear and dynamic. Their scaling factor must be a float greater than 1. The expected format is

68 - `{"type": strategy name, "factor": scaling factor}`. When using this flag, don't update

69 - `max_position_embeddings` to the expected new maximum. See the following thread for more information on how

70 - these scaling strategies behave:

71 - https://www.reddit.com/r/LocalLLaMA/comments/14mrgpr/dynamically_scaled_rope_further_increases/. This is an

72 - experimental feature, subject to breaking API changes in future versions.

73 - attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`):

74 - Whether to use a bias in the query, key, value and output projection layers during self-attention.

75 - attention_dropout (`float`, *optional*, defaults to 0.0):

76 - The dropout ratio for the attention probabilities.

77 - rms_norm_eps (`float`, *optional*, defaults to 1e-05):

78 - The epsilon used by the rms normalization layers.

79 -

80 - ```python

81 - >>> from transformers import Olmo2Model, Olmo2Config

82 -

83 - >>> # Initializing a Olmo2 7B style configuration

84 - >>> configuration = Olmo2Config()

85 -

86 - >>> # Initializing a model from the Olmo2 7B style configuration

87 - >>> model = Olmo2Model(configuration)

88 -

89 - >>> # Accessing the model configuration

90 - >>> configuration = model.config

91 - ```

92 - """

93 -

94 - model_type = "olmo2"

95 - keys_to_ignore_at_inference = ["past_key_values"]

96 -

97 - def __init__(

98 - self,

99 - vocab_size=50304,

100 - hidden_size=4096,

101 - intermediate_size=11008,

102 - num_hidden_layers=32,

103 - num_attention_heads=32,

104 - num_key_value_heads=None,

105 - hidden_act="silu",

106 - max_position_embeddings=2048,

107 - initializer_range=0.02,

108 - use_cache=True,

109 - pad_token_id=1,

110 - bos_token_id=None,

111 - eos_token_id=50279,

112 - tie_word_embeddings=False,

113 - rope_theta=10000.0,

114 - rope_scaling=None,

115 - attention_bias=False,

116 - attention_dropout=0.0,

117 - rms_norm_eps=1e-5,

118 - **kwargs,

119 - ):

120 - super().__init__(

121 - pad_token_id=pad_token_id,

122 - bos_token_id=bos_token_id,

123 - eos_token_id=eos_token_id,

124 - tie_word_embeddings=tie_word_embeddings,

125 - **kwargs,

126 - )

127 - self.vocab_size = vocab_size

128 - self.max_position_embeddings = max_position_embeddings

129 - self.hidden_size = hidden_size

130 - self.intermediate_size = intermediate_size

131 - self.num_hidden_layers = num_hidden_layers

132 - self.num_attention_heads = num_attention_heads

133 -

134 - # for backward compatibility

135 - if num_key_value_heads is None:

136 - num_key_value_heads = num_attention_heads

137 -

138 - self.num_key_value_heads = num_key_value_heads

139 - self.hidden_act = hidden_act

140 - self.initializer_range = initializer_range

141 - self.use_cache = use_cache

142 - self.rope_theta = rope_theta

143 - self.rope_scaling = rope_scaling

144 - self._rope_scaling_validation()

145 - self.attention_bias = attention_bias

146 - self.attention_dropout = attention_dropout

147 -

148 - self.rms_norm_eps = rms_norm_eps

149 -

150 - def _rope_scaling_validation(self):

151 - """

152 - Validate the `rope_scaling` configuration.

153 - """

154 - if self.rope_scaling is None:

155 - return

156 -

157 - if not isinstance(self.rope_scaling, dict) or len(self.rope_scaling) != 2:

158 - raise ValueError(

159 - "`rope_scaling` must be a dictionary with two fields, `type` and `factor`, " f"got {self.rope_scaling}"

160 - )

161 - rope_scaling_type = self.rope_scaling.get("type", None)

162 - rope_scaling_factor = self.rope_scaling.get("factor", None)

163 - if rope_scaling_type is None or rope_scaling_type not in ["linear", "dynamic"]:

164 - raise ValueError(

165 - f"`rope_scaling`'s type field must be one of ['linear', 'dynamic'], got {rope_scaling_type}"

166 - )

167 - if rope_scaling_factor is None or not isinstance(rope_scaling_factor, float) or rope_scaling_factor <= 1.0:

168 - raise ValueError(f"`rope_scaling`'s factor field must be a float > 1, got {rope_scaling_factor}")