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 13, 18:15 UTC
[Chore] Remove Nemotron-Nano-VL config copy (#28126)
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Isotr0py committed
Nov 6, 2025 at 04:06 UTC
ffb08379d8870a1a81ba82b72797f196838d0c86
3 files changed
+1
-63
vllm/transformers_utils/config.py
+1
-1
@@ -81,7 +81,6 @@ _CONFIG_REGISTRY: dict[str, type[PretrainedConfig]] = LazyConfigDict(
81
flex_olmo="FlexOlmoConfig",
82
kimi_linear="KimiLinearConfig",
83
kimi_vl="KimiVLConfig",
84
- Llama_Nemotron_Nano_VL="Nemotron_Nano_VL_Config",
84
RefinedWeb="RWConfig", # For tiiuae/falcon-40b(-instruct)
85
RefinedWebModel="RWConfig", # For tiiuae/falcon-7b(-instruct)
86
jais="JAISConfig",
@@ -106,6 +105,7 @@ _CONFIG_ATTRS_MAPPING: dict[str, str] = {
105
106
_AUTO_CONFIG_KWARGS_OVERRIDES: dict[str, dict[str, Any]] = {
107
"internvl_chat": {"has_no_defaults_at_init": True},
108
+ "Llama_Nemotron_Nano_VL": {"attn_implementation": "eager"},
109
"NVLM_D": {"has_no_defaults_at_init": True},
110
}
111
vllm/transformers_utils/configs/__init__.py
-2
@@ -28,7 +28,6 @@ from vllm.transformers_utils.configs.mlp_speculator import MLPSpeculatorConfig
28
from vllm.transformers_utils.configs.moonvit import MoonViTConfig
29
from vllm.transformers_utils.configs.nemotron import NemotronConfig
30
from vllm.transformers_utils.configs.nemotron_h import NemotronHConfig
31
-from vllm.transformers_utils.configs.nemotron_vl import Nemotron_Nano_VL_Config
31
from vllm.transformers_utils.configs.olmo3 import Olmo3Config
32
from vllm.transformers_utils.configs.ovis import OvisConfig
33
from vllm.transformers_utils.configs.qwen3_next import Qwen3NextConfig
@@ -59,7 +58,6 @@ __all__ = [
58
"KimiVLConfig",
59
"NemotronConfig",
60
"NemotronHConfig",
62
- "Nemotron_Nano_VL_Config",
61
"Olmo3Config",
62
"OvisConfig",
63
"RadioConfig",
vllm/transformers_utils/configs/nemotron_vl.py
deleted
-60
@@ -1,60 +0,0 @@
1
-# SPDX-License-Identifier: Apache-2.0
2
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
-
4
-# ruff: noqa: E501
5
-# Adapted from
6
-# https://huggingface.co/nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1/blob/main/configuration.py
7
-# --------------------------------------------------------
8
-# Adapted from https://huggingface.co/OpenGVLab/InternVL2-Llama3-76B under MIT License
9
-# LICENSE is in incl_licenses directory.
10
-# --------------------------------------------------------
11
-
12
-from transformers import LlamaConfig
13
-from transformers.configuration_utils import PretrainedConfig
14
-from transformers.dynamic_module_utils import get_class_from_dynamic_module
15
-
16
-
17
-class Nemotron_Nano_VL_Config(PretrainedConfig):
18
- model_type = "Llama_Nemotron_Nano_VL"
19
- is_composition = True
20
-
21
- def __init__(
22
- self,
23
- vision_config=None,
24
- llm_config=None,
25
- force_image_size=None,
26
- downsample_ratio=0.5,
27
- template=None,
28
- ps_version="v1",
29
- image_tag_type="internvl",
30
- projector_hidden_size=4096,
31
- vit_hidden_size=1280,
32
- **kwargs,
33
- ):
34
- super().__init__(**kwargs)
35
-
36
- if vision_config is not None:
37
- assert (
38
- "auto_map" in vision_config
39
- and "AutoConfig" in vision_config["auto_map"]
40
- )
41
- vision_auto_config = get_class_from_dynamic_module(
42
- *vision_config["auto_map"]["AutoConfig"].split("--")[::-1]
43
- )
44
- self.vision_config = vision_auto_config(**vision_config)
45
- else:
46
- self.vision_config = PretrainedConfig()
47
-
48
- if llm_config is None:
49
- self.text_config = LlamaConfig()
50
- else:
51
- self.text_config = LlamaConfig(**llm_config)
52
-
53
- # Assign configuration values
54
- self.force_image_size = force_image_size
55
- self.downsample_ratio = downsample_ratio
56
- self.template = template # TODO move out of here and into the tokenizer
57
- self.ps_version = ps_version # Pixel shuffle version
58
- self.image_tag_type = image_tag_type # TODO: into the tokenizer too?
59
- self.projector_hidden_size = projector_hidden_size
60
- self.vit_hidden_size = vit_hidden_size