[Bugfix] Fix nightly transformers CI failure (#21427)

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

tests/models/registry.py

@@ -443,6 +443,12 @@ _MULTIMODAL_EXAMPLE_MODELS = {
                                                         hf_overrides={"architectures": ["TarsierForConditionalGeneration"]}),  # noqa: E501
     "Tarsier2ForConditionalGeneration": _HfExamplesInfo("omni-research/Tarsier2-Recap-7b",  # noqa: E501
                                                         hf_overrides={"architectures": ["Tarsier2ForConditionalGeneration"]}),  # noqa: E501
+    "VoxtralForConditionalGeneration": _HfExamplesInfo(
+        "mistralai/Voxtral-Mini-3B-2507",
+        min_transformers_version="4.54",
+        # disable this temporarily until we support HF format
+        is_available_online=False,
+    ),
     # [Encoder-decoder]
     # Florence-2 uses BartFastTokenizer which can't be loaded from AutoTokenizer
     # Therefore, we borrow the BartTokenizer from the original Bart model
@@ -450,13 +456,7 @@ _MULTIMODAL_EXAMPLE_MODELS = {
                                                          tokenizer="Isotr0py/Florence-2-tokenizer",  # noqa: E501
                                                          trust_remote_code=True),  # noqa: E501
     "MllamaForConditionalGeneration": _HfExamplesInfo("meta-llama/Llama-3.2-11B-Vision-Instruct"),  # noqa: E501
-    "VoxtralForConditionalGeneration": _HfExamplesInfo(
-        "mistralai/Voxtral-Mini-3B-2507",
-        tokenizer_mode="mistral",
-        min_transformers_version="4.54"
-    ),
     "WhisperForConditionalGeneration": _HfExamplesInfo("openai/whisper-large-v3"),  # noqa: E501
-
     # [Cross-encoder]
     "JinaVLForRanking": _HfExamplesInfo("jinaai/jina-reranker-m0"),   # noqa: E501
 }

vllm/model_executor/models/tarsier.py

@@ -13,8 +13,7 @@ from transformers import LlavaConfig as HfLlavaConfig
 from transformers import PretrainedConfig, SiglipVisionConfig
 from transformers.image_utils import ImageInput, get_image_size, to_numpy_array
 from transformers.models.llava import LlavaProcessor
-from transformers.processing_utils import (ProcessingKwargs, Unpack,
-                                           _validate_images_text_input_order)
+from transformers.processing_utils import ProcessingKwargs, Unpack
 from transformers.tokenization_utils_base import PreTokenizedInput, TextInput
 
 from vllm.config import VllmConfig
@@ -94,9 +93,6 @@ class TarsierProcessor(LlavaProcessor):
             raise ValueError(
                 "You have to specify at least one of `images` or `text`.")
 
-        # check if images and text inputs are reversed for BC
-        images, text = _validate_images_text_input_order(images, text)
-
         output_kwargs = self._merge_kwargs(
             TarsierProcessorKwargs,
             tokenizer_init_kwargs=self.tokenizer.init_kwargs,

vllm/transformers_utils/config.py

@@ -37,6 +37,7 @@ from vllm.transformers_utils.configs import (ChatGLMConfig, Cohere2Config,
                                              MiniMaxText01Config,
                                              MiniMaxVL01Config, MllamaConfig,
                                              MLPSpeculatorConfig, MPTConfig,
+                                             Nemotron_Nano_VL_Config,
                                              NemotronConfig, NVLM_D_Config,
                                              OvisConfig, RWConfig,
                                              SkyworkR1VChatConfig, SolarConfig,
@@ -80,6 +81,7 @@ _CONFIG_REGISTRY: dict[str, type[PretrainedConfig]] = {
     "dbrx": DbrxConfig,
     "deepseek_vl_v2": DeepseekVLV2Config,
     "kimi_vl": KimiVLConfig,
+    "Llama_Nemotron_Nano_VL": Nemotron_Nano_VL_Config,
     "mpt": MPTConfig,
     "RefinedWeb": RWConfig,  # For tiiuae/falcon-40b(-instruct)
     "RefinedWebModel": RWConfig,  # For tiiuae/falcon-7b(-instruct)

vllm/transformers_utils/configs/__init__.py

@@ -23,6 +23,7 @@ from vllm.transformers_utils.configs.moonvit import MoonViTConfig
 from vllm.transformers_utils.configs.mpt import MPTConfig
 from vllm.transformers_utils.configs.nemotron import NemotronConfig
 from vllm.transformers_utils.configs.nemotron_h import NemotronHConfig
+from vllm.transformers_utils.configs.nemotron_vl import Nemotron_Nano_VL_Config
 from vllm.transformers_utils.configs.nvlm_d import NVLM_D_Config
 from vllm.transformers_utils.configs.ovis import OvisConfig
 from vllm.transformers_utils.configs.skyworkr1v import SkyworkR1VChatConfig
@@ -50,6 +51,7 @@ __all__ = [
     "KimiVLConfig",
     "NemotronConfig",
     "NemotronHConfig",
+    "Nemotron_Nano_VL_Config",
     "NVLM_D_Config",
     "OvisConfig",
     "SkyworkR1VChatConfig",

vllm/transformers_utils/configs/nemotron_vl.py

@@ -0,0 +1,56 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+# yapf: disable
+# ruff: noqa: E501
+# Adapted from
+# https://huggingface.co/nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1/blob/main/configuration.py
+# --------------------------------------------------------
+# Adapted from https://huggingface.co/OpenGVLab/InternVL2-Llama3-76B under MIT License
+#     LICENSE is in incl_licenses directory.
+# --------------------------------------------------------
+
+from transformers import LlamaConfig
+from transformers.configuration_utils import PretrainedConfig
+from transformers.dynamic_module_utils import get_class_from_dynamic_module
+
+
+class Nemotron_Nano_VL_Config(PretrainedConfig):
+    model_type = 'Llama_Nemotron_Nano_VL'
+    is_composition = True
+
+    def __init__(
+        self,
+        vision_config=None,
+        llm_config=None,
+        force_image_size=None,
+        downsample_ratio=0.5,
+        template=None,
+        ps_version='v1',
+        image_tag_type="internvl",
+        projector_hidden_size=4096,
+        vit_hidden_size=1280,
+        **kwargs
+    ):
+        super().__init__(**kwargs)
+
+        if vision_config is not None:
+            assert "auto_map" in vision_config and "AutoConfig" in vision_config["auto_map"]
+            vision_auto_config = get_class_from_dynamic_module(*vision_config["auto_map"]["AutoConfig"].split("--")[::-1])
+            self.vision_config = vision_auto_config(**vision_config)
+        else:
+            self.vision_config = PretrainedConfig()
+
+        if llm_config is None:
+            self.text_config = LlamaConfig()
+        else:
+            self.text_config = LlamaConfig(**llm_config)
+
+        # Assign configuration values
+        self.force_image_size = force_image_size
+        self.downsample_ratio = downsample_ratio
+        self.template = template  # TODO move out of here and into the tokenizer
+        self.ps_version = ps_version  # Pixel shuffle version
+        self.image_tag_type = image_tag_type # TODO: into the tokenizer too?
+        self.projector_hidden_size = projector_hidden_size
+        self.vit_hidden_size = vit_hidden_size