[New model support]Support Tarsier2 (#19887)

Signed-off-by: 汪志鹏 <wangzhipeng628@gmail.com>

docs/models/supported_models.md

@@ -562,6 +562,7 @@ Specified using `--task generate`.
 | `SkyworkR1VChatModel`                        | Skywork-R1V-38B                                                          | T + I                                                                 | `Skywork/Skywork-R1V-38B`                                                                                                                               |                        | ✅︎                          | ✅︎                    |
 | `SmolVLMForConditionalGeneration`            | SmolVLM2                                                                 | T + I                                                                 | `SmolVLM2-2.2B-Instruct`                                                                                                                                | ✅︎                     |                             | ✅︎                    |
 | `TarsierForConditionalGeneration`            | Tarsier                                                                  | T + I<sup>E+</sup>                                                    | `omni-search/Tarsier-7b`,`omni-search/Tarsier-34b`                                                                                                      |                        | ✅︎                          | ✅︎                    |
+| `Tarsier2ForConditionalGeneration`<sup>^</sup>            | Tarsier2                                                                  | T + I<sup>E+</sup> + V<sup>E+</sup>                                                    | `omni-research/Tarsier2-Recap-7b`,`omni-research/Tarsier2-7b-0115`                                                                                                      |                        | ✅︎                          | ✅︎                    |
 
 <sup>^</sup> You need to set the architecture name via `--hf-overrides` to match the one in vLLM.  
 &nbsp;&nbsp;&nbsp;&nbsp;• For example, to use DeepSeek-VL2 series models:  

examples/offline_inference/vision_language.py

@@ -1040,6 +1040,37 @@ def run_qwen2_5_omni(questions: list[str], modality: str):
     )
 
 
+def run_tarsier2(questions: list[str], modality: str) -> ModelRequestData:
+    model_name = "omni-research/Tarsier2-Recap-7b"
+
+    engine_args = EngineArgs(
+        model=model_name,
+        max_model_len=4096,
+        hf_overrides={"architectures": ["Tarsier2ForConditionalGeneration"]},
+        limit_mm_per_prompt={modality: 1},
+    )
+
+    if modality == "image":
+        placeholder = "<|image_pad|>"
+    elif modality == "video":
+        placeholder = "<|video_pad|>"
+
+    prompts = [
+        (
+            "<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n"
+            f"<|im_start|>user\n<|vision_start|>{placeholder}<|vision_end|>"
+            f"{question}<|im_end|>\n"
+            "<|im_start|>assistant\n"
+        )
+        for question in questions
+    ]
+
+    return ModelRequestData(
+        engine_args=engine_args,
+        prompts=prompts,
+    )
+
+
 # SkyworkR1V
 def run_skyworkr1v(questions: list[str], modality: str) -> ModelRequestData:
     assert modality == "image"
@@ -1112,6 +1143,7 @@ model_example_map = {
     "skywork_chat": run_skyworkr1v,
     "smolvlm": run_smolvlm,
     "tarsier": run_tarsier,
+    "tarsier2": run_tarsier2,
 }
 
 

examples/offline_inference/vision_language_multi_image.py

@@ -828,6 +828,32 @@ def load_tarsier(question: str, image_urls: list[str]) -> ModelRequestData:
     )
 
 
+def load_tarsier2(question: str, image_urls: list[str]) -> ModelRequestData:
+    model_name = "omni-research/Tarsier2-Recap-7b"
+
+    engine_args = EngineArgs(
+        model=model_name,
+        trust_remote_code=True,
+        max_model_len=32768,
+        limit_mm_per_prompt={"image": len(image_urls)},
+        hf_overrides={"architectures": ["Tarsier2ForConditionalGeneration"]},
+    )
+
+    prompt = (
+        "<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n"
+        f"<|im_start|>user\n<|vision_start|>{'<|image_pad|>' * len(image_urls)}"
+        f"<|vision_end|>{question}<|im_end|>\n"
+        "<|im_start|>assistant\n"
+    )
+    image_data = [fetch_image(url) for url in image_urls]
+
+    return ModelRequestData(
+        engine_args=engine_args,
+        prompt=prompt,
+        image_data=image_data,
+    )
+
+
 model_example_map = {
     "aria": load_aria,
     "aya_vision": load_aya_vision,
@@ -853,6 +879,7 @@ model_example_map = {
     "qwen2_5_vl": load_qwen2_5_vl,
     "smolvlm": load_smolvlm,
     "tarsier": load_tarsier,
+    "tarsier2": load_tarsier2,
 }
 
 

tests/models/multimodal/processing/test_common.py

@@ -284,6 +284,7 @@ def _test_processing_correctness_one(
     "fixie-ai/ultravox-v0_5-llama-3_2-1b",
     "openai/whisper-large-v3",
     "omni-research/Tarsier-7b",
+    "omni-research/Tarsier2-Recap-7b"
 ])
 @pytest.mark.parametrize("hit_rate", [0.3, 0.5, 1.0])
 @pytest.mark.parametrize("num_batches", [32])

tests/models/registry.py

@@ -398,6 +398,8 @@ _MULTIMODAL_EXAMPLE_MODELS = {
                                      trust_remote_code=True),
     "TarsierForConditionalGeneration": _HfExamplesInfo("omni-research/Tarsier-7b",  # noqa: E501
                                                         hf_overrides={"architectures": ["TarsierForConditionalGeneration"]}),  # noqa: E501
+    "Tarsier2ForConditionalGeneration": _HfExamplesInfo("omni-research/Tarsier2-Recap-7b",  # noqa: E501
+                                                        hf_overrides={"architectures": ["Tarsier2ForConditionalGeneration"]}),  # noqa: E501
     # [Encoder-decoder]
     # Florence-2 uses BartFastTokenizer which can't be loaded from AutoTokenizer
     # Therefore, we borrow the BartTokenizer from the original Bart model

vllm/model_executor/models/qwen2_vl.py

@@ -32,12 +32,14 @@ import torch
 import torch.nn as nn
 import torch.nn.functional as F
 from einops import rearrange, repeat
-from transformers import BatchFeature
+from transformers import AutoConfig, BatchFeature
 from transformers.models.qwen2_vl import (Qwen2VLImageProcessor,
                                           Qwen2VLProcessor)
 from transformers.models.qwen2_vl.configuration_qwen2_vl import (
     Qwen2VLConfig, Qwen2VLVisionConfig)
 from transformers.models.qwen2_vl.image_processing_qwen2_vl import smart_resize
+from transformers.models.qwen2_vl.video_processing_qwen2_vl import (
+    Qwen2VLVideoProcessor)
 
 from vllm.config import VllmConfig
 from vllm.distributed import parallel_state, tensor_model_parallel_all_gather
@@ -69,6 +71,7 @@ from vllm.sequence import IntermediateTensors
 from vllm.transformers_utils.config import uses_mrope
 from vllm.transformers_utils.processor import (
     cached_image_processor_from_config)
+from vllm.transformers_utils.tokenizer import AnyTokenizer
 
 from .interfaces import (MultiModalEmbeddings, SupportsLoRA,
                          SupportsMultiModal, SupportsPP)
@@ -1405,3 +1408,87 @@ class Qwen2VLForConditionalGeneration(nn.Module, SupportsMultiModal,
             connector="visual.merger.",
             tower_model="visual.",
         )
+
+
+class Tarsier2MultiModalProcessor(Qwen2VLMultiModalProcessor):
+    pass
+
+
+class Tarsier2ImageProcessor(Qwen2VLImageProcessor):
+
+    def __init__(
+        self,
+        size: Optional[dict[str, int]] = None,
+        **kwargs,
+    ) -> None:
+        if size is not None and "min_pixels" in size and "max_pixels" in size:
+            # Remap if Tarsier2-specific format is provided
+            remapped_size = {
+                "shortest_edge": size["min_pixels"],
+                "longest_edge": size["max_pixels"]
+            }
+            super().__init__(size=remapped_size, **kwargs)
+        else:
+            super().__init__(size=size, **kwargs)
+
+
+class Tarsier2Processor(Qwen2VLProcessor):
+
+    def __init__(
+        self,
+        vision_config: dict,
+        tokenizer: AnyTokenizer,
+        **kwargs,
+    ):
+        self.image_processor = Tarsier2ImageProcessor(**vision_config)
+        super().__init__(image_processor=self.image_processor,
+                         tokenizer=tokenizer,
+                         video_processor=Qwen2VLVideoProcessor(),
+                         chat_template=None,
+                         **kwargs)
+
+
+class Tarsier2ProcessingInfo(Qwen2VLProcessingInfo):
+
+    def get_hf_config(self) -> Qwen2VLConfig:
+        model_path = self.ctx.model_config.model
+        original_config = AutoConfig.from_pretrained(model_path)
+        config_dict = original_config.to_dict()
+        correct_config = Qwen2VLConfig.from_dict(config_dict)
+
+        return correct_config
+
+    def get_hf_processor(self, **kwargs: object) -> Tarsier2Processor:
+        return Tarsier2Processor(
+            vision_config=self.ctx.get_hf_image_processor_config(),
+            tokenizer=self.get_tokenizer(),
+            **kwargs,
+        )
+
+    def get_image_processor(self) -> Tarsier2ImageProcessor:
+        return Tarsier2ImageProcessor(
+            **self.ctx.get_hf_image_processor_config())
+
+
+@MULTIMODAL_REGISTRY.register_processor(Tarsier2MultiModalProcessor,
+                                        info=Tarsier2ProcessingInfo,
+                                        dummy_inputs=Qwen2VLDummyInputsBuilder)
+class Tarsier2ForConditionalGeneration(Qwen2VLForConditionalGeneration):
+    hf_to_vllm_mapper = WeightsMapper(orig_to_new_prefix={
+        "vision_tower.": "visual.",
+    })
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        # Tarsier2 uses llava as model_type, which will create a Qwen2VLConfig
+        # as text_config, we need to reconstruct Qwen2VLConfig from LlavaConfig.
+        config = vllm_config.model_config.hf_config
+        qwen2vl_config = config.text_config
+        qwen2vl_config.architectures = config.architectures
+        vllm_config.model_config.hf_config = qwen2vl_config
+        super().__init__(vllm_config=vllm_config, prefix=prefix)
+
+    def load_weights(self, weights: Iterable[tuple[str,
+                                                   torch.Tensor]]) -> set[str]:
+
+        loader = AutoWeightsLoader(self)
+        return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)

vllm/model_executor/models/registry.py

@@ -217,6 +217,7 @@ _MULTIMODAL_MODELS = {
     "UltravoxModel": ("ultravox", "UltravoxModel"),
     "Phi4MMForCausalLM": ("phi4mm", "Phi4MMForCausalLM"),
     "TarsierForConditionalGeneration": ("tarsier", "TarsierForConditionalGeneration"),  # noqa: E501
+    "Tarsier2ForConditionalGeneration": ("qwen2_vl", "Tarsier2ForConditionalGeneration"),  # noqa: E501
     # [Encoder-decoder]
     "Florence2ForConditionalGeneration": ("florence2", "Florence2ForConditionalGeneration"),  # noqa: E501
     "MllamaForConditionalGeneration": ("mllama", "MllamaForConditionalGeneration"),  # noqa: E501