Add tarsier model support (#18985)

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

tests/models/multimodal/processing/test_common.py

@@ -282,6 +282,7 @@ def _test_processing_correctness_one(
     "Skywork/Skywork-R1V-38B",
     "fixie-ai/ultravox-v0_5-llama-3_2-1b",
     "openai/whisper-large-v3",
+    "omni-research/Tarsier-7b",
 ])
 @pytest.mark.parametrize("hit_rate", [0.3, 0.5, 1.0])
 @pytest.mark.parametrize("num_batches", [32])

tests/models/registry.py

@@ -406,6 +406,8 @@ _MULTIMODAL_EXAMPLE_MODELS = {
     "SmolVLMForConditionalGeneration": _HfExamplesInfo("HuggingFaceTB/SmolVLM2-2.2B-Instruct"),  # noqa: E501
     "UltravoxModel": _HfExamplesInfo("fixie-ai/ultravox-v0_5-llama-3_2-1b",  # noqa: E501
                                      trust_remote_code=True),
+    "TarsierForConditionalGeneration": _HfExamplesInfo("omni-research/Tarsier-7b",  # noqa: E501
+                                                        hf_overrides={"architectures": ["TarsierForConditionalGeneration"]}),  # 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