[Model] Add smolvlm support (#16017)

Signed-off-by: chaunceyjiang <chaunceyjiang@gmail.com>

tests/models/decoder_only/vision_language/test_models.py

@@ -493,6 +493,16 @@ VLM_TEST_SETTINGS = {
         patch_hf_runner=model_utils.skyworkr1v_patch_hf_runner,
         marks=[large_gpu_mark(min_gb=80)],
     ),
+    "smolvlm": VLMTestInfo(
+        models=["HuggingFaceTB/SmolVLM2-2.2B-Instruct"],
+        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
+        prompt_formatter=lambda img_prompt:f"<|im_start|>User:{img_prompt}<end_of_utterance>\nAssistant:",  # noqa: E501
+        img_idx_to_prompt=lambda idx: "<image>",
+        max_model_len=8192,
+        max_num_seqs=2,
+        auto_cls=AutoModelForImageTextToText,
+        hf_output_post_proc=model_utils.smolvlm_trunc_hf_output,
+    ),
     ### Tensor parallel / multi-gpu broadcast tests
     "chameleon-broadcast": VLMTestInfo(
         models=["facebook/chameleon-7b"],

tests/models/decoder_only/vision_language/vlm_utils/model_utils.py

@@ -204,6 +204,12 @@ def idefics3_trunc_hf_output(hf_output: RunnerOutput,
     return output_ids, output_str, out_logprobs
 
 
+def smolvlm_trunc_hf_output(hf_output: RunnerOutput,
+                            model: str) -> RunnerOutput:
+    # Based on Idefics3
+    return idefics3_trunc_hf_output(hf_output, model)
+
+
 def minicpmv_trunc_hf_output(hf_output: RunnerOutput,
                              model: str) -> RunnerOutput:
     output_ids, output_str, out_logprobs = hf_output

tests/models/multimodal/processing/test_common.py

@@ -257,6 +257,7 @@ def _test_processing_correctness_mistral(
     "h2oai/h2ovl-mississippi-800m",
     "OpenGVLab/InternVL2-1B",
     "HuggingFaceM4/Idefics3-8B-Llama3",
+    "HuggingFaceTB/SmolVLM2-2.2B-Instruct",
     "meta-llama/Llama-4-Scout-17B-16E-Instruct",
     "llava-hf/llava-1.5-7b-hf",
     "llava-hf/llava-v1.6-mistral-7b-hf",

tests/models/multimodal/processing/test_smolvlm.py

@@ -0,0 +1,65 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Tests for smolvlm's multimodal preprocessing kwargs."""
+import pytest
+from transformers import SmolVLMConfig
+
+from vllm.multimodal import MULTIMODAL_REGISTRY
+
+from ....conftest import _ImageAssets
+from ...utils import build_model_context
+
+
+@pytest.mark.parametrize("model_id", ["HuggingFaceTB/SmolVLM2-2.2B-Instruct"])
+# yapf: disable
+@pytest.mark.parametrize(
+    ("mm_processor_kwargs", "expected_toks_per_img"),
+    [
+        ({"max_image_size": {"longest_edge": 384}}, 1377),
+        ({"max_image_size": {"longest_edge": 768}}, 405),
+    ])
+# yapf: enable
+@pytest.mark.parametrize("num_imgs", [1, 2])
+@pytest.mark.parametrize("kwargs_on_init", [True, False])
+def test_processor_override(
+    image_assets: _ImageAssets,
+    model_id: str,
+    mm_processor_kwargs: dict[str, object],
+    expected_toks_per_img: int,
+    num_imgs: int,
+    kwargs_on_init: bool,
+):
+    """Ensure Idefics3MultiModalProcessor handles num_crops properly."""
+    # Same as the previous test - don't initialize mm_processor_kwargs
+    # in this test and assume that the kwargs will be correctly expanded by
+    # the partial when calling the custom input processor.
+    ctx = build_model_context(
+        model_id,
+        mm_processor_kwargs=mm_processor_kwargs if kwargs_on_init else None,
+        limit_mm_per_prompt={"image": num_imgs},
+    )
+    processor = MULTIMODAL_REGISTRY.create_processor(ctx.model_config)
+    hf_processor_mm_kwargs = {} if kwargs_on_init else mm_processor_kwargs
+
+    # Build the image str / prompt based on the number of images we pass
+    placeholders = "<image>" if num_imgs == 1 else "\n".join(
+        f"Image-{i}: <image>\n" for i in range(1, num_imgs + 1))
+    prompt = f"<|im_start|>User:{placeholders}\n<end_of_utterance>\nAssistant:"  # noqa: E501
+
+    # Build mm_data
+    image_size = ctx.get_hf_config(SmolVLMConfig).vision_config.image_size
+    dummy_image_size = (image_size * 4, image_size * 4)
+    dummy_image = image_assets[0].pil_image.resize(dummy_image_size)
+    mm_data = {"image": [dummy_image] * num_imgs}
+
+    processed_inputs = processor.apply(prompt, mm_data, hf_processor_mm_kwargs)
+
+    # Ensure the placeholders format are correct
+    hf_processor = processor.info.get_hf_processor(**hf_processor_mm_kwargs)
+    hf_processed_inputs = hf_processor(text=prompt, images=mm_data["image"])
+    assert processed_inputs["prompt_token_ids"] == hf_processed_inputs[
+        "input_ids"][0]
+
+    # Ensure we have the right number of placeholders per num_crops size
+    image_token_id = ctx.get_hf_config().image_token_id
+    img_tok_count = processed_inputs["prompt_token_ids"].count(image_token_id)
+    assert img_tok_count == expected_toks_per_img * num_imgs

tests/models/registry.py

@@ -344,6 +344,7 @@ _MULTIMODAL_EXAMPLE_MODELS = {
     "Qwen2_5_VLForConditionalGeneration": _HfExamplesInfo("Qwen/Qwen2.5-VL-3B-Instruct",  # noqa: E501
                                                           min_transformers_version="4.49"),  # noqa: E501
     "SkyworkR1VChatModel": _HfExamplesInfo("Skywork/Skywork-R1V-38B"),
+    "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,
                                      max_transformers_version="4.50"),