[Model] Siglip Embedding Support (#27324)

Signed-off-by: piood <2477084691@qq.com>

docs/models/supported_models.md

@@ -800,12 +800,13 @@ These models primarily support the [`LLM.embed`](./pooling_models.md#llmembed) A
 
 The following table lists those that are tested in vLLM.
 
-| Architecture | Models | Inputs | Example HF Models | [LoRA](../features/lora.md) | [PP](../serving/parallelism_scaling.md) |
-|--------------|--------|--------|-------------------|----------------------|---------------------------|
-| `CLIPModel` | CLIP | T / I | `openai/clip-vit-base-patch32`, `openai/clip-vit-large-patch14`, etc. | | |
-| `LlavaNextForConditionalGeneration`<sup>C</sup> | LLaVA-NeXT-based | T / I | `royokong/e5-v` | | ✅︎ |
-| `Phi3VForCausalLM`<sup>C</sup> | Phi-3-Vision-based | T + I | `TIGER-Lab/VLM2Vec-Full` | | ✅︎ |
-| `*ForConditionalGeneration`<sup>C</sup>, `*ForCausalLM`<sup>C</sup>, etc. | Generative models | \* | N/A | \* | \* |
+| Architecture | Models | Inputs | Example HF Models | [LoRA](../features/lora.md) | [PP](../serving/parallelism_scaling.md) | [V1](gh-issue:8779) |
+|--------------|--------|--------|-------------------|----------------------|---------------------------|---------------------|
+| `CLIPModel` | CLIP | T / I | `openai/clip-vit-base-patch32`, `openai/clip-vit-large-patch14`, etc. | | | ✅︎ |
+| `LlavaNextForConditionalGeneration`<sup>C</sup> | LLaVA-NeXT-based | T / I | `royokong/e5-v` | | ✅︎ | ✅︎ |
+| `Phi3VForCausalLM`<sup>C</sup> | Phi-3-Vision-based | T + I | `TIGER-Lab/VLM2Vec-Full` | | ✅︎ | ✅︎ |
+| `SiglipModel` | SigLIP | T / I | `google/siglip-base-patch16-224` | | | ✅︎ |
+| `*ForConditionalGeneration`<sup>C</sup>, `*ForCausalLM`<sup>C</sup>, etc. | Generative models | \* | N/A | \* | \* | \* |
 
 <sup>C</sup> Automatically converted into an embedding model via `--convert embed`. ([details](./pooling_models.md#model-conversion))  
 \* Feature support is the same as that of the original model.

examples/offline_inference/vision_language_pooling.py

@@ -110,6 +110,53 @@ def run_e5_v(query: Query) -> ModelRequestData:
     )
 
 
+def run_jinavl_reranker(query: Query) -> ModelRequestData:
+    if query["modality"] != "text+images":
+        raise ValueError(f"Unsupported query modality: '{query['modality']}'")
+
+    engine_args = EngineArgs(
+        model="jinaai/jina-reranker-m0",
+        runner="pooling",
+        max_model_len=32768,
+        trust_remote_code=True,
+        mm_processor_kwargs={
+            "min_pixels": 3136,
+            "max_pixels": 602112,
+        },
+        limit_mm_per_prompt={"image": 1},
+    )
+
+    return ModelRequestData(
+        engine_args=engine_args,
+        query=query["text"],
+        documents=query["image"],
+    )
+
+
+def run_siglip(query: Query) -> ModelRequestData:
+    if query["modality"] == "text":
+        prompt = query["text"]
+        image = None
+    elif query["modality"] == "image":
+        prompt = ""  # For image input, make sure that the prompt text is empty
+        image = query["image"]
+    else:
+        modality = query["modality"]
+        raise ValueError(f"Unsupported query modality: '{modality}'")
+
+    engine_args = EngineArgs(
+        model="google/siglip-base-patch16-224",
+        runner="pooling",
+        limit_mm_per_prompt={"image": 1},
+    )
+
+    return ModelRequestData(
+        engine_args=engine_args,
+        prompt=prompt,
+        image=image,
+    )
+
+
 def _get_vlm2vec_prompt_image(query: Query, image_token: str):
     if query["modality"] == "text":
         text = query["text"]
@@ -211,29 +258,6 @@ def run_vlm2vec_qwen2vl(query: Query) -> ModelRequestData:
     )
 
 
-def run_jinavl_reranker(query: Query) -> ModelRequestData:
-    if query["modality"] != "text+images":
-        raise ValueError(f"Unsupported query modality: '{query['modality']}'")
-
-    engine_args = EngineArgs(
-        model="jinaai/jina-reranker-m0",
-        runner="pooling",
-        max_model_len=32768,
-        trust_remote_code=True,
-        mm_processor_kwargs={
-            "min_pixels": 3136,
-            "max_pixels": 602112,
-        },
-        limit_mm_per_prompt={"image": 1},
-    )
-
-    return ModelRequestData(
-        engine_args=engine_args,
-        query=query["text"],
-        documents=query["image"],
-    )
-
-
 def get_query(modality: QueryModality):
     if modality == "text":
         return TextQuery(modality="text", text="A dog sitting in the grass")
@@ -328,9 +352,10 @@ def run_score(model: str, modality: QueryModality, seed: int | None):
 model_example_map = {
     "clip": run_clip,
     "e5_v": run_e5_v,
+    "jinavl_reranker": run_jinavl_reranker,
+    "siglip": run_siglip,
     "vlm2vec_phi3v": run_vlm2vec_phi3v,
     "vlm2vec_qwen2vl": run_vlm2vec_qwen2vl,
-    "jinavl_reranker": run_jinavl_reranker,
 }
 
 

examples/online_serving/pooling/openai_chat_embedding_client_for_multimodal.py

@@ -83,25 +83,29 @@ def run_clip(client: OpenAI, model: str):
     print("Text embedding output:", response.data[0].embedding)
 
 
-def run_vlm2vec(client: OpenAI, model: str):
+def run_dse_qwen2_vl(client: OpenAI, model: str):
     """
     Start the server using:
 
-    vllm serve TIGER-Lab/VLM2Vec-Full \
+    vllm serve MrLight/dse-qwen2-2b-mrl-v1 \
         --runner pooling \
         --trust-remote-code \
-        --max-model-len 4096 \
-        --chat-template examples/template_vlm2vec_phi3v.jinja
+        --max-model-len 8192 \
+        --chat-template examples/template_dse_qwen2_vl.jinja
     """
-
     response = create_chat_embeddings(
         client,
         messages=[
             {
                 "role": "user",
                 "content": [
-                    {"type": "image_url", "image_url": {"url": image_url}},
-                    {"type": "text", "text": "Represent the given image."},
+                    {
+                        "type": "image_url",
+                        "image_url": {
+                            "url": image_url,
+                        },
+                    },
+                    {"type": "text", "text": "What is shown in this image?"},
                 ],
             }
         ],
@@ -111,17 +115,26 @@ def run_vlm2vec(client: OpenAI, model: str):
 
     print("Image embedding output:", response.data[0].embedding)
 
+    # MrLight/dse-qwen2-2b-mrl-v1 requires a placeholder image
+    # of the minimum input size
+    buffer = io.BytesIO()
+    image_placeholder = Image.new("RGB", (56, 56))
+    image_placeholder.save(buffer, "png")
+    buffer.seek(0)
+    image_placeholder = base64.b64encode(buffer.read()).decode("utf-8")
     response = create_chat_embeddings(
         client,
         messages=[
             {
                 "role": "user",
                 "content": [
-                    {"type": "image_url", "image_url": {"url": image_url}},
                     {
-                        "type": "text",
-                        "text": "Represent the given image with the following question: What is in the image.",
+                        "type": "image_url",
+                        "image_url": {
+                            "url": f"data:image/jpeg;base64,{image_placeholder}",
+                        },
                     },
+                    {"type": "text", "text": "Query: What is the weather like today?"},
                 ],
             }
         ],
@@ -129,7 +142,16 @@ def run_vlm2vec(client: OpenAI, model: str):
         encoding_format="float",
     )
 
-    print("Image+Text embedding output:", response.data[0].embedding)
+    print("Text embedding output:", response.data[0].embedding)
+
+
+def run_siglip(client: OpenAI, model: str):
+    """
+    Start the server using:
+
+    vllm serve google/siglip-base-patch16-224 \
+        --runner pooling
+    """
 
     response = create_chat_embeddings(
         client,
@@ -137,7 +159,23 @@ def run_vlm2vec(client: OpenAI, model: str):
             {
                 "role": "user",
                 "content": [
-                    {"type": "text", "text": "A cat and a dog"},
+                    {"type": "image_url", "image_url": {"url": image_url}},
+                ],
+            }
+        ],
+        model=model,
+        encoding_format="float",
+    )
+
+    print("Image embedding output:", response.data[0].embedding)
+
+    response = create_chat_embeddings(
+        client,
+        messages=[
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "a photo of a cat"},
                 ],
             }
         ],
@@ -148,29 +186,25 @@ def run_vlm2vec(client: OpenAI, model: str):
     print("Text embedding output:", response.data[0].embedding)
 
 
-def run_dse_qwen2_vl(client: OpenAI, model: str):
+def run_vlm2vec(client: OpenAI, model: str):
     """
     Start the server using:
 
-    vllm serve MrLight/dse-qwen2-2b-mrl-v1 \
+    vllm serve TIGER-Lab/VLM2Vec-Full \
         --runner pooling \
         --trust-remote-code \
-        --max-model-len 8192 \
-        --chat-template examples/template_dse_qwen2_vl.jinja
+        --max-model-len 4096 \
+        --chat-template examples/template_vlm2vec_phi3v.jinja
     """
+
     response = create_chat_embeddings(
         client,
         messages=[
             {
                 "role": "user",
                 "content": [
-                    {
-                        "type": "image_url",
-                        "image_url": {
-                            "url": image_url,
-                        },
-                    },
-                    {"type": "text", "text": "What is shown in this image?"},
+                    {"type": "image_url", "image_url": {"url": image_url}},
+                    {"type": "text", "text": "Represent the given image."},
                 ],
             }
         ],
@@ -180,26 +214,33 @@ def run_dse_qwen2_vl(client: OpenAI, model: str):
 
     print("Image embedding output:", response.data[0].embedding)
 
-    # MrLight/dse-qwen2-2b-mrl-v1 requires a placeholder image
-    # of the minimum input size
-    buffer = io.BytesIO()
-    image_placeholder = Image.new("RGB", (56, 56))
-    image_placeholder.save(buffer, "png")
-    buffer.seek(0)
-    image_placeholder = base64.b64encode(buffer.read()).decode("utf-8")
     response = create_chat_embeddings(
         client,
         messages=[
             {
                 "role": "user",
                 "content": [
+                    {"type": "image_url", "image_url": {"url": image_url}},
                     {
-                        "type": "image_url",
-                        "image_url": {
-                            "url": f"data:image/jpeg;base64,{image_placeholder}",
-                        },
+                        "type": "text",
+                        "text": "Represent the given image with the following question: What is in the image.",
                     },
-                    {"type": "text", "text": "Query: What is the weather like today?"},
+                ],
+            }
+        ],
+        model=model,
+        encoding_format="float",
+    )
+
+    print("Image+Text embedding output:", response.data[0].embedding)
+
+    response = create_chat_embeddings(
+        client,
+        messages=[
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "A cat and a dog"},
                 ],
             }
         ],
@@ -212,8 +253,9 @@ def run_dse_qwen2_vl(client: OpenAI, model: str):
 
 model_example_map = {
     "clip": run_clip,
-    "vlm2vec": run_vlm2vec,
     "dse_qwen2_vl": run_dse_qwen2_vl,
+    "siglip": run_siglip,
+    "vlm2vec": run_vlm2vec,
 }
 
 

tests/models/multimodal/pooling/test_siglip.py

+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import pytest
+from transformers import SiglipModel
+
+from ....conftest import IMAGE_ASSETS, HfRunner, PromptImageInput, VllmRunner
+from ...utils import check_embeddings_close
+
+HF_TEXT_PROMPTS = [
+    "a photo of a stop sign",
+    "a photo of a cherry blossom",
+]
+
+HF_IMAGE_PROMPTS = IMAGE_ASSETS.prompts(
+    {
+        "stop_sign": "",
+        "cherry_blossom": "",
+    }
+)
+
+MODELS = ["google/siglip-base-patch16-224"]
+
+
+def _run_test(
+    hf_runner: type[HfRunner],
+    vllm_runner: type[VllmRunner],
+    input_texts: list[str],
+    input_images: PromptImageInput,
+    model: str,
+    *,
+    dtype: str,
+) -> None:
+    with vllm_runner(
+        model, runner="pooling", dtype=dtype, enforce_eager=True, max_model_len=64
+    ) as vllm_model:
+        vllm_outputs = vllm_model.embed(input_texts, images=input_images)
+
+    with hf_runner(model, dtype=dtype, auto_cls=SiglipModel) as hf_model:
+        all_inputs = hf_model.get_inputs(input_texts, images=input_images)
+
+        all_outputs = []
+        for inputs in all_inputs:
+            inputs = hf_model.wrap_device(inputs)
+
+            if "pixel_values" in inputs:
+                pooled_output = hf_model.model.get_image_features(
+                    pixel_values=inputs.pixel_values,
+                ).squeeze(0)
+            else:
+                pooled_output = hf_model.model.get_text_features(
+                    input_ids=inputs.input_ids,
+                ).squeeze(0)
+
+            all_outputs.append(pooled_output.tolist())
+
+        hf_outputs = all_outputs
+
+    check_embeddings_close(
+        embeddings_0_lst=hf_outputs,
+        embeddings_1_lst=vllm_outputs,
+        name_0="hf",
+        name_1="vllm",
+    )
+
+
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("dtype", ["float"])
+def test_models_text(
+    hf_runner,
+    vllm_runner,
+    image_assets,
+    model: str,
+    dtype: str,
+) -> None:
+    input_texts_images = [(text, None) for text in HF_TEXT_PROMPTS]
+    input_texts = [text for text, _ in input_texts_images]
+    input_images = [image for _, image in input_texts_images]
+
+    _run_test(
+        hf_runner,
+        vllm_runner,
+        input_texts,
+        input_images,  # type: ignore
+        model,
+        dtype=dtype,
+    )
+
+
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("dtype", ["float"])
+def test_models_image(
+    hf_runner,
+    vllm_runner,
+    image_assets,
+    model: str,
+    dtype: str,
+) -> None:
+    input_texts_images = [
+        (text, asset.pil_image) for text, asset in zip(HF_IMAGE_PROMPTS, image_assets)
+    ]
+    input_texts = [text for text, _ in input_texts_images]
+    input_images = [image for _, image in input_texts_images]
+
+    _run_test(
+        hf_runner,
+        vllm_runner,
+        input_texts,
+        input_images,
+        model,
+        dtype=dtype,
+    )
+
+
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("dtype", ["float"])
+def test_models_text_image_no_crash(
+    vllm_runner,
+    image_assets,
+    model: str,
+    dtype: str,
+) -> None:
+    texts = [HF_TEXT_PROMPTS[0]]
+    images = [image_assets[0].pil_image]
+
+    with vllm_runner(
+        model,
+        runner="pooling",
+        dtype=dtype,
+        enforce_eager=True,
+        max_model_len=64,
+    ) as vllm_model:
+        with pytest.raises(ValueError, match="not both"):
+            vllm_model.embed(texts, images=images)
+
+        vllm_model.embed(texts)
+        vllm_model.embed([""], images=images)

tests/models/registry.py

@@ -471,6 +471,7 @@ _EMBEDDING_EXAMPLE_MODELS = {
         "TIGER-Lab/VLM2Vec-Full", trust_remote_code=True
     ),
     "Qwen2VLForConditionalGeneration": _HfExamplesInfo("MrLight/dse-qwen2-2b-mrl-v1"),
+    "SiglipModel": _HfExamplesInfo("google/siglip-base-patch16-224"),
     "PrithviGeoSpatialMAE": _HfExamplesInfo(
         "ibm-nasa-geospatial/Prithvi-EO-2.0-300M-TL-Sen1Floods11",
         dtype="float16",

vllm/model_executor/models/registry.py

@@ -209,6 +209,7 @@ _EMBEDDING_MODELS = {
     ),
     "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
     "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),  # noqa: E501
+    "SiglipModel": ("siglip", "SiglipEmbeddingModel"),
     # Technically Terratorch models work on images, both in
     # input and output. I am adding it here because it piggy-backs on embedding
     # models for the time being.

vllm/transformers_utils/chat_templates/registry.py

@@ -31,14 +31,15 @@ def _get_minicpmv_chat_template_fallback(tokenizer_name_or_path: str) -> Path |
 
 _MODEL_TYPE_TO_CHAT_TEMPLATE_FALLBACK: dict[str, ChatTemplatePath] = {
     "blip-2": CHAT_TEMPLATES_DIR / "template_blip2.jinja",
-    "clip": CHAT_TEMPLATES_DIR / "template_basic.jinja",
     "chameleon": CHAT_TEMPLATES_DIR / "template_basic.jinja",
-    "deepseek_vl_v2": CHAT_TEMPLATES_DIR / "template_deepseek_vl2.jinja",
+    "clip": CHAT_TEMPLATES_DIR / "template_basic.jinja",
     "deepseek_ocr": CHAT_TEMPLATES_DIR / "template_deepseek_ocr.jinja",
+    "deepseek_vl_v2": CHAT_TEMPLATES_DIR / "template_deepseek_vl2.jinja",
     "fuyu": CHAT_TEMPLATES_DIR / "template_fuyu.jinja",
     "minicpmv": _get_minicpmv_chat_template_fallback,
     "paligemma": CHAT_TEMPLATES_DIR / "template_basic.jinja",
     "qwen": _get_qwen_chat_template_fallback,
+    "siglip": CHAT_TEMPLATES_DIR / "template_basic.jinja",
 }