[Bugfix] Fix Qwen3-VL-Reranker load. (#33298)

Signed-off-by: wang.yuqi <yuqi.wang@daocloud.io>
Signed-off-by: wang.yuqi <noooop@126.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
(cherry picked from commit abb34ac4)

examples/pooling/score/vision_rerank_api_online.py

@@ -18,48 +18,32 @@ e.g.
 """
 
 import argparse
-import base64
-import json
+import pprint
 
 import requests
 
-
-def encode_base64_content_from_url(content_url: str) -> dict[str, str]:
-    """Encode a content retrieved from a remote url to base64 format."""
-
-    with requests.get(content_url, headers=headers) as response:
-        response.raise_for_status()
-        result = base64.b64encode(response.content).decode("utf-8")
-
-    return {"url": f"data:image/jpeg;base64,{result}"}
-
-
-headers = {"accept": "application/json", "Content-Type": "application/json"}
+from vllm.multimodal.utils import encode_image_url, fetch_image
 
 query = "A woman playing with her dog on a beach at sunset."
-documents = {
-    "content": [
-        {
-            "type": "text",
-            "text": (
+document = (
     "A woman shares a joyful moment with her golden retriever on a sun-drenched beach at sunset, "
     "as the dog offers its paw in a heartwarming display of companionship and trust."
-            ),
+)
+image_url = "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg"
+documents = [
+    {
+        "type": "text",
+        "text": document,
     },
     {
         "type": "image_url",
-            "image_url": {
-                "url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg"
-            },
+        "image_url": {"url": image_url},
     },
     {
         "type": "image_url",
-            "image_url": encode_base64_content_from_url(
-                "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg"
-            ),
+        "image_url": {"url": encode_image_url(fetch_image(image_url))},
     },
-    ]
-}
+]
 
 
 def parse_args():
@@ -74,23 +58,36 @@ def main(args):
     models_url = base_url + "/v1/models"
     rerank_url = base_url + "/rerank"
 
-    response = requests.get(models_url, headers=headers)
+    response = requests.get(models_url)
     model = response.json()["data"][0]["id"]
 
-    data = {
+    print("Query: string & Document: list of string")
+    prompt = {"model": model, "query": query, "documents": [document]}
+    response = requests.post(rerank_url, json=prompt)
+    pprint.pprint(response.json())
+
+    print("Query: string & Document: text")
+    prompt = {"model": model, "query": query, "documents": {"content": [documents[0]]}}
+    response = requests.post(rerank_url, json=prompt)
+    pprint.pprint(response.json())
+
+    print("Query: string & Document: image url")
+    prompt = {
+        "model": model,
+        "query": query,
+        "documents": {"content": [documents[1]]},
+    }
+    response = requests.post(rerank_url, json=prompt)
+    pprint.pprint(response.json())
+
+    print("Query: string & Document: image base64")
+    prompt = {
         "model": model,
         "query": query,
-        "documents": documents,
+        "documents": {"content": [documents[2]]},
     }
-    response = requests.post(rerank_url, headers=headers, json=data)
-
-    # Check the response
-    if response.status_code == 200:
-        print("Request successful!")
-        print(json.dumps(response.json(), indent=2))
-    else:
-        print(f"Request failed with status code: {response.status_code}")
-        print(response.text)
+    response = requests.post(rerank_url, json=prompt)
+    pprint.pprint(response.json())
 
 
 if __name__ == "__main__":

examples/pooling/score/vision_score_api_online.py

@@ -17,48 +17,32 @@ e.g.
 """
 
 import argparse
-import base64
-import json
 import pprint
 
 import requests
 
+from vllm.multimodal.utils import encode_image_url, fetch_image
 
-def encode_base64_content_from_url(content_url: str) -> dict[str, str]:
-    """Encode a content retrieved from a remote url to base64 format."""
-
-    with requests.get(content_url, headers=headers) as response:
-        response.raise_for_status()
-        result = base64.b64encode(response.content).decode("utf-8")
-
-    return {"url": f"data:image/jpeg;base64,{result}"}
-
-
-headers = {"accept": "application/json", "Content-Type": "application/json"}
-
-queries = "slm markdown"
-documents = {
-    "content": [
+query = "A woman playing with her dog on a beach at sunset."
+document = (
+    "A woman shares a joyful moment with her golden retriever on a sun-drenched beach at sunset, "
+    "as the dog offers its paw in a heartwarming display of companionship and trust."
+)
+image_url = "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg"
+documents = [
     {
-            "type": "image_url",
-            "image_url": {
-                "url": "https://raw.githubusercontent.com/jina-ai/multimodal-reranker-test/main/handelsblatt-preview.png"
-            },
+        "type": "text",
+        "text": document,
     },
     {
         "type": "image_url",
-            "image_url": {
-                "url": "https://raw.githubusercontent.com/jina-ai/multimodal-reranker-test/main/paper-11.png"
-            },
+        "image_url": {"url": image_url},
     },
     {
         "type": "image_url",
-            "image_url": encode_base64_content_from_url(
-                "https://raw.githubusercontent.com/jina-ai/multimodal-reranker-test/main/paper-11.png"
-            ),
+        "image_url": {"url": encode_image_url(fetch_image(image_url))},
     },
-    ]
-}
+]
 
 
 def parse_args():
@@ -73,15 +57,40 @@ def main(args):
     models_url = base_url + "/v1/models"
     score_url = base_url + "/score"
 
-    response = requests.get(models_url, headers=headers)
+    response = requests.get(models_url)
     model = response.json()["data"][0]["id"]
 
-    prompt = {"model": model, "queries": queries, "documents": documents}
-    response = requests.post(score_url, headers=headers, json=prompt)
-    print("\nPrompt when queries is string and documents is a image list:")
-    pprint.pprint(prompt)
-    print("\nScore Response:")
-    print(json.dumps(response.json(), indent=2))
+    print("Query: string & Document: string")
+    prompt = {"model": model, "queries": query, "documents": document}
+    response = requests.post(score_url, json=prompt)
+    pprint.pprint(response.json())
+
+    print("Query: string & Document: text")
+    prompt = {
+        "model": model,
+        "queries": query,
+        "documents": {"content": [documents[0]]},
+    }
+    response = requests.post(score_url, json=prompt)
+    pprint.pprint(response.json())
+
+    print("Query: string & Document: image url")
+    prompt = {
+        "model": model,
+        "queries": query,
+        "documents": {"content": [documents[1]]},
+    }
+    response = requests.post(score_url, json=prompt)
+    pprint.pprint(response.json())
+
+    print("Query: string & Document: image base64")
+    prompt = {
+        "model": model,
+        "queries": query,
+        "documents": {"content": [documents[2]]},
+    }
+    response = requests.post(score_url, json=prompt)
+    pprint.pprint(response.json())
 
 
 if __name__ == "__main__":

tests/entrypoints/pooling/classify/test_online_vision.py

@@ -5,9 +5,9 @@ import json
 import pytest
 import requests
 
-from tests.entrypoints.test_utils import encode_base64_content_from_url
 from tests.utils import RemoteOpenAIServer
 from vllm.entrypoints.pooling.classify.protocol import ClassificationResponse
+from vllm.multimodal.utils import encode_image_url, fetch_image
 
 MODEL_NAME = "muziyongshixin/Qwen2.5-VL-7B-for-VideoCls"
 MAXIMUM_VIDEOS = 1
@@ -19,7 +19,7 @@ HF_OVERRIDES = {
 }
 input_text = "This product was excellent and exceeded my expectations"
 image_url = "https://vllm-public-assets.s3.us-west-2.amazonaws.com/multimodal_asset/cat_snow.jpg"
-image_base64 = encode_base64_content_from_url(image_url)
+image_base64 = {"url": encode_image_url(fetch_image(image_url))}
 video_url = "https://www.bogotobogo.com/python/OpenCV_Python/images/mean_shift_tracking/slow_traffic_small.mp4"
 
 

tests/entrypoints/pooling/score/test_online_score_vision.py

+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import pytest
+import requests
+
+from tests.utils import VLLM_PATH, RemoteOpenAIServer
+from vllm.entrypoints.pooling.score.protocol import ScoreResponse
+from vllm.multimodal.utils import encode_image_url, fetch_image
+
+MODEL_NAME = "Qwen/Qwen3-VL-Reranker-2B"
+HF_OVERRIDES = {
+    "architectures": ["Qwen3VLForSequenceClassification"],
+    "classifier_from_token": ["no", "yes"],
+    "is_original_qwen3_reranker": True,
+}
+
+query = "A cat standing in the snow."
+image_url = "https://vllm-public-assets.s3.us-west-2.amazonaws.com/multimodal_asset/cat_snow.jpg"
+documents = [
+    {
+        "type": "text",
+        "text": query,
+    },
+    {
+        "type": "image_url",
+        "image_url": {"url": image_url},
+    },
+    {
+        "type": "image_url",
+        "image_url": {"url": encode_image_url(fetch_image(image_url))},
+    },
+]
+
+
+@pytest.fixture(scope="module")
+def server():
+    args = [
+        "--enforce-eager",
+        "--max-model-len",
+        "8192",
+        "--chat-template",
+        str(VLLM_PATH / "examples/pooling/score/template/qwen3_vl_reranker.jinja"),
+    ]
+
+    with RemoteOpenAIServer(
+        MODEL_NAME, args, override_hf_configs=HF_OVERRIDES
+    ) as remote_server:
+        yield remote_server
+
+
+def test_score_api_queries_str_documents_str(server: RemoteOpenAIServer):
+    queries = "What is the capital of France?"
+    documents = "The capital of France is Paris."
+
+    score_response = requests.post(
+        server.url_for("score"),
+        json={
+            "model": MODEL_NAME,
+            "queries": queries,
+            "documents": documents,
+        },
+    )
+    score_response.raise_for_status()
+    score = ScoreResponse.model_validate(score_response.json())
+
+    assert score.id is not None
+    assert score.data is not None
+    assert len(score.data) == 1
+
+
+def test_score_api_queries_str_documents_text_content(server: RemoteOpenAIServer):
+    score_response = requests.post(
+        server.url_for("score"),
+        json={
+            "model": MODEL_NAME,
+            "queries": query,
+            "documents": {"content": [documents[0]]},
+        },
+    )
+    score_response.raise_for_status()
+    score = ScoreResponse.model_validate(score_response.json())
+
+    assert score.id is not None
+    assert score.data is not None
+    assert len(score.data) == 1
+
+
+def test_score_api_queries_str_documents_image_url_content(server: RemoteOpenAIServer):
+    score_response = requests.post(
+        server.url_for("score"),
+        json={
+            "model": MODEL_NAME,
+            "queries": query,
+            "documents": {"content": [documents[1]]},
+        },
+    )
+    score_response.raise_for_status()
+    score = ScoreResponse.model_validate(score_response.json())
+
+    assert score.id is not None
+    assert score.data is not None
+    assert len(score.data) == 1
+
+
+def test_score_api_queries_str_documents_image_base64_content(
+    server: RemoteOpenAIServer,
+):
+    score_response = requests.post(
+        server.url_for("score"),
+        json={
+            "model": MODEL_NAME,
+            "queries": query,
+            "documents": {"content": [documents[2]]},
+        },
+    )
+    score_response.raise_for_status()
+    score = ScoreResponse.model_validate(score_response.json())
+
+    assert score.id is not None
+    assert score.data is not None
+    assert len(score.data) == 1

tests/entrypoints/test_utils.py

 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-import base64
-
-import requests
-
 from vllm.entrypoints.utils import sanitize_message
 
 
@@ -12,11 +8,3 @@ def test_sanitize_message():
         sanitize_message("<_io.BytesIO object at 0x7a95e299e750>")
         == "<_io.BytesIO object>"
     )
-
-
-def encode_base64_content_from_url(content_url: str) -> dict[str, str]:
-    with requests.get(content_url) as response:
-        response.raise_for_status()
-        result = base64.b64encode(response.content).decode("utf-8")
-
-    return {"url": f"data:image/jpeg;base64,{result}"}

vllm/model_executor/models/adapters.py

@@ -466,6 +466,7 @@ def load_weights_using_from_2_way_softmax(
 
     language_model = _get_language_model_for_seq_cls(model)
     is_vlm = language_model is not model
+    using_vlm_head = is_vlm and hasattr(language_model, "score")
 
     language_model.lm_head = ParallelLMHead(
         text_config.vocab_size, text_config.hidden_size, quant_config=quant_config
@@ -506,14 +507,16 @@ def load_weights_using_from_2_way_softmax(
         torch.float32
     ) - lm_head_weight.data[[false_id]].to(torch.float32)
 
-    score_layer = language_model.score if is_vlm else model.score
+    score_layer = language_model.score if using_vlm_head else model.score
     param = score_layer.weight
     weight_loader = getattr(param, "weight_loader", default_weight_loader)
     weight_loader(param, score_weight)
 
     del language_model.lm_head
 
-    score_weight_name = "language_model.score.weight" if is_vlm else "score.weight"
+    score_weight_name = (
+        "language_model.score.weight" if using_vlm_head else "score.weight"
+    )
     loaded_weights.add(score_weight_name)
 
     lm_head_name = "lm_head.weight"
@@ -537,6 +540,7 @@ def load_weights_no_post_processing(model, weights: Iterable[tuple[str, torch.Te
 
     language_model = _get_language_model_for_seq_cls(model)
     is_vlm = language_model is not model
+    using_vlm_head = is_vlm and hasattr(language_model, "score")
 
     language_model.lm_head = ParallelLMHead(
         text_config.vocab_size, text_config.hidden_size, quant_config=quant_config
@@ -572,14 +576,16 @@ def load_weights_no_post_processing(model, weights: Iterable[tuple[str, torch.Te
     token_ids = [tokenizer.convert_tokens_to_ids(t) for t in tokens]
     score_weight = language_model.lm_head.weight.data[token_ids]
 
-    score_layer = language_model.score if is_vlm else model.score
+    score_layer = language_model.score if using_vlm_head else model.score
     param = score_layer.weight
     weight_loader = getattr(param, "weight_loader", default_weight_loader)
     weight_loader(param, score_weight)
 
     del language_model.lm_head
 
-    score_weight_name = "language_model.score.weight" if is_vlm else "score.weight"
+    score_weight_name = (
+        "language_model.score.weight" if using_vlm_head else "score.weight"
+    )
     loaded_weights.add(score_weight_name)
 
     lm_head_name = "lm_head.weight"