[Model][2/N] Improve all pooling task | Support multi-vector retrieval

examples/offline_inference/pooling/README.md

@@ -26,6 +26,12 @@ python examples/offline_inference/pooling/embed_jina_embeddings_v3.py
 python examples/offline_inference/pooling/embed_matryoshka_fy.py
 ```
 
+## Multi vector retrieval usage
+
+```bash
+python examples/offline_inference/pooling/multi_vector_retrieval.py
+```
+
 ## Named Entity Recognition (NER) usage
 
 ```bash

examples/offline_inference/pooling/multi_vector_retrieval.py

@@ -0,0 +1,56 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from argparse import Namespace
+
+from vllm import LLM, EngineArgs
+from vllm.utils import FlexibleArgumentParser
+
+
+def parse_args():
+    parser = FlexibleArgumentParser()
+    parser = EngineArgs.add_cli_args(parser)
+    # Set example specific arguments
+    parser.set_defaults(
+        model="BAAI/bge-m3",
+        runner="pooling",
+        enforce_eager=True,
+    )
+    return parser.parse_args()
+
+
+def main(args: Namespace):
+    # Sample prompts.
+    prompts = [
+        "Hello, my name is",
+        "The president of the United States is",
+        "The capital of France is",
+        "The future of AI is",
+    ]
+
+    # Create an LLM.
+    # You should pass runner="pooling" for embedding models
+    llm = LLM(**vars(args))
+
+    # Generate embedding. The output is a list of EmbeddingRequestOutputs.
+    outputs = llm.embed(prompts)
+
+    # Print the outputs.
+    print("\nGenerated Outputs:\n" + "-" * 60)
+    for prompt, output in zip(prompts, outputs):
+        embeds = output.outputs.embedding
+        print(len(embeds))
+
+    # Generate embedding for each token. The output is a list of PoolingRequestOutput.
+    outputs = llm.encode(prompts, pooling_task="token_embed")
+
+    # Print the outputs.
+    print("\nGenerated Outputs:\n" + "-" * 60)
+    for prompt, output in zip(prompts, outputs):
+        multi_vector = output.outputs.data
+        print(multi_vector.shape)
+
+
+if __name__ == "__main__":
+    args = parse_args()
+    main(args)

examples/offline_inference/prithvi_geospatial_mae_io_processor.py

@@ -40,7 +40,7 @@ def main():
         model_impl="terratorch",
     )
 
-    pooling_params = PoolingParams(task="encode", softmax=False)
+    pooling_params = PoolingParams(task="token_classify", activation=False)
     pooler_output = llm.encode(
         img_prompt,
         pooling_params=pooling_params,

examples/online_serving/pooling/README.md

@@ -18,6 +18,12 @@ python examples/online_serving/pooling/embedding_embed_dtype_client.py
 python examples/online_serving/pooling/jinaai_rerank_client.py
 ```
 
+## Multi vector retrieval usage
+
+```bash
+python examples/online_serving/pooling/multi_vector_retrieval_client.py
+```
+
 ## Named Entity Recognition (NER) usage
 
 ```bash

examples/online_serving/pooling/multi_vector_retrieval_client.py

@@ -0,0 +1,54 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+"""
+Example online usage of Pooling API for multi vector retrieval.
+
+Run `vllm serve <model> --runner pooling`
+to start up the server in vLLM. e.g.
+
+vllm serve BAAI/bge-m3
+"""
+
+import argparse
+
+import requests
+import torch
+
+
+def post_http_request(prompt: dict, api_url: str) -> requests.Response:
+    headers = {"User-Agent": "Test Client"}
+    response = requests.post(api_url, headers=headers, json=prompt)
+    return response
+
+
+def parse_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--host", type=str, default="localhost")
+    parser.add_argument("--port", type=int, default=8000)
+    parser.add_argument("--model", type=str, default="BAAI/bge-m3")
+
+    return parser.parse_args()
+
+
+def main(args):
+    api_url = f"http://{args.host}:{args.port}/pooling"
+    model_name = args.model
+
+    prompts = [
+        "Hello, my name is",
+        "The president of the United States is",
+        "The capital of France is",
+        "The future of AI is",
+    ]
+    prompt = {"model": model_name, "input": prompts}
+
+    pooling_response = post_http_request(prompt=prompt, api_url=api_url)
+    for output in pooling_response.json()["data"]:
+        multi_vector = torch.tensor(output["data"])
+        print(multi_vector.shape)
+
+
+if __name__ == "__main__":
+    args = parse_args()
+    main(args)

tests/conftest.py

@@ -1011,8 +1011,12 @@ def embed(
         req_outputs = self.llm.embed(inputs, *args, **kwargs)
         return [req_output.outputs.embedding for req_output in req_outputs]
 
-    def encode(self, prompts: list[str]) -> list[list[float]]:
-        req_outputs = self.llm.encode(prompts)
+    def token_embed(self, prompts: list[str]) -> list[list[float]]:
+        req_outputs = self.llm.encode(prompts, pooling_task="token_embed")
+        return [req_output.outputs.data for req_output in req_outputs]
+
+    def token_classify(self, prompts: list[str]) -> list[list[float]]:
+        req_outputs = self.llm.encode(prompts, pooling_task="token_classify")
         return [req_output.outputs.data for req_output in req_outputs]
 
     def reward(self, prompts: list[str]) -> list[list[float]]:

tests/entrypoints/pooling/llm/test_classify.py

@@ -63,7 +63,7 @@ def test_encode_api(llm: LLM):
     # chunked prefill does not support all pooling
     err_msg = "pooling_task must be one of.+"
     with pytest.raises(ValueError, match=err_msg):
-        llm.encode(prompts, use_tqdm=False)
+        llm.encode(prompts, pooling_task="token_classify", use_tqdm=False)
 
 
 def test_score_api(llm: LLM):

tests/entrypoints/pooling/llm/test_embedding.py

@@ -35,6 +35,13 @@ def llm():
     cleanup_dist_env_and_memory()
 
 
+@pytest.mark.skip_global_cleanup
+def test_encode_api(llm: LLM):
+    outputs = llm.encode(prompts, pooling_task="token_embed", use_tqdm=False)
+    multi_vector = outputs[0].outputs.data
+    assert multi_vector.shape == (11, 384)
+
+
 def test_pooling_params(llm: LLM):
     def get_outputs(normalize):
         outputs = llm.embed(

tests/entrypoints/pooling/llm/test_encode.py

@@ -57,20 +57,24 @@ def test_multiple_pooling_params(llm: LLM):
     ]
 
     # Multiple PoolingParams should be matched with each prompt
-    outputs = llm.encode(PROMPTS, pooling_params=pooling_params)
+    outputs = llm.encode(PROMPTS, pooling_params=pooling_params, pooling_task="embed")
     assert len(PROMPTS) == len(outputs)
 
     # Exception raised, if the size of params does not match the size of prompts
     with pytest.raises(ValueError):
-        outputs = llm.encode(PROMPTS, pooling_params=pooling_params[:3])
+        outputs = llm.encode(
+            PROMPTS, pooling_params=pooling_params[:3], pooling_task="embed"
+        )
 
     # Single PoolingParams should be applied to every prompt
     single_pooling_params = PoolingParams()
-    outputs = llm.encode(PROMPTS, pooling_params=single_pooling_params)
+    outputs = llm.encode(
+        PROMPTS, pooling_params=single_pooling_params, pooling_task="embed"
+    )
     assert len(PROMPTS) == len(outputs)
 
     # pooling_params is None, default params should be applied
-    outputs = llm.encode(PROMPTS, pooling_params=None)
+    outputs = llm.encode(PROMPTS, pooling_params=None, pooling_task="embed")
     assert len(PROMPTS) == len(outputs)
 
 

tests/entrypoints/pooling/llm/test_reward.py

@@ -36,22 +36,23 @@ def llm():
     cleanup_dist_env_and_memory()
 
 
-@pytest.mark.skip_global_cleanup
 def test_pooling_params(llm: LLM):
-    def get_outputs(softmax):
+    def get_outputs(activation):
         outputs = llm.reward(
-            prompts, pooling_params=PoolingParams(softmax=softmax), use_tqdm=False
+            prompts, pooling_params=PoolingParams(activation=activation), use_tqdm=False
         )
         return torch.cat([x.outputs.data for x in outputs])
 
-    default = get_outputs(softmax=None)
-    w_softmax = get_outputs(softmax=True)
-    wo_softmax = get_outputs(softmax=False)
+    default = get_outputs(activation=None)
+    w_activation = get_outputs(activation=True)
+    wo_activation = get_outputs(activation=False)
 
-    assert torch.allclose(default, w_softmax, atol=1e-2), "Default should use softmax."
-    assert not torch.allclose(w_softmax, wo_softmax, atol=1e-2), (
-        "wo_softmax should not use softmax."
+    assert torch.allclose(default, w_activation, atol=1e-2), (
+        "Default should use activation."
     )
-    assert torch.allclose(softmax(wo_softmax), w_softmax, atol=1e-2), (
-        "w_softmax should be close to softmax(wo_softmax)."
+    assert not torch.allclose(w_activation, wo_activation, atol=1e-2), (
+        "wo_activation should not use activation."
+    )
+    assert torch.allclose(softmax(wo_activation), w_activation, atol=1e-2), (
+        "w_activation should be close to activation(wo_activation)."
     )

tests/entrypoints/pooling/openai/test_embedding.py

@@ -17,6 +17,7 @@
 from vllm.entrypoints.openai.protocol import (
     EMBED_DTYPE_TO_TORCH_DTYPE,
     EmbeddingResponse,
+    PoolingResponse,
 )
 from vllm.transformers_utils.tokenizer import get_tokenizer
 
@@ -509,3 +510,20 @@ async def get_outputs(normalize):
     assert torch.allclose(w_normal, F.normalize(wo_normal, p=2, dim=-1), atol=1e-2), (
         "w_normal should be close to normal(wo_normal)."
     )
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+async def test_pooling(server: RemoteOpenAIServer, model_name: str):
+    input_text = ["The chef prepared a delicious meal."]
+
+    response = requests.post(
+        server.url_for("pooling"),
+        json={"model": model_name, "input": input_text, "encoding_format": "float"},
+    )
+
+    poolings = PoolingResponse.model_validate(response.json())
+
+    assert len(poolings.data) == 1
+    assert len(poolings.data[0].data) == 11
+    assert len(poolings.data[0].data[0]) == 384

tests/entrypoints/pooling/openai/test_rerank.py

@@ -7,7 +7,7 @@
 import torch.nn.functional as F
 
 from tests.utils import RemoteOpenAIServer
-from vllm.entrypoints.openai.protocol import RerankResponse
+from vllm.entrypoints.openai.protocol import PoolingResponse, RerankResponse
 
 MODEL_NAME = "BAAI/bge-reranker-base"
 DTYPE = "bfloat16"
@@ -159,3 +159,20 @@ async def get_outputs(activation):
     assert torch.allclose(F.sigmoid(wo_activation), w_activation, atol=1e-2), (
         "w_activation should be close to activation(wo_activation)."
     )
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+async def test_pooling(server: RemoteOpenAIServer, model_name: str):
+    input_text = ["The chef prepared a delicious meal."]
+
+    response = requests.post(
+        server.url_for("pooling"),
+        json={"model": model_name, "input": input_text, "encoding_format": "float"},
+    )
+
+    poolings = PoolingResponse.model_validate(response.json())
+
+    assert len(poolings.data) == 1
+    assert len(poolings.data[0].data) == 11
+    assert len(poolings.data[0].data[0]) == 1

tests/models/language/pooling/test_multi_vector_retrieval.py

@@ -0,0 +1,45 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import pytest
+import torch
+from transformers import AutoModel
+
+from tests.models.utils import check_embeddings_close
+
+
+@pytest.mark.parametrize(
+    "model",
+    ["BAAI/bge-m3"],
+)
+@pytest.mark.parametrize("dtype", ["half"])
+@torch.inference_mode
+def test_embed_models(hf_runner, vllm_runner, example_prompts, model: str, dtype: str):
+    with vllm_runner(
+        model,
+        runner="pooling",
+        max_model_len=None,
+    ) as vllm_model:
+        vllm_outputs = vllm_model.token_embed(example_prompts)
+
+    with hf_runner(
+        model,
+        auto_cls=AutoModel,
+    ) as hf_model:
+        tokenizer = hf_model.tokenizer
+        hf_outputs = []
+        for prompt in example_prompts:
+            inputs = tokenizer([prompt], return_tensors="pt")
+            inputs = hf_model.wrap_device(inputs)
+            output = hf_model.model(**inputs)
+            embedding = output.last_hidden_state[0].float()
+            # normal
+            hf_outputs.append(embedding.cpu())
+
+    for hf_output, vllm_output in zip(hf_outputs, vllm_outputs):
+        check_embeddings_close(
+            embeddings_0_lst=hf_output,
+            embeddings_1_lst=vllm_output,
+            name_0="hf",
+            name_1="vllm",
+            tol=1e-2,
+        )