[FEATURE] Enables offline /score for embedding models

tests/models/embedding/language/test_scoring.py

@@ -5,12 +5,18 @@
 import math
 
 import pytest
+import torch
+import torch.nn.functional as F
 
 MODELS = [
     "cross-encoder/ms-marco-MiniLM-L-6-v2",  # Bert
     "BAAI/bge-reranker-v2-m3",  # Roberta
 ]
 
+EMBEDDING_MODELS = [
+    "sentence-transformers/all-MiniLM-L12-v2",
+]
+
 TEXTS_1 = [
     "What is the capital of France?",
     "What is the capital of Germany?",
@@ -87,3 +93,97 @@ def test_llm_N_to_N(vllm_runner, hf_runner, model_name, dtype: str):
 
     assert math.isclose(hf_outputs[0], vllm_outputs[0], rel_tol=0.01)
     assert math.isclose(hf_outputs[1], vllm_outputs[1], rel_tol=0.01)
+
+
+@pytest.fixture(scope="module", params=EMBEDDING_MODELS)
+def emb_model_name(request):
+    yield request.param
+
+
+@pytest.mark.parametrize("dtype", ["half"])
+def test_llm_1_to_1_embedding(vllm_runner, hf_runner, emb_model_name,
+                              dtype: str):
+
+    text_pair = [TEXTS_1[0], TEXTS_2[0]]
+
+    with hf_runner(emb_model_name, dtype=dtype,
+                   is_sentence_transformer=True) as hf_model:
+        hf_embeddings = hf_model.encode(text_pair)
+        hf_outputs = [
+            F.cosine_similarity(*map(torch.tensor, hf_embeddings), dim=0)
+        ]
+
+    with vllm_runner(emb_model_name,
+                     task="embed",
+                     dtype=dtype,
+                     max_model_len=None) as vllm_model:
+        vllm_outputs = vllm_model.score(text_pair[0], text_pair[1])
+
+    assert len(vllm_outputs) == 1
+    assert len(hf_outputs) == 1
+
+    assert math.isclose(hf_outputs[0], vllm_outputs[0], rel_tol=0.01)
+
+
+@pytest.mark.parametrize("dtype", ["half"])
+def test_llm_1_to_N_embedding(vllm_runner, hf_runner, emb_model_name,
+                              dtype: str):
+
+    text_pairs = [
+        [TEXTS_1[0], TEXTS_2[0]],
+        [TEXTS_1[0], TEXTS_2[1]],
+    ]
+
+    with hf_runner(emb_model_name, dtype=dtype,
+                   is_sentence_transformer=True) as hf_model:
+        hf_embeddings = [
+            hf_model.encode(text_pair) for text_pair in text_pairs
+        ]
+        hf_outputs = [
+            F.cosine_similarity(*map(torch.tensor, pair), dim=0)
+            for pair in hf_embeddings
+        ]
+
+    with vllm_runner(emb_model_name,
+                     task="embed",
+                     dtype=dtype,
+                     max_model_len=None) as vllm_model:
+        vllm_outputs = vllm_model.score(TEXTS_1[0], TEXTS_2)
+
+    assert len(vllm_outputs) == 2
+    assert len(hf_outputs) == 2
+
+    assert math.isclose(hf_outputs[0], vllm_outputs[0], rel_tol=0.01)
+    assert math.isclose(hf_outputs[1], vllm_outputs[1], rel_tol=0.01)
+
+
+@pytest.mark.parametrize("dtype", ["half"])
+def test_llm_N_to_N_embedding(vllm_runner, hf_runner, emb_model_name,
+                              dtype: str):
+
+    text_pairs = [
+        [TEXTS_1[0], TEXTS_2[0]],
+        [TEXTS_1[1], TEXTS_2[1]],
+    ]
+
+    with hf_runner(emb_model_name, dtype=dtype,
+                   is_sentence_transformer=True) as hf_model:
+        hf_embeddings = [
+            hf_model.encode(text_pair) for text_pair in text_pairs
+        ]
+        hf_outputs = [
+            F.cosine_similarity(*map(torch.tensor, pair), dim=0)
+            for pair in hf_embeddings
+        ]
+
+    with vllm_runner(emb_model_name,
+                     task="embed",
+                     dtype=dtype,
+                     max_model_len=None) as vllm_model:
+        vllm_outputs = vllm_model.score(TEXTS_1, TEXTS_2)
+
+    assert len(vllm_outputs) == 2
+    assert len(hf_outputs) == 2
+
+    assert math.isclose(hf_outputs[0], vllm_outputs[0], rel_tol=0.01)
+    assert math.isclose(hf_outputs[1], vllm_outputs[1], rel_tol=0.01)

vllm/entrypoints/llm.py

@@ -5,6 +5,7 @@
                     Tuple, Type, Union, cast, overload)
 
 import cloudpickle
+import torch
 import torch.nn as nn
 from tqdm import tqdm
 from typing_extensions import TypeVar, deprecated
@@ -996,6 +997,93 @@ def classify(
 
         return [ClassificationRequestOutput.from_base(item) for item in items]
 
+    def _embedding_score(self, tokenizer, truncate_prompt_tokens,
+                         text_1: List[str | TextPrompt | TokensPrompt],
+                         text_2: List[str | TextPrompt | TokensPrompt],
+                         use_tqdm, lora_request,
+                         prompt_adapter_request) -> List[ScoringRequestOutput]:
+
+        encoded_output = self.encode(text_1 + text_2)
+        encoded_output_1 = encoded_output[0:len(text_1)]
+        encoded_output_2 = encoded_output[len(text_1):]
+
+        if len(encoded_output_1) == 1:
+            encoded_output_1 = encoded_output_1 * len(encoded_output_2)
+
+        output_pairs = [(t1, t2)
+                        for t1, t2 in zip(encoded_output_1, encoded_output_2)]
+
+        scores = []
+        scorer = torch.nn.CosineSimilarity(0)
+
+        for embed_1, embed_2 in output_pairs:
+            pair_score = scorer(embed_1.outputs.data, embed_2.outputs.data)
+
+            if getattr(tokenizer, "pad_token", None) is None:
+                tokens = embed_1.prompt_token_ids + embed_2.prompt_token_ids
+            else:
+                tokens = embed_1.prompt_token_ids + [
+                    tokenizer.pad_token_type_id
+                ] + embed_2.prompt_token_ids
+
+            scores.append(
+                PoolingRequestOutput(
+                    request_id=f"{embed_1.request_id}_{embed_2.request_id}",
+                    outputs=pair_score,
+                    prompt_token_ids=tokens,
+                    finished=True))
+
+        items = self.engine_class.validate_outputs(scores,
+                                                   PoolingRequestOutput)
+        return [ScoringRequestOutput.from_base(item) for item in items]
+
+    def _cross_encoding_score(
+            self, tokenizer, truncate_prompt_tokens,
+            text_1: List[str | TextPrompt | TokensPrompt],
+            text_2: List[str | TextPrompt | TokensPrompt], use_tqdm,
+            lora_request,
+            prompt_adapter_request) -> List[ScoringRequestOutput]:
+
+        if isinstance(tokenizer, MistralTokenizer):
+            raise ValueError(
+                "MistralTokenizer not supported for cross-encoding")
+
+        if len(text_1) == 1:
+            text_1 = text_1 * len(text_2)
+
+        input_pairs = [(t1, t2) for t1, t2 in zip(text_1, text_2)]
+
+        pooling_params = PoolingParams()
+
+        tokenization_kwargs: Dict[str, Any] = {}
+        if truncate_prompt_tokens is not None:
+            tokenization_kwargs["truncation"] = True
+            tokenization_kwargs["max_length"] = truncate_prompt_tokens
+
+        parsed_prompts = []
+
+        for q, t in input_pairs:
+            prompt_inputs = tokenizer(text=q,
+                                      text_pair=t,
+                                      **tokenization_kwargs)
+            engine_prompt = TokensPrompt(
+                prompt_token_ids=prompt_inputs["input_ids"],
+                token_type_ids=prompt_inputs.get("token_type_ids"))
+            parsed_prompts.append(engine_prompt)
+
+        self._validate_and_add_requests(
+            prompts=parsed_prompts,
+            params=pooling_params,
+            lora_request=lora_request,
+            prompt_adapter_request=prompt_adapter_request,
+        )
+
+        outputs = self._run_engine(use_tqdm=use_tqdm)
+        items = self.engine_class.validate_outputs(outputs,
+                                                   PoolingRequestOutput)
+
+        return [ScoringRequestOutput.from_base(item) for item in items]
+
     def score(
         self,
         text_1: Union[SingletonPrompt, Sequence[SingletonPrompt]],
@@ -1032,6 +1120,7 @@ def score(
             A list of ``ScoringRequestOutput`` objects containing the
             generated scores in the same order as the input prompts.
         """
+
         runner_type = self.llm_engine.model_config.runner_type
         if runner_type != "pooling":
             messages = ["LLM.score() is only supported for pooling models."]
@@ -1047,25 +1136,20 @@ def score(
 
             raise ValueError(" ".join(messages))
 
-        if not self.llm_engine.model_config.is_cross_encoder:
-            raise ValueError("Your model does not support cross encoding")
-        if self.llm_engine.model_config.task != "score":
-            raise ValueError("Score API is only enabled for `--task score`")
-
-        tokenizer = self.llm_engine.get_tokenizer()
-
-        if isinstance(tokenizer, MistralTokenizer):
+        if self.llm_engine.model_config.task not in ("embed", "score"):
             raise ValueError(
-                "MistralTokenizer not supported for cross-encoding")
+                "Score API is only enabled for `--task embed or score`")
 
         # the tokenizer for models such as
         # "cross-encoder/ms-marco-MiniLM-L-6-v2" doesn't support passing
         # lists of tokens to the `text` and `text_pair` kwargs
+        tokenizer = self.llm_engine.get_tokenizer()
+
         def ensure_str(prompt: SingletonPrompt):
             if isinstance(prompt, dict):
                 if "multi_modal_data" in prompt:
                     raise ValueError("Multi-modal prompt is not "
-                                     "supported for cross encoding")
+                                     "supported for scoring")
                 elif "prompt_token_ids" in prompt:
                     prompt = tokenizer.decode(
                         cast(TokensPrompt, prompt)["prompt_token_ids"])
@@ -1091,40 +1175,15 @@ def ensure_str(prompt: SingletonPrompt):
         if len(text_2) == 0:
             raise ValueError("At least one text_pair element must be given")
 
-        if len(text_1) == 1:
-            text_1 = text_1 * len(text_2)
-
-        input_pairs = [(t1, t2) for t1, t2 in zip(text_1, text_2)]
-        pooling_params = PoolingParams()
-
-        tokenization_kwargs: Dict[str, Any] = {}
-        if truncate_prompt_tokens is not None:
-            tokenization_kwargs["truncation"] = True
-            tokenization_kwargs["max_length"] = truncate_prompt_tokens
-
-        parsed_prompts = []
-
-        for q, t in input_pairs:
-            prompt_inputs = tokenizer(text=q,
-                                      text_pair=t,
-                                      **tokenization_kwargs)
-            engine_prompt = TokensPrompt(
-                prompt_token_ids=prompt_inputs["input_ids"],
-                token_type_ids=prompt_inputs.get("token_type_ids"))
-            parsed_prompts.append(engine_prompt)
-
-        self._validate_and_add_requests(
-            prompts=parsed_prompts,
-            params=pooling_params,
-            lora_request=lora_request,
-            prompt_adapter_request=prompt_adapter_request,
-        )
-
-        outputs = self._run_engine(use_tqdm=use_tqdm)
-        items = self.engine_class.validate_outputs(outputs,
-                                                   PoolingRequestOutput)
-
-        return [ScoringRequestOutput.from_base(item) for item in items]
+        if self.llm_engine.model_config.is_cross_encoder:
+            return self._cross_encoding_score(tokenizer,
+                                              truncate_prompt_tokens, text_1,
+                                              text_2, use_tqdm, lora_request,
+                                              prompt_adapter_request)
+        else:
+            return self._embedding_score(tokenizer, truncate_prompt_tokens,
+                                         text_1, text_2, use_tqdm,
+                                         lora_request, prompt_adapter_request)
 
     def start_profile(self) -> None:
         self.llm_engine.start_profile()