Adds truncate_prompt_tokens param for embeddings creation

Signed-off-by: Flavia Beo <[email protected]>

tests/entrypoints/openai/test_embedding.py

@@ -144,3 +144,44 @@ async def test_batch_base64_embedding(embedding_client: openai.AsyncOpenAI,
         0].embedding
     assert responses_float.data[1].embedding == responses_default.data[
         1].embedding
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "model_name",
+    [EMBEDDING_MODEL_NAME],
+)
+async def test_single_embedding_truncation(
+        embedding_client: openai.AsyncOpenAI, model_name: str):
+    input_texts = [
+        "Como o Brasil pode fomentar o desenvolvimento de modelos de IA?",
+    ]
+
+    # test single embedding
+    embeddings = await embedding_client.embeddings.create(
+        model=model_name, input=input_texts, truncate_prompt_tokens=10)
+    assert embeddings.id is not None
+    assert len(embeddings.data) == 1
+    assert len(embeddings.data[0].embedding) == 4096
+    assert embeddings.usage.completion_tokens == 0
+    assert embeddings.usage.prompt_tokens == 10
+    assert embeddings.usage.total_tokens == 10
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "model_name",
+    [EMBEDDING_MODEL_NAME],
+)
+async def test_single_embedding_truncation_invalid(
+        embedding_client: openai.AsyncOpenAI, model_name: str):
+    input_texts = [
+        "Como o Brasil pode fomentar o desenvolvimento de modelos de IA?",
+    ]
+
+    # test single embedding
+    embeddings = await embedding_client.embeddings.create(
+        model=model_name, input=input_texts, truncate_prompt_tokens=8193)
+    assert "error" in embeddings.object
+    assert "truncate_prompt_tokens value is greater than max_model_len. "\
+            "Please, select a smaller truncation size." in embeddings.message

vllm/entrypoints/openai/protocol.py

@@ -614,6 +614,7 @@ class EmbeddingRequest(OpenAIBaseModel):
     encoding_format: Literal["float", "base64"] = "float"
     dimensions: Optional[int] = None
     user: Optional[str] = None
+    truncate_prompt_tokens: Optional[Annotated[int, Field(ge=1)]] = None
 
     # doc: begin-embedding-pooling-params
     additional_data: Optional[Any] = None

vllm/entrypoints/openai/serving_embedding.py

@@ -110,6 +110,17 @@ async def create_embedding(
         request_id = f"embd-{random_uuid()}"
         created_time = int(time.monotonic())
 
+        truncate_prompt_tokens = None
+
+        if request.truncate_prompt_tokens is not None:
+            if request.truncate_prompt_tokens <= self.max_model_len:
+                truncate_prompt_tokens = request.truncate_prompt_tokens
+            else:
+                return self.create_error_response(
+                    "truncate_prompt_tokens value is "
+                    "greater than max_model_len."
+                    " Please, select a smaller truncation size.")
+
         # Schedule the request and get the result generator.
         generators: List[AsyncGenerator[EmbeddingRequestOutput, None]] = []
         try:
@@ -124,11 +135,9 @@ async def create_embedding(
             pooling_params = request.to_pooling_params()
 
             prompts = list(
-                self._tokenize_prompt_input_or_inputs(
-                    request,
-                    tokenizer,
-                    request.input,
-                ))
+                self._tokenize_prompt_input_or_inputs(request, tokenizer,
+                                                      request.input,
+                                                      truncate_prompt_tokens))
 
             for i, prompt_inputs in enumerate(prompts):
                 request_id_item = f"{request_id}-{i}"