[Feature][Frontend]: Continued `stream_options` implementation also in CompletionRequest (#5319)

tests/entrypoints/test_openai_server.py

@@ -478,8 +478,6 @@ async def test_completion_streaming(server, client: openai.AsyncOpenAI,
         temperature=0.0,
     )
     single_output = single_completion.choices[0].text
-    single_usage = single_completion.usage
-
     stream = await client.completions.create(model=model_name,
                                              prompt=prompt,
                                              max_tokens=5,
@@ -495,7 +493,6 @@ async def test_completion_streaming(server, client: openai.AsyncOpenAI,
     assert finish_reason_count == 1
     assert chunk.choices[0].finish_reason == "length"
     assert chunk.choices[0].text
-    assert chunk.usage == single_usage
     assert "".join(chunks) == single_output
 
 
@@ -550,6 +547,138 @@ async def test_chat_streaming(server, client: openai.AsyncOpenAI,
     assert "".join(chunks) == output
 
 
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "model_name",
+    ["HuggingFaceH4/zephyr-7b-beta", "zephyr-lora"],
+)
+async def test_chat_completion_stream_options(server,
+                                              client: openai.AsyncOpenAI,
+                                              model_name: str):
+    messages = [{
+        "role": "system",
+        "content": "You are a helpful assistant."
+    }, {
+        "role": "user",
+        "content": "What is the capital of France?"
+    }]
+
+    # Test stream=True, stream_options={"include_usage": False}
+    stream = await client.chat.completions.create(
+        model=model_name,
+        messages=messages,
+        max_tokens=10,
+        temperature=0.0,
+        stream=True,
+        stream_options={"include_usage": False})
+    async for chunk in stream:
+        assert chunk.usage is None
+
+    # Test stream=True, stream_options={"include_usage": True}
+    stream = await client.chat.completions.create(
+        model=model_name,
+        messages=messages,
+        max_tokens=10,
+        temperature=0.0,
+        stream=True,
+        stream_options={"include_usage": True})
+
+    async for chunk in stream:
+        if chunk.choices[0].finish_reason is None:
+            assert chunk.usage is None
+        else:
+            assert chunk.usage is None
+            final_chunk = await stream.__anext__()
+            assert final_chunk.usage is not None
+            assert final_chunk.usage.prompt_tokens > 0
+            assert final_chunk.usage.completion_tokens > 0
+            assert final_chunk.usage.total_tokens == (
+                final_chunk.usage.prompt_tokens +
+                final_chunk.usage.completion_tokens)
+            assert final_chunk.choices == []
+
+    # Test stream=False, stream_options={"include_usage": None}
+    with pytest.raises(BadRequestError):
+        await client.chat.completions.create(
+            model=model_name,
+            messages=messages,
+            max_tokens=10,
+            temperature=0.0,
+            stream=False,
+            stream_options={"include_usage": None})
+
+    # Test stream=False, stream_options={"include_usage": True}
+    with pytest.raises(BadRequestError):
+        await client.chat.completions.create(
+            model=model_name,
+            messages=messages,
+            max_tokens=10,
+            temperature=0.0,
+            stream=False,
+            stream_options={"include_usage": True})
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "model_name",
+    ["HuggingFaceH4/zephyr-7b-beta", "zephyr-lora"],
+)
+async def test_completion_stream_options(server, client: openai.AsyncOpenAI,
+                                         model_name: str):
+    prompt = "What is the capital of France?"
+
+    # Test stream=True, stream_options={"include_usage": False}
+    stream = await client.completions.create(
+        model=model_name,
+        prompt=prompt,
+        max_tokens=5,
+        temperature=0.0,
+        stream=True,
+        stream_options={"include_usage": False})
+    async for chunk in stream:
+        assert chunk.usage is None
+
+    # Test stream=True, stream_options={"include_usage": True}
+    stream = await client.completions.create(
+        model=model_name,
+        prompt=prompt,
+        max_tokens=5,
+        temperature=0.0,
+        stream=True,
+        stream_options={"include_usage": True})
+    async for chunk in stream:
+        if chunk.choices[0].finish_reason is None:
+            assert chunk.usage is None
+        else:
+            assert chunk.usage is None
+            final_chunk = await stream.__anext__()
+            assert final_chunk.usage is not None
+            assert final_chunk.usage.prompt_tokens > 0
+            assert final_chunk.usage.completion_tokens > 0
+            assert final_chunk.usage.total_tokens == (
+                final_chunk.usage.prompt_tokens +
+                final_chunk.usage.completion_tokens)
+            assert final_chunk.choices == []
+
+    # Test stream=False, stream_options={"include_usage": None}
+    with pytest.raises(BadRequestError):
+        await client.completions.create(model=model_name,
+                                        prompt=prompt,
+                                        max_tokens=5,
+                                        temperature=0.0,
+                                        stream=False,
+                                        stream_options={"include_usage": None})
+
+    # Test stream=False, stream_options={"include_usage": True}
+    with pytest.raises(BadRequestError):
+        await client.completions.create(model=model_name,
+                                        prompt=prompt,
+                                        max_tokens=5,
+                                        temperature=0.0,
+                                        stream=False,
+                                        stream_options={"include_usage": True})
+
+
 @pytest.mark.asyncio
 @pytest.mark.parametrize(
     # just test 1 lora hereafter
@@ -1343,106 +1472,5 @@ async def test_batch_embedding(embedding_server, client: openai.AsyncOpenAI,
     assert embeddings.usage.total_tokens == 17
 
 
-@pytest.mark.parametrize(
-    "model_name",
-    [MODEL_NAME],
-)
-async def test_stream_options(server, client: openai.AsyncOpenAI,
-                              model_name: str):
-    prompt = "What is the capital of France?"
-
-    # Test stream=True, stream_options=None
-    stream = await client.completions.create(
-        model=model_name,
-        prompt=prompt,
-        max_tokens=5,
-        temperature=0.0,
-        stream=True,
-        stream_options=None,
-    )
-    chunks = []
-    async for chunk in stream:
-        chunks.append(chunk.choices[0].text)
-    assert len(chunks) > 0
-    assert "usage" not in chunk
-
-    # Test stream=True, stream_options={"include_usage": False}
-    stream = await client.completions.create(
-        model=model_name,
-        prompt=prompt,
-        max_tokens=5,
-        temperature=0.0,
-        stream=True,
-        stream_options={"include_usage": False},
-    )
-    chunks = []
-    async for chunk in stream:
-        chunks.append(chunk.choices[0].text)
-    assert len(chunks) > 0
-    assert "usage" not in chunk
-
-    # Test stream=True, stream_options={"include_usage": True}
-    stream = await client.completions.create(
-        model=model_name,
-        prompt=prompt,
-        max_tokens=5,
-        temperature=0.0,
-        stream=True,
-        stream_options={"include_usage": True},
-    )
-    chunks = []
-    finish_reason_count = 0
-    async for chunk in stream:
-        if chunk.choices[0].finish_reason is None:
-            assert chunk.usage is None
-            chunks.append(chunk.choices[0].text)
-        else:
-            assert chunk.usage is None
-            finish_reason_count += 1
-
-    # The last message should have usage and no choices
-    last_message = await stream.__anext__()
-    assert last_message.usage is not None
-    assert last_message.usage.prompt_tokens > 0
-    assert last_message.usage.completion_tokens > 0
-    assert last_message.usage.total_tokens == (
-        last_message.usage.prompt_tokens +
-        last_message.usage.completion_tokens)
-    assert last_message.choices == []
-
-    # Test stream=False, stream_options={"include_usage": None}
-    with pytest.raises(BadRequestError):
-        await client.completions.create(
-            model=model_name,
-            prompt=prompt,
-            max_tokens=5,
-            temperature=0.0,
-            stream=False,
-            stream_options={"include_usage": None},
-        )
-
-    # Test stream=False, stream_options={"include_usage": False}
-    with pytest.raises(BadRequestError):
-        await client.completions.create(
-            model=model_name,
-            prompt=prompt,
-            max_tokens=5,
-            temperature=0.0,
-            stream=False,
-            stream_options={"include_usage": False},
-        )
-
-    # Test stream=False, stream_options={"include_usage": True}
-    with pytest.raises(BadRequestError):
-        await client.completions.create(
-            model=model_name,
-            prompt=prompt,
-            max_tokens=5,
-            temperature=0.0,
-            stream=False,
-            stream_options={"include_usage": True},
-        )
-
-
 if __name__ == "__main__":
     pytest.main([__file__])

vllm/entrypoints/openai/protocol.py

@@ -346,6 +346,7 @@ class CompletionRequest(OpenAIBaseModel):
                                 le=torch.iinfo(torch.long).max)
     stop: Optional[Union[str, List[str]]] = Field(default_factory=list)
     stream: Optional[bool] = False
+    stream_options: Optional[StreamOptions] = None
     suffix: Optional[str] = None
     temperature: Optional[float] = 1.0
     top_p: Optional[float] = 1.0
@@ -482,6 +483,14 @@ class CompletionRequest(OpenAIBaseModel):
                               " in the interval [0, 5]."))
         return data
 
+    @model_validator(mode="before")
+    @classmethod
+    def validate_stream_options(cls, data):
+        if data.get("stream_options") and not data.get("stream"):
+            raise ValueError(
+                "Stream options can only be defined when stream is True.")
+        return data
+
 
 class EmbeddingRequest(BaseModel):
     # Ordered by official OpenAI API documentation

vllm/entrypoints/openai/serving_chat.py

@@ -441,25 +441,24 @@ class OpenAIServingChat(OpenAIServing):
                         yield f"data: {data}\n\n"
                         finish_reason_sent[i] = True
 
-                    if (request.stream_options
-                            and request.stream_options.include_usage):
-                        final_usage = UsageInfo(
-                            prompt_tokens=prompt_tokens,
-                            completion_tokens=previous_num_tokens[i],
-                            total_tokens=prompt_tokens +
-                            previous_num_tokens[i],
-                        )
+            if (request.stream_options
+                    and request.stream_options.include_usage):
+                final_usage = UsageInfo(
+                    prompt_tokens=prompt_tokens,
+                    completion_tokens=previous_num_tokens[i],
+                    total_tokens=prompt_tokens + previous_num_tokens[i],
+                )
 
-                        final_usage_chunk = ChatCompletionStreamResponse(
-                            id=request_id,
-                            object=chunk_object_type,
-                            created=created_time,
-                            choices=[],
-                            model=model_name,
-                            usage=final_usage)
-                        final_usage_data = (final_usage_chunk.model_dump_json(
-                            exclude_unset=True, exclude_none=True))
-                        yield f"data: {final_usage_data}\n\n"
+                final_usage_chunk = ChatCompletionStreamResponse(
+                    id=request_id,
+                    object=chunk_object_type,
+                    created=created_time,
+                    choices=[],
+                    model=model_name,
+                    usage=final_usage)
+                final_usage_data = (final_usage_chunk.model_dump_json(
+                    exclude_unset=True, exclude_none=True))
+                yield f"data: {final_usage_data}\n\n"
 
         except ValueError as e:
             # TODO: Use a vllm-specific Validation Error

vllm/entrypoints/openai/serving_completion.py

@@ -264,7 +264,8 @@ class OpenAIServingCompletion(OpenAIServing):
                         )
                     else:
                         final_usage = None
-                    response_json = CompletionStreamResponse(
+
+                    chunk = CompletionStreamResponse(
                         id=request_id,
                         created=created_time,
                         model=model_name,
@@ -276,10 +277,27 @@ class OpenAIServingCompletion(OpenAIServing):
                                 finish_reason=finish_reason,
                                 stop_reason=stop_reason,
                             )
-                        ],
-                        usage=final_usage,
-                    ).model_dump_json(exclude_unset=True)
+                        ])
+                    if (request.stream_options
+                            and request.stream_options.include_usage):
+                        chunk.usage = None
+
+                    response_json = chunk.model_dump_json(exclude_unset=True)
                     yield f"data: {response_json}\n\n"
+
+            if (request.stream_options
+                    and request.stream_options.include_usage):
+                final_usage_chunk = CompletionStreamResponse(
+                    id=request_id,
+                    created=created_time,
+                    model=model_name,
+                    choices=[],
+                    usage=final_usage,
+                )
+                final_usage_data = (final_usage_chunk.model_dump_json(
+                    exclude_unset=True, exclude_none=True))
+                yield f"data: {final_usage_data}\n\n"
+
         except ValueError as e:
             # TODO: Use a vllm-specific Validation Error
             data = self.create_streaming_error_response(str(e))