llama-stack-mirror/tests/integration/inference/test_openai_completion.py

# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.


import time
import unicodedata

import pytest

from ..test_cases.test_case import TestCase


def _normalize_text(text: str) -> str:
    """
    Normalize Unicode text by removing diacritical marks for comparison.

    The test case streaming_01 expects the answer "Sol" for the question "What's the name of the Sun
    in latin?", but the model is returning "sōl" (with a macron over the 'o'), which is the correct
    Latin spelling. The test is failing because it's doing a simple case-insensitive string search
    for "sol" but the actual response contains the diacritical mark.
    """
    return unicodedata.normalize("NFD", text).encode("ascii", "ignore").decode("ascii").lower()


def provider_from_model(client_with_models, model_id):
    models = {m.identifier: m for m in client_with_models.models.list()}
    models.update({m.provider_resource_id: m for m in client_with_models.models.list()})
    provider_id = models[model_id].provider_id
    providers = {p.provider_id: p for p in client_with_models.providers.list()}
    return providers[provider_id]


def skip_if_model_doesnt_support_openai_completion(client_with_models, model_id):
    provider = provider_from_model(client_with_models, model_id)
    if provider.provider_type in (
        "inline::meta-reference",
        "inline::sentence-transformers",
        "inline::vllm",
        "remote::bedrock",
        "remote::cerebras",
        "remote::databricks",
        # Technically Nvidia does support OpenAI completions, but none of their hosted models
        # support both completions and chat completions endpoint and all the Llama models are
        # just chat completions
        "remote::nvidia",
        "remote::runpod",
        "remote::sambanova",
        "remote::vertexai",
        # {"error":{"message":"Unknown request URL: GET /openai/v1/completions. Please check the URL for typos,
        # or see the docs at https://console.groq.com/docs/","type":"invalid_request_error","code":"unknown_url"}}
        "remote::groq",
        "remote::gemini",  # https://generativelanguage.googleapis.com/v1beta/openai/completions -> 404
        "remote::anthropic",  # at least claude-3-{5,7}-{haiku,sonnet}-* / claude-{sonnet,opus}-4-* are not supported
        "remote::azure",  # {'error': {'code': 'OperationNotSupported', 'message': 'The completion operation
        #  does not work with the specified model, gpt-5-mini. Please choose different model and try
        #  again. You can learn more about which models can be used with each operation here:
        #  https://go.microsoft.com/fwlink/?linkid=2197993.'}}"}
        "remote::watsonx",  # return 404 when hitting the /openai/v1 endpoint
    ):
        pytest.skip(f"Model {model_id} hosted by {provider.provider_type} doesn't support OpenAI completions.")


def skip_if_model_doesnt_support_suffix(client_with_models, model_id):
    # To test `fim` ( fill in the middle ) completion, we need to use a model that supports suffix.
    # Use this to specifically test this API functionality.

    # pytest -sv --stack-config="inference=starter" \
    # tests/integration/inference/test_openai_completion.py \
    # --text-model qwen2.5-coder:1.5b \
    # -k test_openai_completion_non_streaming_suffix

    if model_id != "qwen2.5-coder:1.5b":
        pytest.skip(f"Suffix is not supported for the model: {model_id}.")

    provider = provider_from_model(client_with_models, model_id)
    if provider.provider_type != "remote::ollama":
        pytest.skip(f"Provider {provider.provider_type} doesn't support suffix.")


def skip_if_doesnt_support_n(client_with_models, model_id):
    provider = provider_from_model(client_with_models, model_id)
    if provider.provider_type in (
        "remote::sambanova",
        "remote::ollama",
        # https://console.groq.com/docs/openai#currently-unsupported-openai-features
        # -> Error code: 400 - {'error': {'message': "'n' : number must be at most 1", 'type': 'invalid_request_error'}}
        "remote::groq",
        # Error code: 400 - [{'error': {'code': 400, 'message': 'Only one candidate can be specified in the
        # current model', 'status': 'INVALID_ARGUMENT'}}]
        "remote::gemini",
        # https://docs.anthropic.com/en/api/openai-sdk#simple-fields
        "remote::anthropic",
        "remote::vertexai",
        #  Error code: 400 - [{'error': {'code': 400, 'message': 'Unable to submit request because candidateCount must be 1 but
        #  the entered value was 2. Update the candidateCount value and try again.', 'status': 'INVALID_ARGUMENT'}
        "remote::tgi",  # TGI ignores n param silently
    ):
        pytest.skip(f"Model {model_id} hosted by {provider.provider_type} doesn't support n param.")


def skip_if_model_doesnt_support_openai_chat_completion(client_with_models, model_id):
    provider = provider_from_model(client_with_models, model_id)
    if provider.provider_type in (
        "inline::meta-reference",
        "inline::sentence-transformers",
        "inline::vllm",
        "remote::bedrock",
        "remote::cerebras",
        "remote::databricks",
        "remote::runpod",
        "remote::tgi",
        "remote::watsonx",  # watsonx returns 404 when hitting the /openai/v1 endpoint
    ):
        pytest.skip(f"Model {model_id} hosted by {provider.provider_type} doesn't support OpenAI chat completions.")


def skip_if_provider_isnt_vllm(client_with_models, model_id):
    provider = provider_from_model(client_with_models, model_id)
    if provider.provider_type != "remote::vllm":
        pytest.skip(f"Model {model_id} hosted by {provider.provider_type} doesn't support vllm extra_body parameters.")


def skip_if_provider_isnt_openai(client_with_models, model_id):
    provider = provider_from_model(client_with_models, model_id)
    if provider.provider_type != "remote::openai":
        pytest.skip(
            f"Model {model_id} hosted by {provider.provider_type} doesn't support chat completion calls with base64 encoded files."
        )


@pytest.mark.parametrize(
    "test_case",
    [
        "inference:completion:sanity",
    ],
)
def test_openai_completion_non_streaming(llama_stack_client, client_with_models, text_model_id, test_case):
    skip_if_model_doesnt_support_openai_completion(client_with_models, text_model_id)
    tc = TestCase(test_case)

    # ollama needs more verbose prompting for some reason here...
    prompt = "Respond to this question and explain your answer. " + tc["content"]
    response = llama_stack_client.completions.create(
        model=text_model_id,
        prompt=prompt,
        stream=False,
    )
    assert len(response.choices) > 0
    choice = response.choices[0]
    assert len(choice.text) > 10


@pytest.mark.parametrize(
    "test_case",
    [
        "inference:completion:suffix",
    ],
)
def test_openai_completion_non_streaming_suffix(llama_stack_client, client_with_models, text_model_id, test_case):
    skip_if_model_doesnt_support_openai_completion(client_with_models, text_model_id)
    skip_if_model_doesnt_support_suffix(client_with_models, text_model_id)
    tc = TestCase(test_case)

    # ollama needs more verbose prompting for some reason here...
    response = llama_stack_client.completions.create(
        model=text_model_id,
        prompt=tc["content"],
        stream=False,
        suffix=tc["suffix"],
        max_tokens=10,
    )

    assert len(response.choices) > 0
    choice = response.choices[0]
    assert len(choice.text) > 5
    normalized_text = _normalize_text(choice.text)
    assert "france" in normalized_text


@pytest.mark.parametrize(
    "test_case",
    [
        "inference:completion:sanity",
    ],
)
def test_openai_completion_streaming(llama_stack_client, client_with_models, text_model_id, test_case):
    skip_if_model_doesnt_support_openai_completion(client_with_models, text_model_id)
    tc = TestCase(test_case)

    # ollama needs more verbose prompting for some reason here...
    prompt = "Respond to this question and explain your answer. " + tc["content"]
    response = llama_stack_client.completions.create(
        model=text_model_id,
        prompt=prompt,
        stream=True,
        max_tokens=50,
    )
    streamed_content = [chunk.choices[0].text or "" for chunk in response]
    content_str = "".join(streamed_content).lower().strip()
    assert len(content_str) > 10


@pytest.mark.parametrize(
    "prompt_logprobs",
    [
        1,
        0,
    ],
)
def test_openai_completion_prompt_logprobs(llama_stack_client, client_with_models, text_model_id, prompt_logprobs):
    skip_if_provider_isnt_vllm(client_with_models, text_model_id)

    prompt = "Hello, world!"
    response = llama_stack_client.completions.create(
        model=text_model_id,
        prompt=prompt,
        stream=False,
        prompt_logprobs=prompt_logprobs,
    )
    assert len(response.choices) > 0
    choice = response.choices[0]
    assert len(choice.prompt_logprobs) > 0


def test_openai_completion_guided_choice(llama_stack_client, client_with_models, text_model_id):
    skip_if_provider_isnt_vllm(client_with_models, text_model_id)

    prompt = "I am feeling really sad today."
    response = llama_stack_client.completions.create(
        model=text_model_id,
        prompt=prompt,
        stream=False,
        guided_choice=["joy", "sadness"],
    )
    assert len(response.choices) > 0
    choice = response.choices[0]
    assert choice.text in ["joy", "sadness"]


# Run the chat-completion tests with both the OpenAI client and the LlamaStack client


@pytest.mark.parametrize(
    "test_case",
    [
        "inference:chat_completion:non_streaming_01",
        "inference:chat_completion:non_streaming_02",
    ],
)
def test_openai_chat_completion_non_streaming(compat_client, client_with_models, text_model_id, test_case):
    skip_if_model_doesnt_support_openai_chat_completion(client_with_models, text_model_id)
    tc = TestCase(test_case)
    question = tc["question"]
    expected = tc["expected"]

    response = compat_client.chat.completions.create(
        model=text_model_id,
        messages=[
            {
                "role": "user",
                "content": question,
            }
        ],
        stream=False,
    )
    message_content = response.choices[0].message.content.lower().strip()
    assert len(message_content) > 0
    normalized_expected = _normalize_text(expected)
    normalized_content = _normalize_text(message_content)
    assert normalized_expected in normalized_content


@pytest.mark.parametrize(
    "test_case",
    [
        "inference:chat_completion:streaming_01",
        "inference:chat_completion:streaming_02",
    ],
)
def test_openai_chat_completion_streaming(compat_client, client_with_models, text_model_id, test_case):
    skip_if_model_doesnt_support_openai_chat_completion(client_with_models, text_model_id)
    tc = TestCase(test_case)
    question = tc["question"]
    expected = tc["expected"]

    response = compat_client.chat.completions.create(
        model=text_model_id,
        messages=[{"role": "user", "content": question}],
        stream=True,
        timeout=120,  # Increase timeout to 2 minutes for large conversation history
    )
    streamed_content = []
    for chunk in response:
        # On some providers like Azure, the choices are empty on the first chunk, so we need to check for that
        if chunk.choices and len(chunk.choices) > 0 and chunk.choices[0].delta.content:
            streamed_content.append(chunk.choices[0].delta.content.lower().strip())
    assert len(streamed_content) > 0
    normalized_expected = _normalize_text(expected)
    normalized_content = _normalize_text("".join(streamed_content))
    assert normalized_expected in normalized_content


@pytest.mark.parametrize(
    "test_case",
    [
        "inference:chat_completion:streaming_01",
        "inference:chat_completion:streaming_02",
    ],
)
def test_openai_chat_completion_streaming_with_n(compat_client, client_with_models, text_model_id, test_case):
    skip_if_model_doesnt_support_openai_chat_completion(client_with_models, text_model_id)
    skip_if_doesnt_support_n(client_with_models, text_model_id)

    tc = TestCase(test_case)
    question = tc["question"]
    expected = tc["expected"]

    response = compat_client.chat.completions.create(
        model=text_model_id,
        messages=[{"role": "user", "content": question}],
        stream=True,
        timeout=120,  # Increase timeout to 2 minutes for large conversation history,
        n=2,
    )
    streamed_content = {}
    for chunk in response:
        for choice in chunk.choices:
            if choice.delta.content:
                streamed_content[choice.index] = (
                    streamed_content.get(choice.index, "") + choice.delta.content.lower().strip()
                )
    assert len(streamed_content) == 2
    normalized_expected = _normalize_text(expected)
    for i, content in streamed_content.items():
        normalized_content = _normalize_text(content)
        assert normalized_expected in normalized_content, (
            f"Choice {i}: Expected {normalized_expected} in {normalized_content}"
        )


@pytest.mark.parametrize(
    "stream",
    [
        True,
        False,
    ],
)
def test_inference_store(compat_client, client_with_models, text_model_id, stream):
    skip_if_model_doesnt_support_openai_chat_completion(client_with_models, text_model_id)
    client = compat_client
    # make a chat completion
    message = "Hello, world!"
    response = client.chat.completions.create(
        model=text_model_id,
        messages=[
            {
                "role": "user",
                "content": message,
            }
        ],
        stream=stream,
    )
    if stream:
        # accumulate the streamed content
        content = ""
        response_id = None
        for chunk in response:
            if response_id is None and chunk.id:
                response_id = chunk.id
            if chunk.choices and len(chunk.choices) > 0 and chunk.choices[0].delta.content:
                content += chunk.choices[0].delta.content
    else:
        response_id = response.id
        content = response.choices[0].message.content

    tries = 0
    while tries < 10:
        responses = client.chat.completions.list(limit=1000)
        if response_id in [r.id for r in responses.data]:
            break
        else:
            tries += 1
            time.sleep(0.1)
    assert tries < 10, f"Response {response_id} not found after 1 second"

    retrieved_response = client.chat.completions.retrieve(response_id)
    assert retrieved_response.id == response_id
    assert retrieved_response.choices[0].message.content == content, retrieved_response

    input_content = (
        getattr(retrieved_response.input_messages[0], "content", None)
        or retrieved_response.input_messages[0]["content"]
    )
    assert input_content == message, retrieved_response


@pytest.mark.parametrize(
    "stream",
    [
        True,
        False,
    ],
)
def test_inference_store_tool_calls(compat_client, client_with_models, text_model_id, stream):
    skip_if_model_doesnt_support_openai_chat_completion(client_with_models, text_model_id)
    client = compat_client
    # make a chat completion
    message = "What's the weather in Tokyo? Use the get_weather function to get the weather."
    response = client.chat.completions.create(
        model=text_model_id,
        messages=[
            {
                "role": "user",
                "content": message,
            }
        ],
        stream=stream,
        tools=[
            {
                "type": "function",
                "function": {
                    "name": "get_weather",
                    "description": "Get the weather in a given city",
                    "parameters": {
                        "type": "object",
                        "properties": {
                            "city": {"type": "string", "description": "The city to get the weather for"},
                        },
                    },
                },
            }
        ],
    )
    if stream:
        # accumulate the streamed content
        content = ""
        response_id = None
        for chunk in response:
            if response_id is None and chunk.id:
                response_id = chunk.id
            if chunk.choices and len(chunk.choices) > 0:
                if delta := chunk.choices[0].delta:
                    if delta.content:
                        content += delta.content
    else:
        response_id = response.id
        content = response.choices[0].message.content

    # wait for the response to be stored
    tries = 0
    while tries < 10:
        responses = client.chat.completions.list(limit=1000)
        if response_id in [r.id for r in responses.data]:
            break
        else:
            tries += 1
            time.sleep(0.1)

    assert tries < 10, f"Response {response_id} not found after 1 second"

    responses = client.chat.completions.list(limit=1000)
    assert response_id in [r.id for r in responses.data]

    retrieved_response = client.chat.completions.retrieve(response_id)
    assert retrieved_response.id == response_id
    input_content = (
        getattr(retrieved_response.input_messages[0], "content", None)
        or retrieved_response.input_messages[0]["content"]
    )
    assert input_content == message, retrieved_response
    tool_calls = retrieved_response.choices[0].message.tool_calls
    # sometimes model doesn't output tool calls, but we still want to test that the tool was called
    if tool_calls:
        # because we test with small models, just check that we retrieved
        # a tool call with a name and arguments string, but ignore contents
        assert len(tool_calls) == 1
        assert tool_calls[0].function.name
        assert tool_calls[0].function.arguments
    else:
        # failed tool call parses show up as a message with content, so ensure
        # that the retrieve response content matches the original request
        assert retrieved_response.choices[0].message.content == content


def test_openai_chat_completion_non_streaming_with_file(openai_client, client_with_models, text_model_id):
    skip_if_provider_isnt_openai(client_with_models, text_model_id)

    # Hardcoded base64-encoded PDF with "Hello World" text
    pdf_base64 = "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"

    response = openai_client.chat.completions.create(
        model=text_model_id,
        messages=[
            {
                "role": "user",
                "content": "Describe what you see in this PDF file.",
            },
            {
                "role": "user",
                "content": [
                    {
                        "type": "file",
                        "file": {
                            "filename": "my-temp-hello-world-pdf",
                            "file_data": f"data:application/pdf;base64,{pdf_base64}",
                        },
                    }
                ],
            },
        ],
        stream=False,
    )
    message_content = response.choices[0].message.content.lower().strip()
    normalized_content = _normalize_text(message_content)
    assert "hello world" in normalized_content