test_phi3v.py

import os
import re
from typing import List, Optional, Tuple, Type

import pytest
from transformers import AutoTokenizer

from vllm.multimodal.utils import rescale_image_size
from vllm.sequence import SampleLogprobs
from vllm.utils import is_cpu, is_hip

from ..conftest import IMAGE_ASSETS, HfRunner, VllmRunner, _ImageAssets
from .utils import check_logprobs_close

pytestmark = pytest.mark.vlm

HF_IMAGE_PROMPTS = IMAGE_ASSETS.prompts({
    "stop_sign":
    "<|user|>\n<|image_1|>\nWhat's the content of the image?<|end|>\n<|assistant|>\n",  # noqa: E501
    "cherry_blossom":
    "<|user|>\n<|image_1|>\nWhat is the season?<|end|>\n<|assistant|>\n",
})

models = ["microsoft/Phi-3-vision-128k-instruct"]


def vllm_to_hf_output(vllm_output: Tuple[List[int], str,
                                         Optional[SampleLogprobs]],
                      model: str):
    """Sanitize vllm output to be comparable with hf output."""
    _, output_str, out_logprobs = vllm_output

    output_str_without_image = re.sub(r"(<\|image_\d+\|>)+", "", output_str)
    assert output_str_without_image[0] == " "
    output_str_without_image = output_str_without_image[1:]

    hf_output_str = output_str_without_image + "<|end|><|endoftext|>"

    tokenizer = AutoTokenizer.from_pretrained(model)
    hf_output_ids = tokenizer.encode(output_str_without_image)
    assert hf_output_ids[0] == 1
    hf_output_ids = hf_output_ids[1:]

    return hf_output_ids, hf_output_str, out_logprobs


target_dtype = "half"
if is_cpu():
    target_dtype = "bfloat16"

# ROCm Triton FA can run into shared memory issues with these models,
# use other backends in the meantime
# FIXME (mattwong, gshtrasb, hongxiayan)
if is_hip():
    os.environ["VLLM_USE_TRITON_FLASH_ATTN"] = "0"


def run_test(
    hf_runner: Type[HfRunner],
    vllm_runner: Type[VllmRunner],
    image_assets: _ImageAssets,
    model: str,
    *,
    size_factors: List[float],
    dtype: str,
    max_tokens: int,
    num_logprobs: int,
    tensor_parallel_size: int,
    distributed_executor_backend: Optional[str] = None,
):
    """Inference result should be the same between hf and vllm.

    All the image fixtures for the test is under tests/images.
    For huggingface runner, we provide the PIL images as input.
    For vllm runner, we provide MultiModalDataDict objects 
    and corresponding MultiModalConfig as input.
    Note, the text input is also adjusted to abide by vllm contract.
    The text output is sanitized to be able to compare with hf.
    """
    images = [asset.pil_image for asset in image_assets]

    inputs_per_image = [(
        [prompt for _ in size_factors],
        [
            rescale_image_size(image, factor, transpose=idx)
            for idx, factor in enumerate(size_factors)
        ],
    ) for image, prompt in zip(images, HF_IMAGE_PROMPTS)]

    # NOTE: take care of the order. run vLLM first, and then run HF.
    # vLLM needs a fresh new process without cuda initialization.
    # if we run HF first, the cuda initialization will be done and it
    # will hurt multiprocessing backend with fork method (the default method).

    # max_model_len should be greater than image_feature_size
    with vllm_runner(model,
                     max_model_len=4096,
                     max_num_seqs=1,
                     dtype=dtype,
                     tensor_parallel_size=tensor_parallel_size,
                     distributed_executor_backend=distributed_executor_backend,
                     enforce_eager=True) as vllm_model:
        vllm_outputs_per_image = [
            vllm_model.generate_greedy_logprobs(prompts,
                                                max_tokens,
                                                num_logprobs=num_logprobs,
                                                images=images)
            for prompts, images in inputs_per_image
        ]

    # use eager mode for hf runner, since phi3_v didn't work with flash_attn
    hf_model_kwargs = {"_attn_implementation": "eager"}
    with hf_runner(model, dtype=dtype,
                   model_kwargs=hf_model_kwargs) as hf_model:
        eos_token_id = hf_model.processor.tokenizer.eos_token_id
        hf_outputs_per_image = [
            hf_model.generate_greedy_logprobs_limit(prompts,
                                                    max_tokens,
                                                    num_logprobs=num_logprobs,
                                                    images=images,
                                                    eos_token_id=eos_token_id)
            for prompts, images in inputs_per_image
        ]

    for hf_outputs, vllm_outputs in zip(hf_outputs_per_image,
                                        vllm_outputs_per_image):
        check_logprobs_close(
            outputs_0_lst=hf_outputs,
            outputs_1_lst=[
                vllm_to_hf_output(vllm_output, model)
                for vllm_output in vllm_outputs
            ],
            name_0="hf",
            name_1="vllm",
        )


# Since we use _attn_implementation="eager" for hf_runner, there is more
# significant numerical difference. The basic `logprobs=5` fails to pass.
@pytest.mark.parametrize("model", models)
@pytest.mark.parametrize(
    "size_factors",
    [
        # No image
        [],
        # Single-scale
        [1.0],
        # Single-scale, batched
        [1.0, 1.0, 1.0],
        # Multi-scale
        [0.25, 0.5, 1.0],
    ],
)
@pytest.mark.parametrize("dtype", [target_dtype])
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [10])
def test_models(hf_runner, vllm_runner, image_assets, model, size_factors,
                dtype: str, max_tokens: int, num_logprobs: int) -> None:
    run_test(
        hf_runner,
        vllm_runner,
        image_assets,
        model,
        size_factors=size_factors,
        dtype=dtype,
        max_tokens=max_tokens,
        num_logprobs=num_logprobs,
        tensor_parallel_size=1,
    )