[Model] Initialize Fuyu-8B support (vllm-project#3924)

Co-authored-by: Roger Wang <ywang@roblox.com>

Author: 2 people

docs/source/models/supported_models.rst

@@ -137,6 +137,10 @@ Decoder-only Language Models
     - Phi-3-Small
     - :code:`microsoft/Phi-3-small-8k-instruct`, :code:`microsoft/Phi-3-small-128k-instruct`, etc.
     -
+  * - :code:`PersimmonForCausalLM`
+    - Persimmon
+    - :code:`adept/persimmon-8b-base`, :code:`adept/persimmon-8b-chat`, etc.
+    - 
   * - :code:`QWenLMHeadModel`
     - Qwen
     - :code:`Qwen/Qwen-7B`, :code:`Qwen/Qwen-7B-Chat`, etc.
@@ -178,6 +182,10 @@ Vision Language Models
     - Models
     - Example HuggingFace Models
     - :ref:`LoRA <lora>`
+  * - :code:`FuyuForCausalLM`
+    - Fuyu
+    - :code:`adept/fuyu-8b` etc.
+    - 
   * - :code:`LlavaForConditionalGeneration`
     - LLaVA-1.5
     - :code:`llava-hf/llava-1.5-7b-hf`, :code:`llava-hf/llava-1.5-13b-hf`, etc.

examples/fuyu_example.py

@@ -0,0 +1,31 @@
+import requests
+from PIL import Image
+
+from vllm import LLM, SamplingParams
+
+
+def run_fuyu():
+    llm = LLM(model="adept/fuyu-8b", max_model_len=4096)
+
+    # single-image prompt
+    prompt = "What is the highest life expectancy at of male?\n"
+    url = "https://huggingface.co/adept/fuyu-8b/resolve/main/chart.png"
+    image = Image.open(requests.get(url, stream=True).raw)
+    sampling_params = SamplingParams(temperature=0, max_tokens=64)
+
+    outputs = llm.generate(
+        {
+            "prompt": prompt,
+            "multi_modal_data": {
+                "image": image
+            },
+        },
+        sampling_params=sampling_params)
+
+    for o in outputs:
+        generated_text = o.outputs[0].text
+        print(generated_text)
+
+
+if __name__ == "__main__":
+    run_fuyu()

tests/models/test_fuyu.py

@@ -0,0 +1,142 @@
+from typing import List, Optional, Tuple, Type
+
+import pytest
+
+from vllm.multimodal.utils import rescale_image_size
+from vllm.sequence import SampleLogprobs
+from vllm.utils import is_cpu
+
+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": "What's the content of the image?\n",  # noqa: E501
+    "cherry_blossom": "What is the season?\n",
+    "boardwalk": "What's in this image?\n",
+})
+
+models = ["adept/fuyu-8b"]
+
+
+def vllm_to_hf_output(vllm_output: Tuple[List[int], str,
+                                         Optional[SampleLogprobs]]):
+    """Sanitize vllm output to be comparable with hf output."""
+    output_ids, output_str, out_logprobs = vllm_output
+
+    hf_output_str = output_str.lstrip() + "|ENDOFTEXT|"
+
+    return output_ids, hf_output_str, out_logprobs
+
+
+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 vision language config 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) for factor in 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=2560,
+                     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=vllm_images)
+            for prompts, vllm_images in inputs_per_image
+        ]
+
+    with hf_runner(model, dtype=dtype) as hf_model:
+        hf_model.model.get_output_embeddings = lambda: \
+            hf_model.model.language_model.get_output_embeddings()
+        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=hf_images,
+                                                    eos_token_id=eos_token_id)
+            for prompts, hf_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) for vllm_output in vllm_outputs
+            ],
+            name_0="hf",
+            name_1="vllm",
+        )
+
+
+target_dtype = "half"
+if is_cpu():
+    target_dtype = "bfloat16"
+
+
+@pytest.mark.parametrize("model", models)
+@pytest.mark.parametrize(
+    "size_factors",
+    [
+        # No image
+        [],
+        # Single-scale
+        [0.25],
+        # Single-scale, batched
+        [0.25, 0.25, 0.25],
+        # Multi-scale
+        [0.25, 0.2, 0.15],
+    ],
+)
+@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,
+    )

vllm/model_executor/models/__init__.py

@@ -23,6 +23,7 @@
     "DeepseekForCausalLM": ("deepseek", "DeepseekForCausalLM"),
     "DeepseekV2ForCausalLM": ("deepseek_v2", "DeepseekV2ForCausalLM"),
     "FalconForCausalLM": ("falcon", "FalconForCausalLM"),
+    "FuyuForCausalLM": ("fuyu", "FuyuForCausalLM"),
     "GemmaForCausalLM": ("gemma", "GemmaForCausalLM"),
     "Gemma2ForCausalLM": ("gemma2", "Gemma2ForCausalLM"),
     "GPT2LMHeadModel": ("gpt2", "GPT2LMHeadModel"),
@@ -49,6 +50,7 @@
     "OlmoForCausalLM": ("olmo", "OlmoForCausalLM"),
     "OPTForCausalLM": ("opt", "OPTForCausalLM"),
     "OrionForCausalLM": ("orion", "OrionForCausalLM"),
+    "PersimmonForCausalLM": ("persimmon", "PersimmonForCausalLM"),
     "PaliGemmaForConditionalGeneration":
     ("paligemma", "PaliGemmaForConditionalGeneration"),
     "PhiForCausalLM": ("phi", "PhiForCausalLM"),