[Model] Initialize Phi-3-vision support (vllm-project#4986)

docs/source/models/supported_models.rst

@@ -135,6 +135,10 @@ Alongside each architecture, we include some popular models that use it.
     - Phi-3-Small
     - :code:`microsoft/Phi-3-small-8k-instruct`, :code:`microsoft/Phi-3-small-128k-instruct`, etc.
     -
+  * - :code:`Phi3VForCausalLM`
+    - Phi-3-Vision
+    - :code:`microsoft/Phi-3-vision-128k-instruct`, etc.
+    -
   * - :code:`QWenLMHeadModel`
     - Qwen
     - :code:`Qwen/Qwen-7B`, :code:`Qwen/Qwen-7B-Chat`, etc.

examples/phi3v_example.py

@@ -0,0 +1,57 @@
+import os
+import subprocess
+
+from PIL import Image
+
+from vllm import LLM, SamplingParams
+from vllm.multimodal.image import ImagePixelData
+
+
+def run_phi3v():
+    model_path = "microsoft/Phi-3-vision-128k-instruct"
+    llm = LLM(
+        model=model_path,
+        trust_remote_code=True,
+        max_model_len=4096,
+        image_input_type="pixel_values",
+        image_token_id=32044,
+        image_input_shape="1,3,1008,1344",
+        image_feature_size=1921,
+        disable_image_processor=False,
+    )
+
+    image = Image.open("images/cherry_blossom.jpg")
+
+    # single-image prompt
+    prompt = "<|user|>\n<|image_1|>\nWhat is the season?<|end|>\n<|assistant|>\n"  # noqa: E501
+    prompt = prompt.replace("<|image_1|>", "<|image|>" * 1921 + "<s>")
+
+    sampling_params = SamplingParams(temperature=0, max_tokens=64)
+
+    outputs = llm.generate({
+        "prompt": prompt,
+        "sampling_params": sampling_params,
+        "multi_modal_data": ImagePixelData(image),
+    })
+    for o in outputs:
+        generated_text = o.outputs[0].text
+        print(generated_text)
+
+
+if __name__ == "__main__":
+    s3_bucket_path = "s3://air-example-data-2/vllm_opensource_llava/"
+    local_directory = "images"
+
+    # Make sure the local directory exists or create it
+    os.makedirs(local_directory, exist_ok=True)
+
+    # Use AWS CLI to sync the directory, assume anonymous access
+    subprocess.check_call([
+        "aws",
+        "s3",
+        "sync",
+        s3_bucket_path,
+        local_directory,
+        "--no-sign-request",
+    ])
+    run_phi3v()

requirements-test.txt

@@ -14,6 +14,7 @@ peft
 requests
 ray
 sentence-transformers # required for embedding
+torchvision # required for the image processor of phi3v
 
 # Benchmarking
 aiohttp

tests/conftest.py

@@ -144,6 +144,7 @@ def __init__(
         model_name: str,
         dtype: str = "half",
         *,
+        model_kwargs: Optional[Dict[str, Any]] = None,
         is_embedding_model: bool = False,
         is_vision_model: bool = False,
     ) -> None:
@@ -166,11 +167,13 @@ def __init__(
             else:
                 auto_cls = AutoModelForCausalLM
 
+            model_kwargs = model_kwargs if model_kwargs is not None else {}
             self.model = self.wrap_device(
                 auto_cls.from_pretrained(
                     model_name,
                     torch_dtype=torch_dtype,
                     trust_remote_code=True,
+                    **model_kwargs,
                 ))
 
         self.tokenizer = AutoTokenizer.from_pretrained(

tests/models/test_phi3v.py

@@ -0,0 +1,124 @@
+from typing import List, Tuple
+
+import pytest
+from transformers import AutoTokenizer
+
+from vllm.config import VisionLanguageConfig
+from vllm.utils import is_cpu
+
+from ..conftest import IMAGE_FILES
+
+pytestmark = pytest.mark.llava
+
+# The image token is placed before "user" on purpose so that the test can pass
+HF_IMAGE_PROMPTS = [
+    "<|user|>\n<|image_1|>\nWhat's the content of the image?<|end|>\n<|assistant|>\n",  # noqa: E501
+    "<|user|>\n<|image_1|>\nWhat is the season?<|end|>\n<|assistant|>\n",
+]
+
+assert len(HF_IMAGE_PROMPTS) == len(IMAGE_FILES)
+
+
+def iter_phi3v_configs(model_name: str):
+    image_hw_to_feature_size = {
+        (1008, 1344): 1921,
+    }
+
+    for (h, w), f in image_hw_to_feature_size.items():
+        for input_type, input_shape in [
+            (VisionLanguageConfig.ImageInputType.PIXEL_VALUES, (1, 3, h, w)),
+        ]:
+            yield (model_name,
+                   VisionLanguageConfig(image_input_type=input_type,
+                                        image_feature_size=f,
+                                        image_token_id=32044,
+                                        image_input_shape=input_shape,
+                                        image_processor=model_name,
+                                        image_processor_revision=None))
+
+
+model_and_vl_config = [
+    *iter_phi3v_configs("microsoft/Phi-3-vision-128k-instruct"),
+]
+
+
+def vllm_to_hf_output(vllm_output: Tuple[List[int], str],
+                      vlm_config: VisionLanguageConfig, model_id: str):
+    """Sanitize vllm output to be comparable with hf output.
+    The function reduces `input_ids` from 1, 32000, 32000, ..., 32000,
+    x1, x2, x3 ... to 1, 32000, x1, x2, x3 ...
+    It also reduces `output_str` from "<image><image>bla" to "bla".
+    """
+    input_ids, output_str = vllm_output
+    image_token_id = vlm_config.image_token_id
+
+    tokenizer = AutoTokenizer.from_pretrained(model_id)
+    image_token_str = tokenizer.decode(image_token_id)
+
+    hf_input_ids = [
+        input_id if input_id != image_token_id else 0
+        for idx, input_id in enumerate(input_ids)
+    ]
+    hf_output_str = output_str \
+        .replace(image_token_str * vlm_config.image_feature_size, "") \
+        .replace("<s>", " ").replace("<|user|>", "") \
+        .replace("<|end|>\n<|assistant|>", " ")
+
+    return hf_input_ids, hf_output_str
+
+
+target_dtype = "half"
+if is_cpu():
+    target_dtype = "bfloat16"
+
+
+# TODO: Add test for `tensor_parallel_size` [ref: PR #3883]
+# Since we use _attn_implementation="eager" for hf_runner, here is
+# numeric difference for longer context and test can't pass
+@pytest.mark.parametrize("model_and_config", model_and_vl_config)
+@pytest.mark.parametrize("dtype", [target_dtype])
+@pytest.mark.parametrize("max_tokens", [8])
+def test_models(hf_runner, vllm_runner, hf_images, vllm_images,
+                model_and_config, dtype: str, max_tokens: int) -> 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 MultiModalData 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.
+    """
+    model_id, vlm_config = model_and_config
+
+    # 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_id, dtype=dtype,
+                   model_kwargs=hf_model_kwargs) as hf_model:
+        hf_outputs = hf_model.generate_greedy(HF_IMAGE_PROMPTS,
+                                              max_tokens,
+                                              images=hf_images)
+
+    vllm_image_prompts = [
+        p.replace("<|image_1|>",
+                  "<|image|>" * vlm_config.image_feature_size + "<s>")
+        for p in HF_IMAGE_PROMPTS
+    ]
+
+    with vllm_runner(model_id,
+                     max_model_len=2048,
+                     dtype=dtype,
+                     enforce_eager=True,
+                     **vlm_config.as_cli_args_dict()) as vllm_model:
+        vllm_outputs = vllm_model.generate_greedy(vllm_image_prompts,
+                                                  max_tokens,
+                                                  images=vllm_images)
+
+    for i in range(len(HF_IMAGE_PROMPTS)):
+        hf_output_ids, hf_output_str = hf_outputs[i]
+        vllm_output_ids, vllm_output_str = vllm_to_hf_output(
+            vllm_outputs[i], vlm_config, model_id)
+        assert hf_output_str == vllm_output_str, (
+            f"Test{i}:\nHF: {hf_output_str!r}\nvLLM: {vllm_output_str!r}")
+        assert hf_output_ids == vllm_output_ids, (
+            f"Test{i}:\nHF: {hf_output_ids}\nvLLM: {vllm_output_ids}")

vllm/model_executor/models/__init__.py

@@ -49,6 +49,7 @@
     "OrionForCausalLM": ("orion", "OrionForCausalLM"),
     "PhiForCausalLM": ("phi", "PhiForCausalLM"),
     "Phi3ForCausalLM": ("llama", "LlamaForCausalLM"),
+    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
     "QWenLMHeadModel": ("qwen", "QWenLMHeadModel"),
     "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
     "Qwen2MoeForCausalLM": ("qwen2_moe", "Qwen2MoeForCausalLM"),