[Model][VLM] Support multi-images inputs for Phi-3-vision models (#7783)

tests/models/test_phi3v.py

@@ -21,6 +21,7 @@ HF_IMAGE_PROMPTS = IMAGE_ASSETS.prompts({
     "cherry_blossom":
     "<|user|>\n<|image_1|>\nWhat is the season?<|end|>\n<|assistant|>\n",
 })
+HF_MULTIIMAGE_IMAGE_PROMPT = "<|user|>\n<|image_1|>\n<|image_2|>\nDescribe these images.<|end|>\n<|assistant|>\n"  # noqa: E501
 
 models = ["microsoft/Phi-3.5-vision-instruct"]
 
@@ -184,3 +185,113 @@ def test_regression_7840(hf_runner, vllm_runner, image_assets, model,
         num_logprobs=10,
         tensor_parallel_size=1,
     )
+
+
+def run_multi_image_test(
+    hf_runner: Type[HfRunner],
+    vllm_runner: Type[VllmRunner],
+    images: List[Image.Image],
+    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.
+    """
+
+    inputs_per_case = [
+        ([HF_MULTIIMAGE_IMAGE_PROMPT for _ in size_factors],
+         [[rescale_image_size(image, factor) for image in images]
+          for factor in size_factors])
+    ]
+
+    # 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,
+                     limit_mm_per_prompt={"image": len(images)},
+                     dtype=dtype,
+                     tensor_parallel_size=tensor_parallel_size,
+                     distributed_executor_backend=distributed_executor_backend,
+                     enforce_eager=True) as vllm_model:
+        vllm_outputs_per_case = [
+            vllm_model.generate_greedy_logprobs(prompts,
+                                                max_tokens,
+                                                num_logprobs=num_logprobs,
+                                                images=images)
+            for prompts, images in inputs_per_case
+        ]
+
+    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_case = [
+            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_case
+        ]
+
+    for hf_outputs, vllm_outputs in zip(hf_outputs_per_case,
+                                        vllm_outputs_per_case):
+        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",
+        )
+
+
+@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", [5])
+def test_multi_images_models(hf_runner, vllm_runner, image_assets, model,
+                             size_factors, dtype: str, max_tokens: int,
+                             num_logprobs: int) -> None:
+    run_multi_image_test(
+        hf_runner,
+        vllm_runner,
+        [asset.pil_image for asset in image_assets],
+        model,
+        size_factors=size_factors,
+        dtype=dtype,
+        max_tokens=max_tokens,
+        num_logprobs=num_logprobs,
+        tensor_parallel_size=1,
+    )