fix: phi3v crash with unusual image sizes (#916)

aphrodite/modeling/models/phi3v.py

@@ -396,7 +396,6 @@ def input_processor_for_phi3v(ctx: InputContext, llm_inputs: LLMInputs):
     image_data = multi_modal_data["image"]
     if isinstance(image_data, Image.Image):
         w, h = image_data.size
-        w, h = _calc_hd_transform_size(width=w, height=h)
 
         image_feature_size = get_phi3v_image_feature_size(hf_config,
                                                           input_width=w,

tests/models/test_phi3v.py

@@ -3,13 +3,14 @@
 from typing import List, Optional, Tuple, Type
 
 import pytest
+from PIL import Image
 from transformers import AutoTokenizer
 
 from aphrodite.common.sequence import SampleLogprobs
 from aphrodite.common.utils import is_cpu, is_hip
 from aphrodite.multimodal.utils import rescale_image_size
 
-from ..conftest import IMAGE_ASSETS, AphroditeRunner, HfRunner, _ImageAssets
+from ..conftest import IMAGE_ASSETS, AphroditeRunner, HfRunner
 from .utils import check_logprobs_close
 
 pytestmark = pytest.mark.vlm
@@ -58,7 +59,7 @@ def aphrodite_to_hf_output(aphrodite_output: Tuple[List[int], str,
 def run_test(
     hf_runner: Type[HfRunner],
     aphrodite_runner: Type[AphroditeRunner],
-    image_assets: _ImageAssets,
+    images: List[Image.Image],
     model: str,
     *,
     size_factors: List[float],
@@ -77,7 +78,6 @@ def run_test(
     Note, the text input is also adjusted to abide by aphrodite 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],
@@ -112,7 +112,7 @@ def run_test(
     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
+        eos_token_id = hf_model.processor.eos_token_id
         hf_outputs_per_image = [
             hf_model.generate_greedy_logprobs_limit(prompts,
                                                     max_tokens,
@@ -159,11 +159,28 @@ def test_models(hf_runner, aphrodite_runner, image_assets, model, size_factors,
     run_test(
         hf_runner,
         aphrodite_runner,
-        image_assets,
+        [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,
     )
+
+
+@pytest.mark.parametrize("model", models)
+@pytest.mark.parametrize("dtype", [target_dtype])
+def test_regression_7840(hf_runner, aphrodite_runner, image_assets, model,
+                         dtype) -> None:
+    run_test(
+        hf_runner,
+        aphrodite_runner,
+        [image_assets[0].pil_image.resize((465, 226))],
+        model,
+        size_factors=[1.0],
+        dtype=dtype,
+        max_tokens=128,
+        num_logprobs=10,
+        tensor_parallel_size=1,
+    )