From 870f91ab50cc406bc981775a7b10a3c6719fd0be Mon Sep 17 00:00:00 2001
From: zifeitong <zifeitong@gmail.com>
Date: Sat, 24 Aug 2024 18:16:24 -0700
Subject: [PATCH] [Bugfix] Fix Phi-3v crash when input images are of certain
 sizes (#7840)

---
 tests/models/test_phi3v.py          | 27 ++++++++++++++++++++++-----
 vllm/model_executor/models/phi3v.py |  2 --
 2 files changed, 22 insertions(+), 7 deletions(-)

diff --git a/tests/models/test_phi3v.py b/tests/models/test_phi3v.py
index 197e63b1b1e52..40829785d3214 100644
--- a/tests/models/test_phi3v.py
+++ b/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 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 ..conftest import IMAGE_ASSETS, HfRunner, VllmRunner
 from .utils import check_logprobs_close
 
 pytestmark = pytest.mark.vlm
@@ -58,7 +59,7 @@ def vllm_to_hf_output(vllm_output: Tuple[List[int], str,
 def run_test(
     hf_runner: Type[HfRunner],
     vllm_runner: Type[VllmRunner],
-    image_assets: _ImageAssets,
+    images: List[Image.Image],
     model: str,
     *,
     size_factors: List[float],
@@ -77,8 +78,6 @@ def run_test(
     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],
         [
@@ -159,7 +158,7 @@ def test_models(hf_runner, vllm_runner, image_assets, model, size_factors,
     run_test(
         hf_runner,
         vllm_runner,
-        image_assets,
+        [asset.pil_image for asset in image_assets],
         model,
         size_factors=size_factors,
         dtype=dtype,
@@ -167,3 +166,21 @@ def test_models(hf_runner, vllm_runner, image_assets, model, size_factors,
         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, vllm_runner, image_assets, model,
+                         dtype) -> None:
+    # Regression test for #7840.
+    run_test(
+        hf_runner,
+        vllm_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,
+    )
diff --git a/vllm/model_executor/models/phi3v.py b/vllm/model_executor/models/phi3v.py
index 4854377215608..2e52531989232 100644
--- a/vllm/model_executor/models/phi3v.py
+++ b/vllm/model_executor/models/phi3v.py
@@ -400,8 +400,6 @@ 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,
                                                           input_height=h)