[Bugfix] Standardize getting number of image patches/tokens (#34358)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>

tests/kernels/core/test_mrope.py

@@ -4,8 +4,6 @@ from typing import NamedTuple
 
 import pytest
 import torch
-from packaging.version import Version
-from transformers import __version__ as TRANSFORMERS_VERSION
 
 from vllm.model_executor.layers.rotary_embedding import get_rope
 from vllm.platforms import current_platform
@@ -46,31 +44,13 @@ class MRoPETestInfo(NamedTuple):
     marks: list[pytest.MarkDecorator] = []
 
 
-TRANSFORMERS_BASE_VERSION = Version(TRANSFORMERS_VERSION).base_version
-
 MODELS_TO_TEST = [
     MRoPETestInfo(model_name="zai-org/GLM-4.1V-9B-Thinking"),
     MRoPETestInfo(model_name="Qwen/Qwen2-VL-7B-Instruct"),
     MRoPETestInfo(model_name="Qwen/Qwen2-VL-72B-Instruct"),
     MRoPETestInfo(model_name="Qwen/Qwen2.5-VL-72B-Instruct"),
-    MRoPETestInfo(
-        model_name="Qwen/Qwen3-VL-4B-Instruct",
-        marks=[
-            pytest.mark.skipif(
-                Version(TRANSFORMERS_BASE_VERSION) < Version("4.57.0"),
-                reason="Qwen3-VL only available after Transformers v4.57",
-            )
-        ],
-    ),
-    MRoPETestInfo(
-        model_name="Qwen/Qwen3-VL-30B-A3B-Instruct",
-        marks=[
-            pytest.mark.skipif(
-                Version(TRANSFORMERS_BASE_VERSION) < Version("4.57.0"),
-                reason="Qwen3-VL only available after Transformers v4.57",
-            )
-        ],
-    ),
+    MRoPETestInfo(model_name="Qwen/Qwen3-VL-4B-Instruct"),
+    MRoPETestInfo(model_name="Qwen/Qwen3-VL-30B-A3B-Instruct"),
 ]
 
 num_tokens_list = [11, 8192]

tests/models/multimodal/generation/test_common.py

@@ -961,12 +961,6 @@ VLM_TEST_SETTINGS = {
                 limit_mm_per_prompt={"image": 4},
             )
         ],
-        marks=[
-            pytest.mark.skipif(
-                Version(TRANSFORMERS_VERSION) == Version("4.57.1"),
-                reason="This model is broken in Transformers v4.57.1",
-            )
-        ],
     ),
     # regression test for https://github.com/vllm-project/vllm/issues/15122
     "qwen2_5_vl-windows-attention": VLMTestInfo(

tests/models/multimodal/processing/test_gemma3.py

@@ -168,6 +168,7 @@ def test_get_image_size_with_most_features(
         image_width=max_image_size.width,
         image_height=max_image_size.height,
         processor=hf_processor,
+        mm_kwargs=hf_processor_mm_kwargs,
     )
 
     prompt = "<start_of_image>"

tests/models/multimodal/processing/test_idefics3.py

@@ -3,7 +3,9 @@
 """Tests for Idefics3's multimodal preprocessing kwargs."""
 
 import pytest
+from packaging.version import Version
 from transformers import Idefics3Config
+from transformers import __version__ as TRANSFORMERS_VERSION
 
 from vllm.multimodal import MULTIMODAL_REGISTRY
 
@@ -11,6 +13,10 @@ from ....conftest import ImageTestAssets
 from ...utils import build_model_context
 
 
+@pytest.mark.skipif(
+    Version(TRANSFORMERS_VERSION) < Version("5.2.0"),
+    reason="See https://github.com/huggingface/transformers/pull/43948",
+)
 @pytest.mark.parametrize("model_id", ["HuggingFaceM4/Idefics3-8B-Llama3"])
 @pytest.mark.parametrize(
     ("mm_processor_kwargs", "expected_toks_per_img"),
@@ -63,7 +69,11 @@ def test_processor_override(
 
     # Ensure the placeholders format are correct
     hf_processor = processor.info.get_hf_processor(**hf_processor_mm_kwargs)
-    hf_processed_inputs = hf_processor(text=prompt, images=mm_data["image"])
+    hf_processed_inputs = hf_processor(
+        text=prompt,
+        images=mm_data["image"],
+        **processor.info.ctx.get_merged_mm_kwargs(hf_processor_mm_kwargs),
+    )
     assert processed_inputs["prompt_token_ids"] == hf_processed_inputs["input_ids"][0]
 
     # Ensure we have the right number of placeholders per num_crops size

tests/models/multimodal/processing/test_qwen2_vl.py

@@ -82,6 +82,7 @@ def test_get_image_size_with_most_features(
         image_width=max_image_size.width,
         image_height=max_image_size.height,
         image_processor=hf_processor.image_processor,
+        mm_kwargs=hf_processor_mm_kwargs,
     )
 
     prompt = "<|vision_start|><|image_pad|><|vision_end|>"

tests/models/multimodal/processing/test_smolvlm.py

@@ -3,7 +3,9 @@
 """Tests for smolvlm's multimodal preprocessing kwargs."""
 
 import pytest
+from packaging.version import Version
 from transformers import SmolVLMConfig
+from transformers import __version__ as TRANSFORMERS_VERSION
 
 from vllm.multimodal import MULTIMODAL_REGISTRY
 
@@ -11,6 +13,10 @@ from ....conftest import ImageTestAssets
 from ...utils import build_model_context
 
 
+@pytest.mark.skipif(
+    Version(TRANSFORMERS_VERSION) < Version("5.2.0"),
+    reason="See https://github.com/huggingface/transformers/pull/43948",
+)
 @pytest.mark.parametrize("model_id", ["HuggingFaceTB/SmolVLM2-2.2B-Instruct"])
 @pytest.mark.parametrize(
     ("mm_processor_kwargs", "expected_toks_per_img"),
@@ -63,7 +69,11 @@ def test_processor_override(
 
     # Ensure the placeholders format are correct
     hf_processor = processor.info.get_hf_processor(**hf_processor_mm_kwargs)
-    hf_processed_inputs = hf_processor(text=prompt, images=mm_data["image"])
+    hf_processed_inputs = hf_processor(
+        text=prompt,
+        images=mm_data["image"],
+        **processor.info.ctx.get_merged_mm_kwargs(hf_processor_mm_kwargs),
+    )
     assert processed_inputs["prompt_token_ids"] == hf_processed_inputs["input_ids"][0]
 
     # Ensure we have the right number of placeholders per num_crops size