[VLM] Reorganize profiling/processing-related code (vllm-project#11812)

Signed-off-by: DarkLight1337 <[email protected]>

Author: DarkLight1337

tests/models/decoder_only/vision_language/processing/test_llava_next.py

@@ -4,24 +4,17 @@
 import pytest
 from PIL import Image
 from pqdm.threads import pqdm
-from transformers import AutoTokenizer
 
-from vllm.inputs import InputProcessingContext
+from vllm.multimodal import MULTIMODAL_REGISTRY
 from vllm.multimodal.parse import ImageSize
+from vllm.multimodal.processing import BaseMultiModalProcessor
+from vllm.multimodal.utils import cached_get_tokenizer
 
 from ....utils import build_model_context
 
 
-# Fixtures lazy import to avoid initializing CUDA during test collection
-@pytest.fixture()
-def processor_for_llava_next():
-    from vllm.model_executor.models.llava_next import (
-        LlavaNextMultiModalProcessor)
-    return LlavaNextMultiModalProcessor
-
-
 def _validate_image_prompt_replacements_one(
-    processor,
+    processor: BaseMultiModalProcessor,
     num_imgs: int,
     failed_size_excs: list[tuple[ImageSize, Exception]],
     image_size: ImageSize,
@@ -78,20 +71,17 @@ def _test_image_prompt_replacements(
 
 @pytest.mark.parametrize("model_id", ["llava-hf/llava-v1.6-mistral-7b-hf"])
 @pytest.mark.parametrize("num_imgs", [1, 2])
-def test_processor_prompt_replacements_regression(
-    processor_for_llava_next,
-    model_id: str,
-    num_imgs: int,
-):
+def test_processor_prompt_replacements_regression(model_id, num_imgs):
     ctx = build_model_context(
         model_name=model_id,
         tokenizer_name=model_id,
         mm_processor_kwargs=None,
         limit_mm_per_prompt={"image": num_imgs},
     )
-    tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
-    ctx = InputProcessingContext(ctx.model_config, tokenizer)
-    processor = processor_for_llava_next(ctx)
+    processor = MULTIMODAL_REGISTRY.create_processor(
+        ctx.model_config,
+        tokenizer=cached_get_tokenizer(ctx.model_config.tokenizer),
+    )
 
     image_ratios = [(171, 152), (184, 161), (198, 176), (333, 296), (369, 328),
                     (488, 183), (2560, 1669)]
@@ -111,20 +101,17 @@ def test_processor_prompt_replacements_regression(
                   "Comment this out to run it manually.")
 @pytest.mark.parametrize("model_id", ["llava-hf/llava-v1.6-mistral-7b-hf"])
 @pytest.mark.parametrize("num_imgs", [1])
-def test_processor_prompt_replacements_all(
-    processor_for_llava_next,
-    model_id: str,
-    num_imgs: int,
-):
+def test_processor_prompt_replacements_all(model_id, num_imgs):
     ctx = build_model_context(
         model_name=model_id,
         tokenizer_name=model_id,
         mm_processor_kwargs=None,
         limit_mm_per_prompt={"image": num_imgs},
     )
-    tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
-    ctx = InputProcessingContext(ctx.model_config, tokenizer)
-    processor = processor_for_llava_next(ctx)
+    processor = MULTIMODAL_REGISTRY.create_processor(
+        ctx.model_config,
+        tokenizer=cached_get_tokenizer(ctx.model_config.tokenizer),
+    )
 
     seen_aspect_ratios = set[float]()
     image_sizes = list[ImageSize]()

tests/models/decoder_only/vision_language/processing/test_llava_onevision.py

@@ -4,24 +4,17 @@
 import pytest
 from PIL import Image
 from pqdm.threads import pqdm
-from transformers import AutoTokenizer
 
-from vllm.inputs import InputProcessingContext
+from vllm.multimodal import MULTIMODAL_REGISTRY
 from vllm.multimodal.parse import ImageSize
+from vllm.multimodal.processing import BaseMultiModalProcessor
+from vllm.multimodal.utils import cached_get_tokenizer
 
 from ....utils import build_model_context
 
 
-# Fixtures lazy import to avoid initializing CUDA during test collection
-@pytest.fixture()
-def processor_for_llava_onevision():
-    from vllm.model_executor.models.llava_onevision import (
-        LlavaOnevisionMultiModalProcessor)
-    return LlavaOnevisionMultiModalProcessor
-
-
 def _validate_image_prompt_replacements_one(
-    processor,
+    processor: BaseMultiModalProcessor,
     num_imgs: int,
     failed_size_excs: list[tuple[ImageSize, Exception]],
     image_size: ImageSize,
@@ -77,20 +70,17 @@ def _test_image_prompt_replacements(
 @pytest.mark.parametrize("model_id",
                          ["llava-hf/llava-onevision-qwen2-0.5b-ov-hf"])
 @pytest.mark.parametrize("num_imgs", [1, 2])
-def test_processor_prompt_replacements_regression(
-    processor_for_llava_onevision,
-    model_id: str,
-    num_imgs: int,
-):
+def test_processor_prompt_replacements_regression(model_id, num_imgs):
     ctx = build_model_context(
         model_name=model_id,
         tokenizer_name=model_id,
         mm_processor_kwargs=None,
         limit_mm_per_prompt={"image": num_imgs},
     )
-    tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
-    ctx = InputProcessingContext(ctx.model_config, tokenizer)
-    processor = processor_for_llava_onevision(ctx)
+    processor = MULTIMODAL_REGISTRY.create_processor(
+        ctx.model_config,
+        tokenizer=cached_get_tokenizer(ctx.model_config.tokenizer),
+    )
 
     image_ratios = [(171, 152), (184, 161), (198, 176), (333, 296), (369, 328),
                     (488, 183), (2560, 1669)]
@@ -111,20 +101,17 @@ def test_processor_prompt_replacements_regression(
 @pytest.mark.parametrize("model_id",
                          ["llava-hf/llava-onevision-qwen2-0.5b-ov-hf"])
 @pytest.mark.parametrize("num_imgs", [1])
-def test_processor_prompt_replacements_all(
-    processor_for_llava_onevision,
-    model_id: str,
-    num_imgs: int,
-):
+def test_processor_prompt_replacements_all(model_id, num_imgs):
     ctx = build_model_context(
         model_name=model_id,
         tokenizer_name=model_id,
         mm_processor_kwargs=None,
         limit_mm_per_prompt={"image": num_imgs},
     )
-    tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
-    ctx = InputProcessingContext(ctx.model_config, tokenizer)
-    processor = processor_for_llava_onevision(ctx)
+    processor = MULTIMODAL_REGISTRY.create_processor(
+        ctx.model_config,
+        tokenizer=cached_get_tokenizer(ctx.model_config.tokenizer),
+    )
 
     seen_aspect_ratios = set[float]()
     image_sizes = list[ImageSize]()

tests/models/decoder_only/vision_language/processing/test_phi3v.py

@@ -1,21 +1,13 @@
 """Tests for phi3v's multimodal preprocessing kwargs."""
 import pytest
-from transformers import AutoTokenizer
 
-from vllm.inputs import InputProcessingContext
-from vllm.model_executor.models.phi3v import _IMAGE_TOKEN_ID
+from vllm.multimodal import MULTIMODAL_REGISTRY
+from vllm.multimodal.utils import cached_get_tokenizer
 
 from .....conftest import _ImageAssets
 from ....utils import build_model_context
 
 
-# Wrap lazy imports to avoid initializing CUDA during test collection
-@pytest.fixture()
-def processor_for_phi3v():
-    from vllm.model_executor.models.phi3v import Phi3VMultiModalProcessor
-    return Phi3VMultiModalProcessor
-
-
 @pytest.mark.parametrize("model_id", ["microsoft/Phi-3.5-vision-instruct"])
 # yapf: disable
 @pytest.mark.parametrize(
@@ -29,29 +21,33 @@ def processor_for_phi3v():
 # yapf: enable
 @pytest.mark.parametrize("num_imgs", [1, 2])
 def test_processor_override(
-    processor_for_phi3v,
     image_assets: _ImageAssets,
     model_id: str,
     mm_processor_kwargs: dict[str, int],
     expected_toks_per_img: int,
     num_imgs: int,
 ):
     """Ensure input_processor_for_phi3v handles num_crops properly."""
+    # Avoid initializing CUDA early
+    from vllm.model_executor.models.phi3v import _IMAGE_TOKEN_ID
+
     ctx = build_model_context(
         model_name=model_id,
         tokenizer_name=model_id,
         trust_remote_code=True,
         limit_mm_per_prompt={"image": num_imgs},
     )
-    tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
-    ctx = InputProcessingContext(ctx.model_config, tokenizer)
+    tokenizer = cached_get_tokenizer(ctx.model_config.tokenizer)
+    processor = MULTIMODAL_REGISTRY.create_processor(
+        ctx.model_config,
+        tokenizer=tokenizer,
+    )
 
     # Build the image str / prompt based on the number of images we pass
     img_str = "".join([f"<|image_{idx}|>\n" for idx in range(1, num_imgs + 1)])
     prompt = f"<|user|>\n{img_str}<|end|>\n<|assistant|>\n"
     mm_data = {"image": [image_assets[0].pil_image] * num_imgs}
 
-    processor = processor_for_phi3v(ctx)
     processed_inputs = processor.apply(prompt, mm_data, mm_processor_kwargs)
 
     # Ensure we have the right number of placeholders per num_crops size

tests/models/decoder_only/vision_language/processing/test_qwen2_vl.py

@@ -1,19 +1,12 @@
 import pytest
-from transformers import AutoTokenizer
 
-from vllm.inputs import InputProcessingContext
+from vllm.multimodal import MULTIMODAL_REGISTRY
+from vllm.multimodal.utils import cached_get_tokenizer
 
 from .....conftest import _ImageAssets
 from ....utils import build_model_context
 
 
-# Fixtures lazy import to avoid initializing CUDA during test collection
-@pytest.fixture()
-def processor_for_qwen2_vl():
-    from vllm.model_executor.models.qwen2_vl import Qwen2VLMultiModalProcessor
-    return Qwen2VLMultiModalProcessor
-
-
 @pytest.mark.parametrize("model_id", ["Qwen/Qwen2-VL-2B-Instruct"])
 # yapf: disable
 @pytest.mark.parametrize(
@@ -24,7 +17,6 @@ def processor_for_qwen2_vl():
 # yapf: enable
 @pytest.mark.parametrize("num_imgs", [1, 2])
 def test_processor_override(
-    processor_for_qwen2_vl,
     image_assets: _ImageAssets,
     model_id: str,
     mm_processor_kwargs: dict[str, object],
@@ -39,18 +31,20 @@ def test_processor_override(
         mm_processor_kwargs=None,
         limit_mm_per_prompt={"image": num_imgs},
     )
-    tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
-    ctx = InputProcessingContext(ctx.model_config, tokenizer)
+    tokenizer = cached_get_tokenizer(ctx.model_config.tokenizer)
+    processor = MULTIMODAL_REGISTRY.create_processor(
+        ctx.model_config,
+        tokenizer=tokenizer,
+    )
 
     # Build the image str / prompt based on the number of images we pass
     prompt = "<|vision_start|><|image_pad|><|vision_end|>" * num_imgs
     mm_data = {"image": [image_assets[0].pil_image] * num_imgs}
 
-    processor = processor_for_qwen2_vl(ctx)
     processed_inputs = processor.apply(prompt, mm_data, mm_processor_kwargs)
 
     # Ensure we have the right number of placeholders per num_crops size
-    hf_processor = processor._get_hf_processor(**mm_processor_kwargs)
+    hf_processor = processor.info.get_hf_processor(**mm_processor_kwargs)
     image_token_id = tokenizer.convert_tokens_to_ids(hf_processor.image_token)
     img_tok_count = processed_inputs["prompt_token_ids"].count(image_token_id)
     pixel_shape = processed_inputs["mm_kwargs"]["pixel_values"].shape