Implement merged processor for llava-next

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

Author: DarkLight1337

tests/multimodal/test_processing.py

@@ -631,6 +631,7 @@ def _test_processing_cache_correctness(
     ("facebook/chameleon-7b", {"image": False}),
     ("adept/fuyu-8b", {"image": False}),
     ("llava-hf/llava-1.5-7b-hf", {"image": True}),
+    ("llava-hf/llava-v1.6-mistral-7b-hf", {"image": True}),
     ("TIGER-Lab/Mantis-8B-siglip-llama3", {"image": True}),
     ("mistral-community/pixtral-12b", {"image": True}),
     ("Qwen/Qwen2-VL-2B-Instruct", {"image": True, "video": True}),

tests/plugins/vllm_add_dummy_model/vllm_add_dummy_model/my_llava.py

@@ -3,13 +3,11 @@
 import torch
 
 from vllm.model_executor.models.llava import (LlavaForConditionalGeneration,
-                                              LlavaMultiModalProcessor,
-                                              get_max_llava_image_tokens)
+                                              LlavaMultiModalProcessor)
 from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.multimodal import MULTIMODAL_REGISTRY
 
 
-@MULTIMODAL_REGISTRY.register_max_image_tokens(get_max_llava_image_tokens)
 @MULTIMODAL_REGISTRY.register_processor(LlavaMultiModalProcessor)
 class MyLlava(LlavaForConditionalGeneration):
 

vllm/model_executor/models/clip.py

@@ -24,6 +24,8 @@
                                    resolve_visual_encoder_outputs)
 from vllm.sequence import SequenceData
 
+from .vision import VisionEncoderInfo
+
 
 def get_clip_patch_grid_length(*, image_size: int, patch_size: int) -> int:
     assert image_size % patch_size == 0
@@ -149,6 +151,29 @@ def input_processor_for_clip(
                         multi_modal_placeholders={"image": ranges})
 
 
+class CLIPEncoderInfo(VisionEncoderInfo[CLIPVisionConfig]):
+
+    def get_num_image_tokens(
+        self,
+        *,
+        image_width: int,
+        image_height: int,
+    ) -> int:
+        return get_clip_image_feature_size(self.vision_config)
+
+    def get_max_image_tokens(self) -> int:
+        return get_max_clip_image_tokens(self.vision_config)
+
+    def get_num_patches(self) -> int:
+        return get_clip_patch_grid_length(
+            image_size=self.vision_config.image_size,
+            patch_size=self.vision_config.patch_size,
+        )
+
+    def get_image_size(self) -> int:
+        return self.vision_config.image_size
+
+
 # Adapted from https://github.com/huggingface/transformers/blob/v4.39.0/src/transformers/models/clip/modeling_clip.py#L164 # noqa
 class CLIPVisionEmbeddings(nn.Module):
 

vllm/model_executor/models/fuyu.py

@@ -76,7 +76,7 @@ def _get_image_target_size(self) -> ImageSize:
         return ImageSize(width=target_size["width"],
                          height=target_size["height"])
 
-    def _get_image_grid_size(
+    def _get_image_feature_grid_size(
         self,
         *,
         image_width: int,
@@ -99,7 +99,7 @@ def _get_image_grid_size(
     def get_mm_max_tokens_per_item(self) -> Mapping[str, int]:
         target_width, target_height = self._get_image_target_size()
 
-        max_ncols, max_nrows = self._get_image_grid_size(
+        max_ncols, max_nrows = self._get_image_feature_grid_size(
             image_width=target_width,
             image_height=target_height,
         )
@@ -172,7 +172,7 @@ def get_replacement_fuyu(item_idx: int):
             images = mm_items.get_items("image", ImageProcessorItems)
             image_size = images.get_image_size(item_idx)
 
-            ncols, nrows = self._get_image_grid_size(
+            ncols, nrows = self._get_image_feature_grid_size(
                 image_width=image_size.width,
                 image_height=image_size.height,
             )