[Doc] Update LLaVA docs (#5437)

Co-authored-by: Roger Wang <ywang@roblox.com>

docs/source/models/vlm.rst

@@ -20,9 +20,9 @@ The following :ref:`engine arguments <engine_args>` are specific to VLMs:
     Currently, the support for vision language models on vLLM has the following limitations:
 
     * Only single image input is supported per text prompt.
-    * Dynamic ``image_input_shape`` is not supported: the input image will be resized to the static ``image_input_shape``. This means model output might not exactly match the HuggingFace implementation.
+    * Dynamic ``image_input_shape`` is not supported: the input image will be resized to the static ``image_input_shape``. This means our LLaVA-NeXT output may not exactly match the huggingface implementation.
 
-    We are continuously improving user & developer experience for VLMs. Please raise an issue on GitHub if you have any feedback or feature requests.
+    We are continuously improving user & developer experience for VLMs. Please `open an issue on GitHub <https://github.com/vllm-project/vllm/issues/new/choose>`_ if you have any feedback or feature requests.
 
 Offline Batched Inference
 -------------------------

vllm/model_executor/models/llava.py

@@ -227,7 +227,7 @@ class LlavaForConditionalGeneration(VisionLanguageModelBase):
         attn_metadata: AttentionMetadata,
         **kwargs: object,
     ) -> SamplerOutput:
-        """Run forward pass for Llava 1.5.
+        """Run forward pass for LLaVA-1.5.
 
         One key thing to understand is the `input_ids` already accounts for the
         positions of the to-be-inserted image embeddings.
@@ -247,22 +247,25 @@ class LlavaForConditionalGeneration(VisionLanguageModelBase):
         This way, the `positions` and `attn_metadata` are consistent
         with the `input_ids`.
 
-        The model takes two types of image inputs:
+        This model has two modes of image inputs:
-        PIXEL_VALUES and IMAGE_FEATURES.
+        `PIXEL_VALUES` and `IMAGE_FEATURES`.
-        The following shows how each maps to huggingface implementation.
-        PIXEL_VALUES:
-        - https://github.com/huggingface/transformers/blob/07bdbeb/src/transformers/models/llava/modeling_llava.py#L353
-        IMAGE_FEATURES:
-        - https://github.com/huggingface/transformers/blob/07bdbeb/src/transformers/models/llava/modeling_llava.py#L430
-        before going through the multi modal projector.
 
         Args:
             input_ids: Flattened (concatenated) input_ids corresponding to a
                 batch.
-            pixel_values: For PIXEL_VALUES, expects a batch with shape
+            pixel_values: The pixels in each input image.
-                [1, 3, 336, 336].
+                Expects a batch with shape `[1, 3, 336, 336]`.
-            image_features: For IMAGE_FEATURES, expects a batch with shape
+                (Only applicable to `PIXEL_VALUES` mode)
-                [1, 576, 1024].
+            image_features: The image features for each input image outputted by
+                the vision tower before passing to the multi-modal projector.
+                Expects a batch with shape `[1, 576, 1024]`.
+                (Only applicable to `IMAGE_FEATURES` mode)
+
+        See also:
+            Each input maps to huggingface implementation, as follows:
+
+            - `pixel_values`: https://github.com/huggingface/transformers/blob/v4.41.1/src/transformers/models/llava/modeling_llava.py#L360
+            - `image_features`: https://github.com/huggingface/transformers/blob/v4.41.1/src/transformers/models/llava/modeling_llava.py#L437
         """
         image_input = self._parse_and_validate_image_input(**kwargs)
 

vllm/model_executor/models/llava_next.py

@@ -108,15 +108,6 @@ def _image_pixel_processor(
 @MULTIMODAL_REGISTRY.register_image_pixel_input(_image_pixel_processor)
 @MULTIMODAL_REGISTRY.register_dummy_data(_get_dummy_image_data)
 class LlavaNextForConditionalGeneration(VisionLanguageModelBase):
-    """
-    Args to `forward()`:
-        input_ids: Flattened (concatenated) input_ids corresponding to a
-            batch.
-        pixel_values: For PIXEL_VALUES, expects a batch with shape
-            [1, num_patches, 3, 336, 336].
-        image_features: For IMAGE_FEATURES, expects a batch with shape
-            [1, num_patches, 1176, 1024].
-    """
 
     def __init__(self,
                  config: LlavaNextConfig,
@@ -355,7 +346,7 @@ class LlavaNextForConditionalGeneration(VisionLanguageModelBase):
         attn_metadata: AttentionMetadata,
         **kwargs: object,
     ) -> SamplerOutput:
-        """Run forward pass for Llava 1.5.
+        """Run forward pass for LlaVA-NeXT.
 
         One key thing to understand is the `input_ids` already accounts for the
         positions of the to-be-inserted image embeddings.
@@ -375,22 +366,19 @@ class LlavaNextForConditionalGeneration(VisionLanguageModelBase):
         This way, the `positions` and `attn_metadata` are consistent
         with the `input_ids`.
 
-        The model takes two types of image inputs:
-        PIXEL_VALUES and IMAGE_FEATURES.
-        The following shows how each maps to huggingface implementation.
-        PIXEL_VALUES:
-        - https://github.com/huggingface/transformers/blob/07bdbeb/src/transformers/models/llava/modeling_llava.py#L353
-        IMAGE_FEATURES:
-        - https://github.com/huggingface/transformers/blob/07bdbeb/src/transformers/models/llava/modeling_llava.py#L430
-        before going through the multi modal projector.
-
         Args:
             input_ids: Flattened (concatenated) input_ids corresponding to a
                 batch.
-            pixel_values: For PIXEL_VALUES, expects a batch with shape
+            pixel_values: The pixels in each grid patch for each input image.
-                [1, 3, 336, 336].
+                Expects a batch with shape `[1, num_patches, 3, 336, 336]`.
-            image_features: For IMAGE_FEATURES, expects a batch with shape
+            image_sizes: The original `(width, height)` for each input image.
-                [1, 576, 1024].
+                Expects a batch with shape `[1, 2]`.
+
+        See also:
+            Each input maps to huggingface implementation, as follows:
+
+            - `pixel_values`: https://github.com/huggingface/transformers/blob/v4.41.1/src/transformers/models/llava_next/modeling_llava_next.py#L690
+            - `image_sizes`: https://github.com/huggingface/transformers/blob/v4.41.1/src/transformers/models/llava_next/modeling_llava_next.py#L691
         """
         image_input = self._parse_and_validate_image_input(**kwargs)