Support LLaVA v1.5 (#4305)

extensions/multimodal/README.md

@@ -4,6 +4,8 @@
 
 Adds support for multimodality (text+images) to text-generation-webui.
 
+Note: multimodal currently only works for transformers, AutoGPTQ, and GPTQ-for-LLaMa loaders. ExLlama (v1 and v2) and llama.cpp support are planned.
+
 https://user-images.githubusercontent.com/3718215/233817203-69b57e77-0c55-4fd6-b742-3204bb13b8fc.mp4
 
 ## Usage
@@ -11,13 +13,15 @@ https://user-images.githubusercontent.com/3718215/233817203-69b57e77-0c55-4fd6-b
 To run this extension, download a LLM that supports multimodality, and then start server.py with the appropriate `--multimodal-pipeline` argument. Examples:
 
 ```
+python server.py --model liuhaotian_llava-v1.5-13b --multimodal-pipeline llava-v1.5-13b --load-in-4bit
+python server.py --model TheBloke_llava-v1.5-13B-GPTQ_gptq-4bit-32g-actorder_True --multimodal-pipeline llava-v1.5-13b --disable_exllama --loader autogptq
 python server.py --model wojtab_llava-7b-v0-4bit-128g --multimodal-pipeline llava-7b
-python3 server.py --model wojtab_llava-13b-v0-4bit-128g --multimodal-pipeline llava-13b
+python server.py --model wojtab_llava-13b-v0-4bit-128g --multimodal-pipeline llava-13b
 python server.py --model anon8231489123_vicuna-13b-GPTQ-4bit-128g --multimodal-pipeline minigpt4-13b
 python server.py --model llama-7b-4bit --multimodal-pipeline minigpt4-7b
 ```
 
-There is built-in support for LLaVA-v0-13B and LLaVA-v0-7b. To install `minigpt4`:
+There is built-in support for LLaVA-v0-13B, LLaVA-v0-7b, and LLaVA-v1.5-13B. To install `minigpt4`:
 
 - clone https://github.com/Wojtab/minigpt-4-pipeline into `extensions/multimodal/pipelines`
 - install the requirements.txt
@@ -31,6 +35,7 @@ To send an image, just upload it to the extension field below chat, and send a p
 Additionally, there is *Embed all images, not only the last one* checkbox. It modifies the image embeddings, by default (if it's unchecked), all but the most recent images have their embeddings empty, so they are not fed to the network. It seems as if some multimodal networks consider the features in all images at the same time as if they were a single image. Due to this behavior, by default, the extension skips previous images. However, it can lead to sub-par generation on other pipelines. If you want to include all images, just tick this checkbox.
 
 ## Compatibility
+
 As of now, the following multimodal pipelines are supported:
 |Pipeline|`--multimodal-pipeline`|Default LLM|LLM info(for the linked model)|Pipeline repository|
 |-|-|-|-|-|

extensions/multimodal/pipelines/llava/llava.py

@@ -13,6 +13,20 @@ from modules.logging_colors import logger
 from modules.text_generation import encode
 
 
+def expand2square(pil_img: Image.Image, background_color: Tuple[int]) -> Image.Image:
+    width, height = pil_img.size
+    if width == height:
+        return pil_img
+    elif width > height:
+        result = Image.new(pil_img.mode, (width, width), background_color)
+        result.paste(pil_img, (0, (width - height) // 2))
+        return result
+    else:
+        result = Image.new(pil_img.mode, (height, height), background_color)
+        result.paste(pil_img, ((height - width) // 2, 0))
+        return result
+
+
 class LLaVA_v0_Pipeline(AbstractMultimodalPipeline):
     CLIP_REPO = "openai/clip-vit-large-patch14"
 
@@ -27,21 +41,33 @@ class LLaVA_v0_Pipeline(AbstractMultimodalPipeline):
     def _load_models(self):
         start_ts = time.time()
 
-        logger.info(f"LLaVA - Loading CLIP from {LLaVA_v0_Pipeline.CLIP_REPO} as {self.clip_dtype} on {self.clip_device}...")
-        image_processor = CLIPImageProcessor.from_pretrained(LLaVA_v0_Pipeline.CLIP_REPO, torch_dtype=self.clip_dtype)
-        vision_tower = CLIPVisionModel.from_pretrained(LLaVA_v0_Pipeline.CLIP_REPO, torch_dtype=self.clip_dtype).to(self.clip_device)
+        logger.info(f"LLaVA - Loading CLIP from {self.CLIP_REPO} as {self.clip_dtype} on {self.clip_device}...")
+        image_processor = CLIPImageProcessor.from_pretrained(self.CLIP_REPO, torch_dtype=self.clip_dtype)
+        vision_tower = CLIPVisionModel.from_pretrained(self.CLIP_REPO, torch_dtype=self.clip_dtype).to(self.clip_device)
 
         logger.info(f"LLaVA - Loading projector from {self.llava_projector_repo()} as {self.projector_dtype} on {self.projector_device}...")
         projector_path = hf_hub_download(self.llava_projector_repo(), self.llava_projector_filename())
-        mm_projector = torch.nn.Linear(*self.llava_projector_shape())
+        mm_projector = self.build_mm_projector()
         projector_data = torch.load(projector_path)
-        mm_projector.weight = torch.nn.Parameter(projector_data['model.mm_projector.weight'].to(dtype=self.projector_dtype), False)
-        mm_projector.bias = torch.nn.Parameter(projector_data['model.mm_projector.bias'].to(dtype=self.projector_dtype), False)
+        projector_data = {k[19:]: v for k, v in projector_data.items() if k.startswith('model.mm_projector.')}
+        mm_projector.load_state_dict(projector_data)
         mm_projector = mm_projector.to(self.projector_device)
 
         logger.info(f"LLaVA supporting models loaded, took {time.time() - start_ts:.2f} seconds")
         return image_processor, vision_tower, mm_projector
 
+    def build_mm_projector(self) -> torch.nn.Module:
+        projector_shape = self.llava_projector_shape()
+        if len(projector_shape) == 2:
+            return torch.nn.Linear(*projector_shape)
+        else:
+            modules = []
+            modules.append(torch.nn.Linear(projector_shape[0], projector_shape[1]))
+            for i in range(2, len(projector_shape)):
+                modules.append(torch.nn.GELU())
+                modules.append(torch.nn.Linear(projector_shape[i-1], projector_shape[i]))
+            return torch.nn.Sequential(*modules)
+
     @staticmethod
     def image_start() -> str:
         return "<im_start>"
@@ -175,3 +201,50 @@ class LLaVA_LLaMA_2_13B_Pipeline(LLaVA_v0_13B_Pipeline):
     @staticmethod
     def placeholder_embeddings() -> torch.Tensor:
         return LLaVA_v0_Pipeline.embed_tokens(encode("<unk>"*256, add_bos_token=False)[0])
+
+
+class LLaVA_v1_5_13B_Pipeline(LLaVA_v0_13B_Pipeline):
+    CLIP_REPO = "openai/clip-vit-large-patch14-336"
+
+    def __init__(self, params: dict) -> None:
+        super().__init__(params)
+
+    @staticmethod
+    def name() -> str:
+        return "llava-v1.5-13b"
+
+    @staticmethod
+    def llava_projector_shape() -> Tuple[int, int]:
+        return (1024, 5120, 5120)
+
+    @staticmethod
+    def placeholder_token_id() -> int:
+        return 0
+
+    @staticmethod
+    def llava_projector_repo() -> str:
+        return "liuhaotian/llava-v1.5-13b"
+
+    @staticmethod
+    def image_start() -> str:
+        return ""
+
+    @staticmethod
+    def image_end() -> str:
+        return ""
+
+    @staticmethod
+    def num_image_embeds() -> int:
+        return 576
+
+    def embed_images(self, images: List[Image.Image]) -> torch.Tensor:
+        # pad it to square first
+        images = [
+            expand2square(image, tuple(int(x*255) for x in self.image_processor.image_mean))
+            for image in images
+        ]
+        return super().embed_images(images)
+
+    @staticmethod
+    def placeholder_embeddings() -> torch.Tensor:
+        return LLaVA_v0_Pipeline.embed_tokens(encode("<unk>"*576, add_bos_token=False)[0])

extensions/multimodal/pipelines/llava/pipelines.py

@@ -2,7 +2,7 @@ from typing import Optional
 
 from extensions.multimodal.abstract_pipeline import AbstractMultimodalPipeline
 
-available_pipelines = ['llava-7b', 'llava-13b', 'llava-llama-2-13b']
+available_pipelines = ['llava-7b', 'llava-13b', 'llava-llama-2-13b', 'llava-v1.5-13b']
 
 
 def get_pipeline(name: str, params: dict) -> Optional[AbstractMultimodalPipeline]:
@@ -15,6 +15,9 @@ def get_pipeline(name: str, params: dict) -> Optional[AbstractMultimodalPipeline
     if name == 'llava-llama-2-13b':
         from .llava import LLaVA_LLaMA_2_13B_Pipeline
         return LLaVA_LLaMA_2_13B_Pipeline(params)
+    if name == 'llava-v1.5-13b':
+        from .llava import LLaVA_v1_5_13B_Pipeline
+        return LLaVA_v1_5_13B_Pipeline(params)
     return None
 
 
@@ -25,10 +28,15 @@ def get_pipeline_from_model_name(model_name: str, params: dict) -> Optional[Abst
         if '13b' in model_name.lower():
             from .llava import LLaVA_LLaMA_2_13B_Pipeline
             return LLaVA_LLaMA_2_13B_Pipeline(params)
-    if '7b' in model_name.lower():
-        from .llava import LLaVA_v0_7B_Pipeline
-        return LLaVA_v0_7B_Pipeline(params)
-    if '13b' in model_name.lower():
-        from .llava import LLaVA_v0_13B_Pipeline
-        return LLaVA_v0_13B_Pipeline(params)
+    elif 'llava-v1.5' in model_name.lower():
+        if '13b' in model_name.lower():
+            from .llava import LLaVA_v1_5_13B_Pipeline
+            return LLaVA_v1_5_13B_Pipeline(params)
+    else:
+        if '7b' in model_name.lower():
+            from .llava import LLaVA_v0_7B_Pipeline
+            return LLaVA_v0_7B_Pipeline(params)
+        if '13b' in model_name.lower():
+            from .llava import LLaVA_v0_13B_Pipeline
+            return LLaVA_v0_13B_Pipeline(params)
     return None

extensions/multimodal/script.py

@@ -46,23 +46,24 @@ def chat_input_modifier(text, visible_text, state):
 
 def add_chat_picture(picture, text, visible_text):
     # resize the image, so that shortest edge is at least 224 (size for CLIP), and at most 300 (to keep history manageable)
+    # Adjusted to 336 for the values here, due to the increased resolution in llava-v1.5
     max_hw, min_hw = max(picture.size), min(picture.size)
     aspect_ratio = max_hw / min_hw
-    shortest_edge = int(max(300 / aspect_ratio, 224))
+    shortest_edge = int(max(336 / aspect_ratio, 336))
     longest_edge = int(shortest_edge * aspect_ratio)
     w = shortest_edge if picture.width < picture.height else longest_edge
     h = shortest_edge if picture.width >= picture.height else longest_edge
     picture = picture.resize((w, h))
 
     buffer = BytesIO()
-    picture.save(buffer, format="JPEG")
+    picture.save(buffer, format="PNG")
     img_str = base64.b64encode(buffer.getvalue()).decode('utf-8')
     image = f'<img src="data:image/jpeg;base64,{img_str}">'
 
     if '<image>' in text:
         text = text.replace('<image>', image)
     else:
-        text = text + '\n' + image
+        text = image + '\n' + text
 
     if visible_text == '' or visible_text is None:
         visible_text = text

instruction-templates/LLaVA-v1.yaml

@@ -0,0 +1,4 @@
+user: "USER:"
+bot: "ASSISTANT:"
+turn_template: "<|user|> <|user-message|>\n<|bot|> <|bot-message|></s>\n"
+context: "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions.\n\n"

models/config.yaml

@@ -39,6 +39,8 @@
 .*llava:
   instruction_template: 'LLaVA'
   custom_stopping_strings: '"\n###"'
+.*llava.*1.5:
+  instruction_template: 'LLaVA-v1'
 .*wizard.*mega:
   instruction_template: 'Wizard-Mega'
   custom_stopping_strings: '"</s>"'