LLaVA support (#1487)

extensions/llava/README.md

@@ -0,0 +1,49 @@
+# LLaVA
+
+## Description
+Adds [LLaVA](https://github.com/haotian-liu/LLaVA) multimodality support to text-generation-webui.
+
+https://user-images.githubusercontent.com/3718215/233817203-69b57e77-0c55-4fd6-b742-3204bb13b8fc.mp4
+
+## Usage
+To run this extension, download LLaVA weights, for example from [here](https://huggingface.co/wojtab/llava-13b-v0-4bit-128g), and then start server.py with `--extensions llava` argument.
+
+When in ui, go to instruct mode, and select LLaVA template, you also should add `"\n###"` to "Custom stopping strings" in parameters tab.
+
+Do note, that each image takes up 258 tokens, so adjust max_new_tokens to be at most 1700 (recommended value is between 200 to 500), so the images don't get truncated.
+
+To send an image, just upload it to the extension field below chat, and send a prompt as always. The image will be added to the end of your message. If you wish to modify the placement, include a string `<image>` in your prompt.
+
+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. From initial testing, it seems as LLaVA considers the features in all images at the same time, so by default the extension skips previous images. If you want to include them anyway, just tick this checkbox.
+
+## Extension config
+This extension uses following parameters (from settings.json):
+|Parameter|Description|
+|---------|-----------|
+|`llava-clip_bits`|Number of bits to load CLIP feature extractor in (either 32 or 16, default=32)|
+|`llava-clip_device`|Torch device to run the extractor on, for example `cpu` or `cuda:0`, by default `cuda:0` if available|
+|`llava-projector_bits`|Number of bits to load CLIP->LLaMA feature projector in (either 32 or 16, default=32)|
+|`llava-projector_bits`|Torch device to run the CLIP->LLaMA feature projector on, for example `cpu` or `cuda:0`, by default `cuda:0` if available|
+|`llava-add_all_images_to_prompt`|Default value of "Embed all images, not only the last one" checkbox|
+
+## Technical description
+
+### Original LLaVA
+The default LLaVA implementation uses modified `transformers` library, however this extension forgoes this requirement. The transformers are modified in LLaVA in such a way, that the entire LLaVA model gets loaded, and the inference now looks as follows:
+```
+images --> CLIP --> projector --> input embeddings for images --> | 
+                                                                  | --> LLaMA
+prompt -------------------------> input embeddings for text ----> |
+```
+The images are represented in the prompt by the following token IDs:
+- 32000 - `<im_patch>` - placeholder token for embeddings from projector
+- 32001 - `<im_start>` - token marking start of an image
+- 32002 - `<im_end>` - token marking end of an image
+
+By default, image will be represented as `<im_start><im_patch>*256<im_end>`. The input embeddings for an image are converted with a single linear layer of the projector, then they are placed instead of `<im_patch>` tokens.
+The concatenated prompt then gets fed to fine-tuned LLaMA.
+
+### In this extension
+
+Using default transformers, they only load the LLaMA part of LLaVA, ignoring the added projector weights, and not loading CLIP. We then reconstruct the `images -> CLIP -> projector` pipeline ourselves, then concatenate the input embeddings, and feed it to LLaMA loaded by transformers. This allows us to use normal flow from webui to load this model, and just hijack the model input with additional features.
+Splitting it to 3 separate models, allows us to configure each of them, and to move them to different devices(for example we can run CLIP+projector on CPU and LLaMA on GPU). Also, it enables us to use 4-bit GPTQ quantization for LLaVA, massively cutting down the VRAM requirement (it should be possible to fit on 12GB of VRAM with full context size by moving CLIP and projector to CPU).

extensions/llava/script.py

@@ -0,0 +1,279 @@
+import base64
+import re
+import time
+from dataclasses import dataclass
+from functools import partial
+from io import BytesIO
+
+import gradio as gr
+import torch
+from huggingface_hub import hf_hub_download
+from PIL import Image
+from transformers import CLIPImageProcessor, CLIPVisionModel
+
+from modules import shared
+from modules.extensions import apply_extensions
+from modules.text_generation import encode, get_max_prompt_length
+
+params = {
+    "add_all_images_to_prompt": False,
+    # device to run CLIP on
+    "clip_device": None,
+    # bits to load clip in either 32 or 16 (it doesn't support 8-bit)
+    "clip_bits": 32,
+    # device to run projector on
+    "projector_device": None,
+    # projector bits, either 32 or 16
+    "projector_bits": 32
+}
+
+
+# If 'state' is True, will hijack the next chat generation
+input_hijack = {
+    'state': False,
+    'value': ["", ""]
+}
+
+
+# initialized in ui, so that params are loaded from settings
+llava_embedder = None
+
+
+@dataclass
+class Token:
+    token: str
+    id: int
+
+
+class LLaVAEmbedder:
+    IM_PATCH = Token("<im_patch>", 32000)
+    IM_START = Token("<im_start>", 32001)
+    IM_END = Token("<im_end>", 32002)
+    CLIP_VIT_HUB_NAME = 'openai/clip-vit-large-patch14'
+    PROJECTOR_HUB_NAME = 'liuhaotian/LLaVA-13b-pretrain-projector-v0'
+    PROJECTOR_FILE = 'LLaVA-13b-pretrain-projector-v0-CC3M-595K-original_caption.bin'
+
+    def __init__(self):
+        self.clip_device = self._get_device("clip_device")
+        self.clip_dtype = self._get_dtype("clip_bits")
+        self.projector_device = self._get_device("projector_device")
+        self.projector_dtype = self._get_dtype("projector_bits")
+        self.image_processor, self.vision_tower, self.mm_projector = self._load_models()
+
+    def _get_device(self, setting_name):
+        if params[setting_name] is None:
+            return torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
+        return torch.device(params[setting_name])
+
+    def _get_dtype(self, setting_name):
+        return torch.float32 if int(params[setting_name]) == 32 else torch.float16
+
+    def _load_models(self):
+        start_ts = time.time()
+
+        print(f"LLaVA - Loading {LLaVAEmbedder.CLIP_VIT_HUB_NAME} as {self.clip_dtype} on {self.clip_device}...")
+        image_processor = CLIPImageProcessor.from_pretrained(LLaVAEmbedder.CLIP_VIT_HUB_NAME, torch_dtype=self.clip_dtype)
+        vision_tower = CLIPVisionModel.from_pretrained(LLaVAEmbedder.CLIP_VIT_HUB_NAME, torch_dtype=self.clip_dtype).to(self.clip_device)
+
+        print(f"LLaVA - Loading {LLaVAEmbedder.PROJECTOR_HUB_NAME} as {self.projector_dtype} on {self.projector_device}...")
+        projector_path = hf_hub_download(LLaVAEmbedder.PROJECTOR_HUB_NAME, LLaVAEmbedder.PROJECTOR_FILE)
+        mm_projector = torch.nn.Linear(1024, 5120)
+        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)
+        mm_projector = mm_projector.to(self.projector_device)
+
+        print(f"LLaVA supporting models loaded, took {time.time() - start_ts:.2f} seconds")
+        return image_processor, vision_tower, mm_projector
+
+    def _update_prompt(self, prompt, images):
+        for _ in images:
+            # replace the image token with the image patch token in the prompt (each occurrence)
+            replace_token = LLaVAEmbedder.IM_PATCH.token * 256
+            replace_token = LLaVAEmbedder.IM_START.token + replace_token + LLaVAEmbedder.IM_END.token
+            prompt = re.sub(r"<image:([A-Za-z0-9+/=]+)>", replace_token, prompt, 1)
+        return prompt
+
+    def _extract_image_features(self, images):
+        images = self.image_processor(images, return_tensors='pt')['pixel_values']
+        images = images.to(self.clip_device, dtype=self.clip_dtype)
+
+        with torch.no_grad():
+            image_forward_outs = self.vision_tower(images, output_hidden_states=True)
+            select_hidden_state_layer = -2
+            select_hidden_state = image_forward_outs.hidden_states[select_hidden_state_layer]
+            image_features = select_hidden_state[:, 1:].to(self.projector_device, dtype=self.projector_dtype)
+            image_features = self.mm_projector(image_features)
+        return image_features
+
+    def forward(self, prompt, images, state):
+        prompt = self._update_prompt(prompt, images)
+        input_ids = encode(prompt, add_bos_token=state['add_bos_token'], truncation_length=get_max_prompt_length(state))[0]
+        input_embeds = shared.model.model.embed_tokens(input_ids).to(self.projector_device)
+
+        if input_ids[0] == LLaVAEmbedder.IM_PATCH.id:
+            # prompt got truncated in the middle of an image, remove the image data
+            im_end = torch.where(input_ids == LLaVAEmbedder.IM_END.id)[0][0]
+            input_ids = input_ids[im_end+1:]
+            input_embeds = input_embeds[im_end+1:]
+            leftover_images = torch.where(input_ids == LLaVAEmbedder.IM_START.id)[0].shape[0]
+            print(f"LLaVA - WARNING: removed {len(images) - leftover_images} image(s) from prompt. The generation might be broken, try decreasing max_new_tokens")
+            images = images[-leftover_images:]
+            if len(images) == 0:
+                return prompt, input_ids, input_embeds, 0
+
+        total_embedded = 0
+        image_features = self._extract_image_features(images).to(self.projector_device)
+        image_start_tokens = torch.where(input_ids == LLaVAEmbedder.IM_START.id)[0]
+
+        if not torch.any(input_ids == LLaVAEmbedder.IM_PATCH.id) or len(image_start_tokens) == 0:
+            # multimodal LLM, but the current prompt is not multimodal/truncated
+            return prompt, input_ids, input_embeds, total_embedded
+
+        cur_image_idx = 0
+        if not params['add_all_images_to_prompt']:
+            image_start_tokens = [image_start_tokens[-1]]
+            cur_image_idx = -1
+
+        for image_start_token_pos in image_start_tokens:
+            cur_image_features = image_features[cur_image_idx]
+            num_patches = cur_image_features.shape[0]
+            input_embeds = torch.cat((input_embeds[:image_start_token_pos+1], cur_image_features, input_embeds[image_start_token_pos + num_patches + 1:]), dim=0)
+            cur_image_idx += 1
+            total_embedded += 1
+
+        return prompt, input_ids, input_embeds, total_embedded
+
+    @staticmethod
+    def len_in_tokens(text):
+        images = re.findall(r"<image:[A-Za-z0-9+/=]+>", text)
+        image_tokens = 0
+        for _ in images:
+            image_tokens += 258
+        return len(encode(re.sub(r"<image:[A-Za-z0-9+/=]+>", '', text))[0]) + image_tokens
+
+
+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)
+    max_hw, min_hw = max(picture.size), min(picture.size)
+    aspect_ratio = max_hw / min_hw
+    shortest_edge = int(max(300 / aspect_ratio, 224))
+    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")
+    img_str = base64.b64encode(buffer.getvalue()).decode('utf-8')
+    visible = f'<img src="data:image/jpeg;base64,{img_str}">'
+    internal = f'<image:{img_str}>'
+
+    if visible_text == '' or visible_text is None:
+        visible_text = text
+
+    if '<image>' in text:
+        text = text.replace('<image>', internal)
+    else:
+        text = text + '\n' + internal
+
+    if '<image>' in visible_text:
+        visible_text = visible_text.replace('<image>', visible)
+    else:
+        visible_text = visible_text + '\n' + visible
+
+    return text, visible_text
+
+
+def fix_picture_after_remove_last(text, visible_text):
+    image = re.search(r'<img src="data:image/jpeg;base64,([A-Za-z0-9+/=]+)">', text)
+    if image is None:
+        return text, visible_text
+    if visible_text is None:
+        visible_text = text
+    text = re.sub(r'<img src="data:image/jpeg;base64,([A-Za-z0-9+/=]+)">', "<image:\\1>", text)
+    return text, visible_text
+
+
+def custom_generate_chat_prompt(user_input, state, **kwargs):
+    impersonate = kwargs['impersonate'] if 'impersonate' in kwargs else False
+    _continue = kwargs['_continue'] if '_continue' in kwargs else False
+    also_return_rows = kwargs['also_return_rows'] if 'also_return_rows' in kwargs else False
+    rows = [f"{state['context'].strip()}\n"]
+    min_rows = 3
+
+    # Finding the maximum prompt size
+    chat_prompt_size = state['chat_prompt_size']
+    if shared.soft_prompt:
+        chat_prompt_size -= shared.soft_prompt_tensor.shape[1]
+    max_length = min(get_max_prompt_length(state), chat_prompt_size)
+
+    prefix1 = f"{state['name1']}: "
+    prefix2 = f"{state['name2']}: "
+
+    i = len(shared.history['internal']) - 1
+    while i >= 0 and LLaVAEmbedder.len_in_tokens(''.join(rows)) < max_length:
+        if _continue and i == len(shared.history['internal']) - 1:
+            rows.insert(1, f"{prefix2}{shared.history['internal'][i][1]}")
+        else:
+            rows.insert(1, f"{prefix2}{shared.history['internal'][i][1].strip()}{state['end_of_turn']}\n")
+
+        string = shared.history['internal'][i][0]
+        if string != '':
+            rows.insert(1, f"{prefix1}{string.strip()}{state['end_of_turn']}\n")
+
+        i -= 1
+
+    if impersonate:
+        min_rows = 2
+        rows.append(f"{prefix1}")
+    elif not _continue:
+        # Adding the user message
+        if len(user_input) > 0:
+            rows.append(f"{prefix1}{user_input}{state['end_of_turn']}\n")
+
+        # Adding the Character prefix
+        rows.append(apply_extensions("bot_prefix", f"{prefix2}"))
+
+    while len(rows) > min_rows and LLaVAEmbedder.len_in_tokens(''.join(rows)) >= max_length:
+        rows.pop(1)
+    prompt = ''.join(rows)
+
+    if also_return_rows:
+        return prompt, rows
+    else:
+        return prompt
+
+
+def tokenizer_modifier(state, prompt, input_ids, input_embeds):
+    global params
+    start_ts = time.time()
+    image_matches = re.finditer(r"<image:([A-Za-z0-9+/=]+)>", prompt)
+    images = [Image.open(BytesIO(base64.b64decode(match.group(1)))) for match in image_matches]
+
+    if len(images) == 0:
+        return prompt, input_ids, input_embeds
+
+    prompt, input_ids, input_embeds, total_embedded = llava_embedder.forward(prompt, images, state)
+    print(f'LLaVA - Embedded {total_embedded} image(s) in {time.time()-start_ts:.2f}s')
+    return prompt, input_ids.unsqueeze(0).to(shared.model.device), input_embeds.unsqueeze(0).to(shared.model.device)
+
+
+def ui():
+    global llava_embedder
+    llava_embedder = LLaVAEmbedder()
+    with gr.Column():
+        picture_select = gr.Image(label='Send a picture', type='pil')
+        # I found that it doesn't deal super well with multiple images, and demo ui had a bug where it included only the last image anyway
+        single_image_checkbox = gr.Checkbox(False, label='Embed all images, not only the last one')
+    # Prepare the input hijack
+    picture_select.upload(
+        lambda picture: input_hijack.update({"state": True, "value": partial(add_chat_picture, picture)}),
+        [picture_select],
+        None
+    )
+    picture_select.clear(lambda: input_hijack.update({"state": False, "value": ["",""]}), None, None)
+    single_image_checkbox.change(lambda x: params.update({"add_all_images_to_prompt": x}), single_image_checkbox, None)
+    shared.gradio['Generate'].click(lambda: None, None, picture_select)
+    shared.gradio['textbox'].submit(lambda: None, None, picture_select)
+    shared.gradio['Remove last'].click(lambda: input_hijack.update({"state": True, "value": fix_picture_after_remove_last}), None, None)