init

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+*.mp4 filter=lfs diff=lfs merge=lfs -text

.gitattributes copy

@@ -0,0 +1,36 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
+*.mp4 filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1 @@
+__pycache__

app.py

@@ -1,7 +1,126 @@
+import os
+import shutil
+import tempfile
+
+import spaces
 import gradio as gr
+import torch
+
+from llava.conversation import Conversation, conv_templates
+from llava.serve.gradio_utils import (Chat, block_css, learn_more_markdown,
+                                      title_markdown)
+
+
+def save_video_to_local(video_path):
+    filename = os.path.join('temp', next(
+        tempfile._get_candidate_names()) + '.mp4')
+    shutil.copyfile(video_path, filename)
+    return filename
+
+
+@spaces.GPU(duration=60)
+def generate(video, textbox_in, first_run, state, state_):
+    flag = 1
+    if not textbox_in:
+        if len(state_.messages) > 0:
+            textbox_in = state_.messages[-1][1]
+            state_.messages.pop(-1)
+            flag = 0
+        else:
+            return "Please enter instruction"
+
+    video = video if video else "none"
+
+    if type(state) is not Conversation:
+        state = conv_templates[conv_mode].copy()
+        state_ = conv_templates[conv_mode].copy()
+
+    first_run = False if len(state.messages) > 0 else True
+
+    text_en_out, state_ = handler.generate(
+        video, textbox_in, first_run=first_run, state=state_)
+    state_.messages[-1] = (state_.roles[1], text_en_out)
+
+    textbox_out = text_en_out
+
+    if flag:
+        state.append_message(state.roles[0], textbox_in)
+    state.append_message(state.roles[1], textbox_out)
+    torch.cuda.empty_cache()
+    return (state, state_, state.to_gradio_chatbot(), False, gr.update(value=None, interactive=True), gr.update(value=video if os.path.exists(video) else None, interactive=True))
+
+
+def clear_history(state, state_):
+    state = conv_templates[conv_mode].copy()
+    state_ = conv_templates[conv_mode].copy()
+    return (gr.update(value=None, interactive=True),
+            gr.update(value=None, interactive=True),
+            True, state, state_, state.to_gradio_chatbot())
+
+
+conv_mode = "llava_llama_3"
+model_path = 'Lin-Chen/sharegpt4video-8b'
+device = 'cuda'
+load_8bit = False
+load_4bit = False
+dtype = torch.float16
+handler = Chat(model_path, conv_mode=conv_mode,
+               load_8bit=load_8bit, load_4bit=load_8bit, device=device)
+
+textbox = gr.Textbox(
+    show_label=False, placeholder="Enter text and press ENTER", container=False
+)
+with gr.Blocks(title='ShareGPT4Video-8B🚀', theme=gr.themes.Default(), css=block_css) as demo:
+    gr.Markdown(title_markdown)
+    state = gr.State()
+    state_ = gr.State()
+    first_run = gr.State()
+
+    with gr.Row():
+        with gr.Column(scale=3):
+            video = gr.Video(label="Input Video")
+
+            cur_dir = os.path.dirname(os.path.abspath(__file__))
+
+        with gr.Column(scale=7):
+            chatbot = gr.Chatbot(label="ShareGPT4Video-8B",
+                                 bubble_full_width=True)
+            with gr.Row():
+                with gr.Column(scale=8):
+                    textbox.render()
+                with gr.Column(scale=1, min_width=50):
+                    submit_btn = gr.Button(
+                        value="Send", variant="primary", interactive=True
+                    )
+            with gr.Row(elem_id="buttons") as button_row:
+                regenerate_btn = gr.Button(
+                    value="🔄  Regenerate", interactive=True)
+                clear_btn = gr.Button(
+                    value="🗑️  Clear history", interactive=True)
+
+    with gr.Row():
+        gr.Examples(
+            examples=[
+                [
+                    f"{cur_dir}/examples/sample_demo_1.mp4",
+                    "Why is this video funny?",
+                ],
+                [
+                    f"{cur_dir}/examples/C_1_0.mp4",
+                    "Write a poem for this video.",
+                ],
+                [
+                    f"{cur_dir}/examples/yoga.mp4",
+                    "What is happening in this video?",
+                ]
+            ],
+            inputs=[video, textbox],
+        )
+    gr.Markdown(learn_more_markdown)
 
-def greet(name):
-    return "Hello " + name + "!!"
+    submit_btn.click(generate, [video, textbox, first_run, state, state_],
+                     [state, state_, chatbot, first_run, textbox, video])
+    clear_btn.click(clear_history, [state, state_],
+                    [video, textbox, first_run, state, state_, chatbot])
 
-demo = gr.Interface(fn=greet, inputs="text", outputs="text")
-demo.launch()
+demo.launch(server_name='0.0.0.0', server_port=23858, share=True)

examples/C_1_0.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5957657865ccb3c8101e82a5a10bbc7e9c33eb957bd53989098a7c0e42512c70
+size 596629

examples/sample_demo_1.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fc6562a172eb9cb3c760a3c9992349c1faa2c793c112b7b9e50bd5cb17c2164d
+size 1549315

examples/yoga.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:74b65d9bec7f83e487b7f923076c01d476dd2ef7ed83928a696ab6f88c7751b7
+size 776184

llava/__init__.py

@@ -0,0 +1 @@
+# from .model import LlavaLlamaForCausalLM

llava/constants.py

@@ -0,0 +1,13 @@
+CONTROLLER_HEART_BEAT_EXPIRATION = 30
+WORKER_HEART_BEAT_INTERVAL = 15
+
+LOGDIR = "."
+
+# Model Constants
+IGNORE_INDEX = -100
+IMAGE_TOKEN_INDEX = -200
+DEFAULT_IMAGE_TOKEN = "<image>"
+DEFAULT_IMAGE_PATCH_TOKEN = "<im_patch>"
+DEFAULT_IM_START_TOKEN = "<im_start>"
+DEFAULT_IM_END_TOKEN = "<im_end>"
+IMAGE_PLACEHOLDER = "<image-placeholder>"

llava/conversation.py

@@ -0,0 +1,428 @@
+import base64
+import dataclasses
+from enum import Enum, auto
+from io import BytesIO
+from typing import Any, List, Tuple
+
+from PIL import Image
+from transformers import AutoTokenizer
+
+
+class SeparatorStyle(Enum):
+    """Different separator style."""
+    SINGLE = auto()
+    TWO = auto()
+    MPT = auto()
+    PLAIN = auto()
+    LLAMA_2 = auto()
+    LLAMA_3 = auto()
+
+
+@dataclasses.dataclass
+class Conversation:
+    """A class that keeps all conversation history."""
+    system: str
+    roles: List[str]
+    messages: List[List[str]]
+    offset: int
+    sep_style: SeparatorStyle = SeparatorStyle.SINGLE
+    sep: str = "###"
+    sep2: str = None
+    version: str = "Unknown"
+
+    skip_next: bool = False
+
+    def get_prompt(self):
+        messages = self.messages
+        if len(messages) > 0 and type(messages[0][1]) is tuple:
+            messages = self.messages.copy()
+            init_role, init_msg = messages[0].copy()
+            init_msg = init_msg[0].replace("<image>", "").strip()
+            if 'mmtag' in self.version:
+                messages[0] = (init_role, init_msg)
+                messages.insert(0, (self.roles[0], "<Image><image></Image>"))
+                messages.insert(1, (self.roles[1], "Received."))
+            else:
+                messages[0] = (init_role, "<image>\n" + init_msg)
+
+        if self.sep_style == SeparatorStyle.SINGLE:
+            ret = self.system + self.sep
+            for role, message in messages:
+                if message:
+                    if type(message) is tuple:
+                        message, _, _ = message
+                    ret += role + ": " + message + self.sep
+                else:
+                    ret += role + ":"
+        elif self.sep_style == SeparatorStyle.TWO:
+            seps = [self.sep, self.sep2]
+            ret = self.system + seps[0]
+            for i, (role, message) in enumerate(messages):
+                if message:
+                    if type(message) is tuple:
+                        message, _, _ = message
+                    ret += role + ": " + message + seps[i % 2]
+                else:
+                    ret += role + ":"
+        elif self.sep_style == SeparatorStyle.MPT:
+            ret = self.system + self.sep
+            for role, message in messages:
+                if message:
+                    if type(message) is tuple:
+                        message, _, _ = message
+                    ret += role + message + self.sep
+                else:
+                    ret += role
+        elif self.sep_style == SeparatorStyle.LLAMA_2:
+            def wrap_sys(
+                msg): return f"<<SYS>>\n{msg}\n<</SYS>>\n\n" if len(msg) > 0 else msg
+
+            def wrap_inst(msg): return f"[INST] {msg} [/INST]"
+            ret = ""
+
+            for i, (role, message) in enumerate(messages):
+                if i == 0:
+                    assert message, "first message should not be none"
+                    assert role == self.roles[0], "first message should come from user"
+                if message:
+                    if type(message) is tuple:
+                        message, _, _ = message
+                    if i == 0:
+                        message = wrap_sys(self.system) + message
+                    if i % 2 == 0:
+                        message = wrap_inst(message)
+                        ret += self.sep + message
+                    else:
+                        ret += " " + message + " " + self.sep2
+                else:
+                    ret += ""
+            ret = ret.lstrip(self.sep)
+        elif self.sep_style == SeparatorStyle.PLAIN:
+            seps = [self.sep, self.sep2]
+            ret = self.system
+            for i, (role, message) in enumerate(messages):
+                if message:
+                    if type(message) is tuple:
+                        message, _, _ = message
+                    ret += message + seps[i % 2]
+                else:
+                    ret += ""
+        else:
+            raise ValueError(f"Invalid style: {self.sep_style}")
+
+        return ret
+
+    def append_message(self, role, message):
+        self.messages.append([role, message])
+
+    def process_image(self, image, image_process_mode, return_pil=False, image_format='PNG', max_len=1344, min_len=672):
+        if image_process_mode == "Pad":
+            def expand2square(pil_img, background_color=(122, 116, 104)):
+                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
+            image = expand2square(image)
+        elif image_process_mode in ["Default", "Crop"]:
+            pass
+        elif image_process_mode == "Resize":
+            image = image.resize((336, 336))
+        else:
+            raise ValueError(
+                f"Invalid image_process_mode: {image_process_mode}")
+        if max(image.size) > max_len:
+            max_hw, min_hw = max(image.size), min(image.size)
+            aspect_ratio = max_hw / min_hw
+            shortest_edge = int(min(max_len / aspect_ratio, min_len, min_hw))
+            longest_edge = int(shortest_edge * aspect_ratio)
+            W, H = image.size
+            if H > W:
+                H, W = longest_edge, shortest_edge
+            else:
+                H, W = shortest_edge, longest_edge
+            image = image.resize((W, H))
+        if return_pil:
+            return image
+        else:
+            buffered = BytesIO()
+            image.save(buffered, format=image_format)
+            img_b64_str = base64.b64encode(buffered.getvalue()).decode()
+            return img_b64_str
+
+    def get_images(self, return_pil=False):
+        images = []
+        for i, (role, msg) in enumerate(self.messages[self.offset:]):
+            if i % 2 == 0:
+                if type(msg) is tuple:
+                    msg, image, image_process_mode = msg
+                    image = self.process_image(
+                        image, image_process_mode, return_pil=return_pil)
+                    images.append(image)
+        return images
+
+    def to_gradio_chatbot(self):
+        ret = []
+        for i, (role, msg) in enumerate(self.messages[self.offset:]):
+            if i % 2 == 0:
+                if type(msg) is tuple:
+                    msg, image, image_process_mode = msg
+                    img_b64_str = self.process_image(
+                        image, "Default", return_pil=False,
+                        image_format='JPEG')
+                    img_str = f'<img src="data:image/jpeg;base64,{img_b64_str}" alt="user upload image" />'
+                    msg = img_str + msg.replace('<image>', '').strip()
+                    ret.append([msg, None])
+                else:
+                    ret.append([msg, None])
+            else:
+                ret[-1][-1] = msg
+        return ret
+
+    def copy(self):
+        return Conversation(
+            system=self.system,
+            roles=self.roles,
+            messages=[[x, y] for x, y in self.messages],
+            offset=self.offset,
+            sep_style=self.sep_style,
+            sep=self.sep,
+            sep2=self.sep2,
+            version=self.version)
+
+    def dict(self):
+        if len(self.get_images()) > 0:
+            return {
+                "system": self.system,
+                "roles": self.roles,
+                "messages": [[x, y[0] if type(y) is tuple else y] for x, y in self.messages],
+                "offset": self.offset,
+                "sep": self.sep,
+                "sep2": self.sep2,
+            }
+        return {
+            "system": self.system,
+            "roles": self.roles,
+            "messages": self.messages,
+            "offset": self.offset,
+            "sep": self.sep,
+            "sep2": self.sep2,
+        }
+
+
+conv_vicuna_v0 = Conversation(
+    system="A chat between a curious human and an artificial intelligence assistant. "
+           "The assistant gives helpful, detailed, and polite answers to the human's questions.",
+    roles=("Human", "Assistant"),
+    messages=(
+        ("Human", "What are the key differences between renewable and non-renewable energy sources?"),
+        ("Assistant",
+            "Renewable energy sources are those that can be replenished naturally in a relatively "
+            "short amount of time, such as solar, wind, hydro, geothermal, and biomass. "
+            "Non-renewable energy sources, on the other hand, are finite and will eventually be "
+            "depleted, such as coal, oil, and natural gas. Here are some key differences between "
+            "renewable and non-renewable energy sources:\n"
+            "1. Availability: Renewable energy sources are virtually inexhaustible, while non-renewable "
+            "energy sources are finite and will eventually run out.\n"
+            "2. Environmental impact: Renewable energy sources have a much lower environmental impact "
+            "than non-renewable sources, which can lead to air and water pollution, greenhouse gas emissions, "
+            "and other negative effects.\n"
+            "3. Cost: Renewable energy sources can be more expensive to initially set up, but they typically "
+            "have lower operational costs than non-renewable sources.\n"
+            "4. Reliability: Renewable energy sources are often more reliable and can be used in more remote "
+            "locations than non-renewable sources.\n"
+            "5. Flexibility: Renewable energy sources are often more flexible and can be adapted to different "
+            "situations and needs, while non-renewable sources are more rigid and inflexible.\n"
+            "6. Sustainability: Renewable energy sources are more sustainable over the long term, while "
+            "non-renewable sources are not, and their depletion can lead to economic and social instability.\n")
+    ),
+    offset=2,
+    sep_style=SeparatorStyle.SINGLE,
+    sep="###",
+)
+
+conv_vicuna_v1 = Conversation(
+    system="A chat between a curious user and an artificial intelligence assistant. "
+    "The assistant gives helpful, detailed, and polite answers to the user's questions.",
+    roles=("USER", "ASSISTANT"),
+    version="v1",
+    messages=(),
+    offset=0,
+    sep_style=SeparatorStyle.TWO,
+    sep=" ",
+    sep2="</s>",
+)
+
+conv_llama_2 = Conversation(
+    system="""You are a helpful, respectful and honest assistant. Always answer as helpfully as possible, while being safe.  Your answers should not include any harmful, unethical, racist, sexist, toxic, dangerous, or illegal content. Please ensure that your responses are socially unbiased and positive in nature.
+
+If a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. If you don't know the answer to a question, please don't share false information.""",
+    roles=("USER", "ASSISTANT"),
+    version="llama_v2",
+    messages=(),
+    offset=0,
+    sep_style=SeparatorStyle.LLAMA_2,
+    sep="<s>",
+    sep2="</s>",
+)
+
+conv_llava_llama_2 = Conversation(
+    system="You are a helpful language and vision assistant. "
+           "You are able to understand the visual content that the user provides, "
+           "and assist the user with a variety of tasks using natural language.",
+    roles=("USER", "ASSISTANT"),
+    version="llama_v2",
+    messages=(),
+    offset=0,
+    sep_style=SeparatorStyle.LLAMA_2,
+    sep="<s>",
+    sep2="</s>",
+)
+
+conv_mpt = Conversation(
+    system="""<|im_start|>system
+A conversation between a user and an LLM-based AI assistant. The assistant gives helpful and honest answers.""",
+    roles=("<|im_start|>user\n", "<|im_start|>assistant\n"),
+    version="mpt",
+    messages=(),
+    offset=0,
+    sep_style=SeparatorStyle.MPT,
+    sep="<|im_end|>",
+)
+
+conv_llava_plain = Conversation(
+    system="",
+    roles=("", ""),
+    messages=(
+    ),
+    offset=0,
+    sep_style=SeparatorStyle.PLAIN,
+    sep="\n",
+)
+
+conv_llava_v0 = Conversation(
+    system="A chat between a curious human and an artificial intelligence assistant. "
+           "The assistant gives helpful, detailed, and polite answers to the human's questions.",
+    roles=("Human", "Assistant"),
+    messages=(
+    ),
+    offset=0,
+    sep_style=SeparatorStyle.SINGLE,
+    sep="###",
+)
+
+conv_llava_v0_mmtag = Conversation(
+    system="A chat between a curious user and an artificial intelligence assistant. "
+           "The assistant is able to understand the visual content that the user provides, and assist the user with a variety of tasks using natural language."
+           "The visual content will be provided with the following format: <Image>visual content</Image>.",
+    roles=("Human", "Assistant"),
+    messages=(
+    ),
+    offset=0,
+    sep_style=SeparatorStyle.SINGLE,
+    sep="###",
+    version="v0_mmtag",
+)
+
+conv_llava_v1 = Conversation(
+    system="A chat between a curious human and an artificial intelligence assistant. "
+           "The assistant gives helpful, detailed, and polite answers to the human's questions.",
+    roles=("USER", "ASSISTANT"),
+    version="v1",
+    messages=(),
+    offset=0,
+    sep_style=SeparatorStyle.TWO,
+    sep=" ",
+    sep2="</s>",
+)
+
+conv_llava_v1_mmtag = Conversation(
+    system="A chat between a curious user and an artificial intelligence assistant. "
+           "The assistant is able to understand the visual content that the user provides, and assist the user with a variety of tasks using natural language."
+           "The visual content will be provided with the following format: <Image>visual content</Image>.",
+    roles=("USER", "ASSISTANT"),
+    messages=(),
+    offset=0,
+    sep_style=SeparatorStyle.TWO,
+    sep=" ",
+    sep2="</s>",
+    version="v1_mmtag",
+)
+
+conv_mistral_instruct = Conversation(
+    system="",
+    roles=("USER", "ASSISTANT"),
+    version="llama_v2",
+    messages=(),
+    offset=0,
+    sep_style=SeparatorStyle.LLAMA_2,
+    sep="",
+    sep2="</s>",
+)
+
+conv_chatml_direct = Conversation(
+    system="""<|im_start|>system
+Answer the questions.""",
+    roles=("<|im_start|>user\n", "<|im_start|>assistant\n"),
+    version="mpt",
+    messages=(),
+    offset=0,
+    sep_style=SeparatorStyle.MPT,
+    sep="<|im_end|>",
+)
+
+conv_yi = Conversation(
+    system="""<|im_start|>system\nAnswer the questions.""",
+    roles=("<|im_start|>user\n", "<|im_start|>assistant\n"),
+    version="yi",
+    messages=(),
+    offset=0,
+    sep_style=SeparatorStyle.MPT,
+    sep="<|im_end|>\n",
+)
+
+conv_llava_llama_3 = Conversation(
+    system="<|begin_of_text|><|start_header_id|>system<|end_header_id|>\n\nYou are a helpful language and vision assistant. You are able to understand the visual content that the user provides, and assist the user with a variety of tasks using natural language.",
+    roles=("<|start_header_id|>user<|end_header_id|>\n\n",
+           "<|start_header_id|>assistant<|end_header_id|>\n\n"),
+    version="llama3",
+    messages=[],
+    offset=0,
+    sep_style=SeparatorStyle.MPT,
+    sep="<|eot_id|>",
+)
+
+default_conversation = conv_vicuna_v1
+conv_templates = {
+    "default": conv_vicuna_v0,
+    "v0": conv_vicuna_v0,
+    "v1": conv_vicuna_v1,
+    "vicuna_v1": conv_vicuna_v1,
+    "llama_2": conv_llama_2,
+    "mistral_instruct": conv_mistral_instruct,
+    "chatml_direct": conv_chatml_direct,
+    "mistral_direct": conv_chatml_direct,
+
+    "plain": conv_llava_plain,
+    "v0_plain": conv_llava_plain,
+    "llava_v0": conv_llava_v0,
+    "v0_mmtag": conv_llava_v0_mmtag,
+    "llava_v1": conv_llava_v1,
+    "v1_mmtag": conv_llava_v1_mmtag,
+    "llava_llama_2": conv_llava_llama_2,
+    "llava_llama_3": conv_llava_llama_3,
+
+    "mpt": conv_mpt,
+}
+
+
+if __name__ == "__main__":
+    print(default_conversation.get_prompt())

llava/mm_utils.py

@@ -0,0 +1,250 @@
+from PIL import Image
+from io import BytesIO
+import base64
+import torch
+import math
+import ast
+
+from transformers import StoppingCriteria
+from llava.constants import IMAGE_TOKEN_INDEX
+
+
+def select_best_resolution(original_size, possible_resolutions):
+    """
+    Selects the best resolution from a list of possible resolutions based on the original size.
+
+    Args:
+        original_size (tuple): The original size of the image in the format (width, height).
+        possible_resolutions (list): A list of possible resolutions in the format [(width1, height1), (width2, height2), ...].
+
+    Returns:
+        tuple: The best fit resolution in the format (width, height).
+    """
+    original_width, original_height = original_size
+    best_fit = None
+    max_effective_resolution = 0
+    min_wasted_resolution = float('inf')
+
+    for width, height in possible_resolutions:
+        scale = min(width / original_width, height / original_height)
+        downscaled_width, downscaled_height = int(original_width * scale), int(original_height * scale)
+        effective_resolution = min(downscaled_width * downscaled_height, original_width * original_height)
+        wasted_resolution = (width * height) - effective_resolution
+
+        if effective_resolution > max_effective_resolution or (effective_resolution == max_effective_resolution and wasted_resolution < min_wasted_resolution):
+            max_effective_resolution = effective_resolution
+            min_wasted_resolution = wasted_resolution
+            best_fit = (width, height)
+
+    return best_fit
+
+
+def resize_and_pad_image(image, target_resolution):
+    """
+    Resize and pad an image to a target resolution while maintaining aspect ratio.
+
+    Args:
+        image (PIL.Image.Image): The input image.
+        target_resolution (tuple): The target resolution (width, height) of the image.
+
+    Returns:
+        PIL.Image.Image: The resized and padded image.
+    """
+    original_width, original_height = image.size
+    target_width, target_height = target_resolution
+
+    scale_w = target_width / original_width
+    scale_h = target_height / original_height
+
+    if scale_w < scale_h:
+        new_width = target_width
+        new_height = min(math.ceil(original_height * scale_w), target_height)
+    else:
+        new_height = target_height
+        new_width = min(math.ceil(original_width * scale_h), target_width)
+
+    # Resize the image
+    resized_image = image.resize((new_width, new_height))
+
+    new_image = Image.new('RGB', (target_width, target_height), (0, 0, 0))
+    paste_x = (target_width - new_width) // 2
+    paste_y = (target_height - new_height) // 2
+    new_image.paste(resized_image, (paste_x, paste_y))
+
+    return new_image
+
+
+def divide_to_patches(image, patch_size):
+    """
+    Divides an image into patches of a specified size.
+
+    Args:
+        image (PIL.Image.Image): The input image.
+        patch_size (int): The size of each patch.
+
+    Returns:
+        list: A list of PIL.Image.Image objects representing the patches.
+    """
+    patches = []
+    width, height = image.size
+    for i in range(0, height, patch_size):
+        for j in range(0, width, patch_size):
+            box = (j, i, j + patch_size, i + patch_size)
+            patch = image.crop(box)
+            patches.append(patch)
+
+    return patches
+
+
+def get_anyres_image_grid_shape(image_size, grid_pinpoints, patch_size):
+    """
+    Calculate the shape of the image patch grid after the preprocessing for images of any resolution.
+
+    Args:
+        image_size (tuple): The size of the input image in the format (width, height).
+        grid_pinpoints (str): A string representation of a list of possible resolutions.
+        patch_size (int): The size of each image patch.
+
+    Returns:
+        tuple: The shape of the image patch grid in the format (width, height).
+    """
+    if type(grid_pinpoints) is list:
+        possible_resolutions = grid_pinpoints
+    else:
+        possible_resolutions = ast.literal_eval(grid_pinpoints)
+    width, height = select_best_resolution(image_size, possible_resolutions)
+    return width // patch_size, height // patch_size
+
+
+def process_anyres_image(image, processor, grid_pinpoints):
+    """
+    Process an image with variable resolutions.
+
+    Args:
+        image (PIL.Image.Image): The input image to be processed.
+        processor: The image processor object.
+        grid_pinpoints (str): A string representation of a list of possible resolutions.
+
+    Returns:
+        torch.Tensor: A tensor containing the processed image patches.
+    """
+    if type(grid_pinpoints) is list:
+        possible_resolutions = grid_pinpoints
+    else:
+        possible_resolutions = ast.literal_eval(grid_pinpoints)
+    best_resolution = select_best_resolution(image.size, possible_resolutions)
+    image_padded = resize_and_pad_image(image, best_resolution)
+
+    patches = divide_to_patches(image_padded, processor.crop_size['height'])
+
+    shortest_edge = processor.size['shortest_edge'] if isinstance(
+        processor.size, dict) else min(processor.size[0], processor.size[1])
+    image_original_resize = image.resize(
+        (shortest_edge, shortest_edge))
+
+    image_patches = [image_original_resize] + patches
+    image_patches = [processor.preprocess(image_patch, return_tensors='pt')['pixel_values'][0]
+                     for image_patch in image_patches]
+    return torch.stack(image_patches, dim=0)
+
+
+def load_image_from_base64(image):
+    return Image.open(BytesIO(base64.b64decode(image)))
+
+
+def expand2square(pil_img, background_color):
+    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
+
+
+def process_images(images, image_processor, model_cfg):
+    image_aspect_ratio = getattr(model_cfg, "image_aspect_ratio", None)
+    new_images = []
+    if image_aspect_ratio == 'pad':
+        for image in images:
+            image = expand2square(image, tuple(int(x*255) for x in image_processor.image_mean))
+            image = image_processor.preprocess(image, return_tensors='pt')['pixel_values'][0]
+            new_images.append(image)
+    elif image_aspect_ratio == "anyres":
+        for image in images:
+            image = process_anyres_image(image, image_processor, model_cfg.image_grid_pinpoints)
+            new_images.append(image)
+    else:
+        return image_processor(images, return_tensors='pt')['pixel_values']
+    if all(x.shape == new_images[0].shape for x in new_images):
+        new_images = torch.stack(new_images, dim=0)
+    return new_images
+
+
+def tokenizer_image_token(prompt, tokenizer, image_token_index=IMAGE_TOKEN_INDEX, return_tensors=None):
+    prompt_chunks = [tokenizer(chunk).input_ids for chunk in prompt.split('<image>')]
+
+    def insert_separator(X, sep):
+        return [ele for sublist in zip(X, [sep]*len(X)) for ele in sublist][:-1]
+
+    input_ids = []
+    offset = 0
+    if len(prompt_chunks) > 0 and len(prompt_chunks[0]) > 0 and prompt_chunks[0][0] == tokenizer.bos_token_id:
+        offset = 1
+        input_ids.append(prompt_chunks[0][0])
+
+    for x in insert_separator(prompt_chunks, [image_token_index] * (offset + 1)):
+        input_ids.extend(x[offset:])
+
+    if return_tensors is not None:
+        if return_tensors == 'pt':
+            return torch.tensor(input_ids, dtype=torch.long)
+        raise ValueError(f'Unsupported tensor type: {return_tensors}')
+    return input_ids
+
+
+def get_model_name_from_path(model_path):
+    model_path = model_path.strip("/")
+    model_paths = model_path.split("/")
+    if model_paths[-1].startswith('checkpoint-'):
+        return model_paths[-2] + "_" + model_paths[-1]
+    else:
+        return model_paths[-1]
+
+class KeywordsStoppingCriteria(StoppingCriteria):
+    def __init__(self, keywords, tokenizer, input_ids):
+        self.keywords = keywords
+        self.keyword_ids = []
+        self.max_keyword_len = 0
+        for keyword in keywords:
+            cur_keyword_ids = tokenizer(keyword).input_ids
+            if len(cur_keyword_ids) > 1 and cur_keyword_ids[0] == tokenizer.bos_token_id:
+                cur_keyword_ids = cur_keyword_ids[1:]
+            if len(cur_keyword_ids) > self.max_keyword_len:
+                self.max_keyword_len = len(cur_keyword_ids)
+            self.keyword_ids.append(torch.tensor(cur_keyword_ids))
+        self.tokenizer = tokenizer
+        self.start_len = input_ids.shape[1]
+    
+    def call_for_batch(self, output_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
+        offset = min(output_ids.shape[1] - self.start_len, self.max_keyword_len)
+        self.keyword_ids = [keyword_id.to(output_ids.device) for keyword_id in self.keyword_ids]
+        for keyword_id in self.keyword_ids:
+            truncated_output_ids = output_ids[0, -keyword_id.shape[0]:]
+            if torch.equal(truncated_output_ids, keyword_id):
+                return True
+        outputs = self.tokenizer.batch_decode(output_ids[:, -offset:], skip_special_tokens=True)[0]
+        for keyword in self.keywords:
+            if keyword in outputs:
+                return True
+        return False
+    
+    def __call__(self, output_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
+        outputs = []
+        for i in range(output_ids.shape[0]):
+            outputs.append(self.call_for_batch(output_ids[i].unsqueeze(0), scores))
+        return all(outputs)

llava/model/__init__.py

@@ -0,0 +1,6 @@
+try:
+    from .language_model.llava_llama import LlavaLlamaForCausalLM, LlavaConfig
+    from .language_model.llava_mpt import LlavaMptForCausalLM, LlavaMptConfig
+    from .language_model.llava_mistral import LlavaMistralForCausalLM, LlavaMistralConfig
+except:
+    pass

llava/model/apply_delta.py

@@ -0,0 +1,48 @@
+"""
+Usage:
+python3 -m fastchat.model.apply_delta --base ~/model_weights/llama-7b --target ~/model_weights/vicuna-7b --delta lmsys/vicuna-7b-delta
+"""
+import argparse
+
+import torch
+from tqdm import tqdm
+from transformers import AutoTokenizer, AutoModelForCausalLM
+from llava import LlavaLlamaForCausalLM
+
+
+def apply_delta(base_model_path, target_model_path, delta_path):
+    print("Loading base model")
+    base = AutoModelForCausalLM.from_pretrained(
+        base_model_path, torch_dtype=torch.float16, low_cpu_mem_usage=True)
+
+    print("Loading delta")
+    delta = LlavaLlamaForCausalLM.from_pretrained(delta_path, torch_dtype=torch.float16, low_cpu_mem_usage=True)
+    delta_tokenizer = AutoTokenizer.from_pretrained(delta_path)
+
+    print("Applying delta")
+    for name, param in tqdm(delta.state_dict().items(), desc="Applying delta"):
+        if name not in base.state_dict():
+            assert name in ['model.mm_projector.weight', 'model.mm_projector.bias'], f'{name} not in base model'
+            continue
+        if param.data.shape == base.state_dict()[name].shape:
+            param.data += base.state_dict()[name]
+        else:
+            assert name in ['model.embed_tokens.weight', 'lm_head.weight'], \
+                f'{name} dimension mismatch: {param.data.shape} vs {base.state_dict()[name].shape}'
+            bparam = base.state_dict()[name]
+            param.data[:bparam.shape[0], :bparam.shape[1]] += bparam
+
+    print("Saving target model")
+    delta.save_pretrained(target_model_path)
+    delta_tokenizer.save_pretrained(target_model_path)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--base-model-path", type=str, required=True)
+    parser.add_argument("--target-model-path", type=str, required=True)
+    parser.add_argument("--delta-path", type=str, required=True)
+
+    args = parser.parse_args()
+
+    apply_delta(args.base_model_path, args.target_model_path, args.delta_path)

llava/model/builder.py

@@ -0,0 +1,204 @@
+#    Copyright 2023 Haotian Liu
+#
+#    Licensed under the Apache License, Version 2.0 (the "License");
+#    you may not use this file except in compliance with the License.
+#    You may obtain a copy of the License at
+#
+#        http://www.apache.org/licenses/LICENSE-2.0
+#
+#    Unless required by applicable law or agreed to in writing, software
+#    distributed under the License is distributed on an "AS IS" BASIS,
+#    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#    See the License for the specific language governing permissions and
+#    limitations under the License.
+
+
+import os
+import shutil
+import warnings
+
+import torch
+from transformers import (AutoConfig, AutoModelForCausalLM, AutoTokenizer,
+                          BitsAndBytesConfig)
+
+from llava.constants import (DEFAULT_IM_END_TOKEN, DEFAULT_IM_START_TOKEN,
+                             DEFAULT_IMAGE_PATCH_TOKEN)
+from llava.model import *
+from llava.train.train import smart_tokenizer_and_embedding_resize
+
+
+def load_pretrained_model(model_path, model_base, model_name, load_8bit=False, load_4bit=False, device_map="auto", device="cuda", use_flash_attn=False, lora_alpha=None, **kwargs):
+    kwargs = {"device_map": device_map, **kwargs}
+
+    if device != "cuda":
+        kwargs['device_map'] = {"": device}
+
+    if load_8bit:
+        kwargs['load_in_8bit'] = True
+    elif load_4bit:
+        kwargs['load_in_4bit'] = True
+        kwargs['quantization_config'] = BitsAndBytesConfig(
+            load_in_4bit=True,
+            bnb_4bit_compute_dtype=torch.float16,
+            bnb_4bit_use_double_quant=True,
+            bnb_4bit_quant_type='nf4'
+        )
+    else:
+        kwargs['torch_dtype'] = torch.float16
+
+    if use_flash_attn:
+        kwargs['attn_implementation'] = 'flash_attention_2'
+
+    if 'llava' or 'sharegpt4video' in model_name.lower():
+        # Load LLaVA model
+        if 'lora' in model_name.lower() and model_base is None:
+            warnings.warn('There is `lora` in model name but no `model_base` is provided. If you are loading a LoRA model, please provide the `model_base` argument. Detailed instruction: https://github.com/haotian-liu/LLaVA#launch-a-model-worker-lora-weights-unmerged.')
+        if 'lora' in model_name.lower() and model_base is not None:
+            from llava.model.language_model.llava_llama import LlavaConfig
+            lora_cfg_pretrained = LlavaConfig.from_pretrained(model_path)
+            tokenizer = AutoTokenizer.from_pretrained(
+                model_base, use_fast=False, model_max_length=lora_cfg_pretrained.tokenizer_model_max_length)
+            print('Loading LLaVA from base model...')
+            model = LlavaLlamaForCausalLM.from_pretrained(
+                model_base, low_cpu_mem_usage=True, config=lora_cfg_pretrained, **kwargs)
+            token_num, tokem_dim = model.lm_head.out_features, model.lm_head.in_features
+            if model.lm_head.weight.shape[0] != token_num:
+                model.lm_head.weight = torch.nn.Parameter(torch.empty(
+                    token_num, tokem_dim, device=model.device, dtype=model.dtype))
+                model.model.embed_tokens.weight = torch.nn.Parameter(torch.empty(
+                    token_num, tokem_dim, device=model.device, dtype=model.dtype))
+            print('Loading additional LLaVA weights...')
+            if os.path.exists(os.path.join(model_path, 'non_lora_trainables.bin')):
+                non_lora_trainables = torch.load(os.path.join(
+                    model_path, 'non_lora_trainables.bin'), map_location='cpu')
+            else:
+                # this is probably from HF Hub
+                from huggingface_hub import hf_hub_download
+
+                def load_from_hf(repo_id, filename, subfolder=None):
+                    cache_file = hf_hub_download(
+                        repo_id=repo_id,
+                        filename=filename,
+                        subfolder=subfolder)
+                    return torch.load(cache_file, map_location='cpu')
+                non_lora_trainables = load_from_hf(
+                    model_path, 'non_lora_trainables.bin')
+            non_lora_trainables = {(k[11:] if k.startswith(
+                'base_model.') else k): v for k, v in non_lora_trainables.items()}
+            if any(k.startswith('model.model.') for k in non_lora_trainables):
+                non_lora_trainables = {(k[6:] if k.startswith(
+                    'model.') else k): v for k, v in non_lora_trainables.items()}
+            model.load_state_dict(non_lora_trainables, strict=False)
+
+            from peft import PeftModel
+            print('Loading LoRA weights...')
+            if lora_alpha is not None:
+                print("Lora Scaling:", lora_alpha/128)
+                model = PeftModel.from_pretrained(
+                    model, model_path, lora_alpha=lora_alpha, torch_device='cpu')
+            else:
+                model = PeftModel.from_pretrained(model, model_path, torch_device='cpu')
+            print('Merging LoRA weights...')
+            model = model.merge_and_unload()
+            print('Model is loaded...')
+        elif model_base is not None:
+            # this may be mm projector only
+            print('Loading LLaVA from base model...')
+            if 'mpt' in model_name.lower():
+                if not os.path.isfile(os.path.join(model_path, 'configuration_mpt.py')):
+                    shutil.copyfile(os.path.join(model_base, 'configuration_mpt.py'), os.path.join(
+                        model_path, 'configuration_mpt.py'))
+                tokenizer = AutoTokenizer.from_pretrained(
+                    model_base, use_fast=True)
+                cfg_pretrained = AutoConfig.from_pretrained(
+                    model_path, trust_remote_code=True)
+                model = LlavaMptForCausalLM.from_pretrained(
+                    model_base, low_cpu_mem_usage=True, config=cfg_pretrained, **kwargs)
+            else:
+                tokenizer = AutoTokenizer.from_pretrained(
+                    model_base, use_fast=False)
+                cfg_pretrained = AutoConfig.from_pretrained(model_path)
+                model = LlavaLlamaForCausalLM.from_pretrained(
+                    model_base, low_cpu_mem_usage=True, config=cfg_pretrained, **kwargs)
+
+            mm_projector_weights = torch.load(os.path.join(
+                model_path, 'mm_projector.bin'), map_location='cpu')
+            mm_projector_weights = {k: v.to(torch.float16)
+                                    for k, v in mm_projector_weights.items()}
+            model.load_state_dict(mm_projector_weights, strict=False)
+        else:
+            if 'mpt' in model_name.lower():
+                tokenizer = AutoTokenizer.from_pretrained(
+                    model_path, use_fast=True)
+                model = LlavaMptForCausalLM.from_pretrained(
+                    model_path, low_cpu_mem_usage=True, **kwargs)
+            elif 'mistral' in model_name.lower():
+                tokenizer = AutoTokenizer.from_pretrained(model_path)
+                model = LlavaMistralForCausalLM.from_pretrained(
+                    model_path,
+                    low_cpu_mem_usage=True,
+                    **kwargs
+                )
+            else:
+                tokenizer = AutoTokenizer.from_pretrained(
+                    model_path, use_fast=False)
+                model = LlavaLlamaForCausalLM.from_pretrained(
+                    model_path,
+                    low_cpu_mem_usage=True,
+                    **kwargs
+                )
+    else:
+        # Load language model
+        if model_base is not None:
+            # PEFT model
+            from peft import PeftModel
+            tokenizer = AutoTokenizer.from_pretrained(
+                model_base, use_fast=False)
+            model = AutoModelForCausalLM.from_pretrained(
+                model_base, low_cpu_mem_usage=True, **kwargs)
+            print(f"Loading LoRA weights from {model_path}")
+            model = PeftModel.from_pretrained(model, model_path)
+            print(f"Merging weights")
+            model = model.merge_and_unload()
+            print('Convert to FP16...')
+            model.to(torch.float16)
+        else:
+            use_fast = False
+            if 'mpt' in model_name.lower():
+                tokenizer = AutoTokenizer.from_pretrained(
+                    model_path, use_fast=True)
+                model = AutoModelForCausalLM.from_pretrained(
+                    model_path, low_cpu_mem_usage=True, trust_remote_code=True, **kwargs)
+            else:
+                tokenizer = AutoTokenizer.from_pretrained(
+                    model_path, use_fast=False)
+                model = AutoModelForCausalLM.from_pretrained(
+                    model_path, low_cpu_mem_usage=True, **kwargs)
+
+    image_processor = None
+
+    if 'llava' or 'sharegpt4video' in model_name.lower():
+        mm_use_im_start_end = getattr(
+            model.config, "mm_use_im_start_end", False)
+        mm_use_im_patch_token = getattr(
+            model.config, "mm_use_im_patch_token", True)
+        if mm_use_im_patch_token:
+            tokenizer.add_tokens(
+                [DEFAULT_IMAGE_PATCH_TOKEN], special_tokens=True)
+        if mm_use_im_start_end:
+            tokenizer.add_tokens(
+                [DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN], special_tokens=True)
+
+        vision_tower = model.get_vision_tower()
+        if not vision_tower.is_loaded:
+            vision_tower.load_model(device_map=device_map)
+        if device_map != 'auto':
+            vision_tower.to(device=device_map, dtype=torch.float16)
+        image_processor = vision_tower.image_processor
+
+    if hasattr(model.config, "max_sequence_length"):
+        context_len = model.config.max_sequence_length
+    else:
+        context_len = 2048
+
+    return tokenizer, model, image_processor, context_len

llava/model/consolidate.py

@@ -0,0 +1,29 @@
+"""
+Usage:
+python3 -m llava.model.consolidate --src ~/model_weights/llava-7b --dst ~/model_weights/llava-7b_consolidate
+"""
+import argparse
+
+import torch
+from transformers import AutoTokenizer, AutoModelForCausalLM
+from llava.model import *
+from llava.model.utils import auto_upgrade
+
+
+def consolidate_ckpt(src_path, dst_path):
+    print("Loading model")
+    auto_upgrade(src_path)
+    src_model = AutoModelForCausalLM.from_pretrained(src_path, torch_dtype=torch.float16, low_cpu_mem_usage=True)
+    src_tokenizer = AutoTokenizer.from_pretrained(src_path, use_fast=False)
+    src_model.save_pretrained(dst_path)
+    src_tokenizer.save_pretrained(dst_path)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--src", type=str, required=True)
+    parser.add_argument("--dst", type=str, required=True)
+
+    args = parser.parse_args()
+
+    consolidate_ckpt(args.src, args.dst)

llava/model/language_model/llava_llama.py

@@ -0,0 +1,159 @@
+#    Copyright 2023 Haotian Liu
+#
+#    Licensed under the Apache License, Version 2.0 (the "License");
+#    you may not use this file except in compliance with the License.
+#    You may obtain a copy of the License at
+#
+#        http://www.apache.org/licenses/LICENSE-2.0
+#
+#    Unless required by applicable law or agreed to in writing, software
+#    distributed under the License is distributed on an "AS IS" BASIS,
+#    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#    See the License for the specific language governing permissions and
+#    limitations under the License.
+
+
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+
+from transformers import AutoConfig, AutoModelForCausalLM, \
+                         LlamaConfig, LlamaModel, LlamaForCausalLM
+
+from transformers.modeling_outputs import CausalLMOutputWithPast
+from transformers.generation.utils import GenerateOutput
+
+from ..llava_arch import LlavaMetaModel, LlavaMetaForCausalLM
+
+
+class LlavaConfig(LlamaConfig):
+    model_type = "llava_llama"
+
+
+class LlavaLlamaModel(LlavaMetaModel, LlamaModel):
+    config_class = LlavaConfig
+
+    def __init__(self, config: LlamaConfig):
+        super(LlavaLlamaModel, self).__init__(config)
+
+
+class LlavaLlamaForCausalLM(LlamaForCausalLM, LlavaMetaForCausalLM):
+    config_class = LlavaConfig
+
+    def __init__(self, config):
+        super(LlamaForCausalLM, self).__init__(config)
+        self.model = LlavaLlamaModel(config)
+        self.pretraining_tp = config.pretraining_tp
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_model(self):
+        return self.model
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        images: Optional[torch.FloatTensor] = None,
+        image_sizes: Optional[List[List[int]]] = None,
+        return_dict: Optional[bool] = None,
+        **kwargs
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+
+        if inputs_embeds is None:
+            (
+                input_ids,
+                position_ids,
+                attention_mask,
+                past_key_values,
+                inputs_embeds,
+                labels
+            ) = self.prepare_inputs_labels_for_multimodal(
+                input_ids,
+                position_ids,
+                attention_mask,
+                past_key_values,
+                labels,
+                images,
+                image_sizes
+            )
+
+        return super().forward(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            labels=labels,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict
+        )
+
+    @torch.no_grad()
+    def generate(
+        self,
+        inputs: Optional[torch.Tensor] = None,
+        images: Optional[torch.Tensor] = None,
+        image_sizes: Optional[torch.Tensor] = None,
+        **kwargs,
+    ) -> Union[GenerateOutput, torch.LongTensor]:
+        position_ids = kwargs.pop("position_ids", None)
+        attention_mask = kwargs.pop("attention_mask", None)
+        if "inputs_embeds" in kwargs:
+            raise NotImplementedError("`inputs_embeds` is not supported")
+
+        if images is not None:
+            (
+                inputs,
+                position_ids,
+                attention_mask,
+                _,
+                inputs_embeds,
+                _
+            ) = self.prepare_inputs_labels_for_multimodal(
+                inputs,
+                position_ids,
+                attention_mask,
+                None,
+                None,
+                images,
+                image_sizes=image_sizes
+            )
+        else:
+            inputs_embeds = self.get_model().embed_tokens(inputs)
+
+        return super().generate(
+            position_ids=position_ids,
+            attention_mask=attention_mask,
+            inputs_embeds=inputs_embeds,
+            **kwargs
+        )
+
+    def prepare_inputs_for_generation(self, input_ids, past_key_values=None,
+                                      inputs_embeds=None, **kwargs):
+        images = kwargs.pop("images", None)
+        image_sizes = kwargs.pop("image_sizes", None)
+        inputs = super().prepare_inputs_for_generation(
+            input_ids, past_key_values=past_key_values, inputs_embeds=inputs_embeds, **kwargs
+        )
+        if images is not None:
+            inputs['images'] = images
+        if image_sizes is not None:
+            inputs['image_sizes'] = image_sizes
+        return inputs
+
+AutoConfig.register("llava_llama", LlavaConfig)
+AutoModelForCausalLM.register(LlavaConfig, LlavaLlamaForCausalLM)

llava/model/language_model/llava_mistral.py

@@ -0,0 +1,158 @@
+#    Copyright 2023 Haotian Liu
+#
+#    Licensed under the Apache License, Version 2.0 (the "License");
+#    you may not use this file except in compliance with the License.
+#    You may obtain a copy of the License at
+#
+#        http://www.apache.org/licenses/LICENSE-2.0
+#
+#    Unless required by applicable law or agreed to in writing, software
+#    distributed under the License is distributed on an "AS IS" BASIS,
+#    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#    See the License for the specific language governing permissions and
+#    limitations under the License.
+
+
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+from torch.nn import CrossEntropyLoss
+
+from transformers import AutoConfig, AutoModelForCausalLM, \
+                         MistralConfig, MistralModel, MistralForCausalLM
+
+from transformers.modeling_outputs import CausalLMOutputWithPast
+from transformers.generation.utils import GenerateOutput
+
+from ..llava_arch import LlavaMetaModel, LlavaMetaForCausalLM
+
+
+class LlavaMistralConfig(MistralConfig):
+    model_type = "llava_mistral"
+
+
+class LlavaMistralModel(LlavaMetaModel, MistralModel):
+    config_class = LlavaMistralConfig
+
+    def __init__(self, config: MistralConfig):
+        super(LlavaMistralModel, self).__init__(config)
+
+
+class LlavaMistralForCausalLM(MistralForCausalLM, LlavaMetaForCausalLM):
+    config_class = LlavaMistralConfig
+
+    def __init__(self, config):
+        super(MistralForCausalLM, self).__init__(config)
+        self.model = LlavaMistralModel(config)
+
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_model(self):
+        return self.model
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        images: Optional[torch.FloatTensor] = None,
+        image_sizes: Optional[List[List[int]]] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+
+        if inputs_embeds is None:
+            (
+                input_ids,
+                position_ids,
+                attention_mask,
+                past_key_values,
+                inputs_embeds,
+                labels
+            ) = self.prepare_inputs_labels_for_multimodal(
+                input_ids,
+                position_ids,
+                attention_mask,
+                past_key_values,
+                labels,
+                images,
+                image_sizes
+            )
+
+        return super().forward(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            labels=labels,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict
+        )
+
+    @torch.no_grad()
+    def generate(
+        self,
+        inputs: Optional[torch.Tensor] = None,
+        images: Optional[torch.Tensor] = None,
+        image_sizes: Optional[torch.Tensor] = None,
+        **kwargs,
+    ) -> Union[GenerateOutput, torch.LongTensor]:
+        position_ids = kwargs.pop("position_ids", None)
+        attention_mask = kwargs.pop("attention_mask", None)
+        if "inputs_embeds" in kwargs:
+            raise NotImplementedError("`inputs_embeds` is not supported")
+
+        if images is not None:
+            (
+                inputs,
+                position_ids,
+                attention_mask,
+                _,
+                inputs_embeds,
+                _
+            ) = self.prepare_inputs_labels_for_multimodal(
+                inputs,
+                position_ids,
+                attention_mask,
+                None,
+                None,
+                images,
+                image_sizes=image_sizes
+            )
+        else:
+            inputs_embeds = self.get_model().embed_tokens(inputs)
+
+        return super().generate(
+            position_ids=position_ids,
+            attention_mask=attention_mask,
+            inputs_embeds=inputs_embeds,
+            **kwargs
+        )
+
+    def prepare_inputs_for_generation(self, input_ids, past_key_values=None,
+                                      inputs_embeds=None, **kwargs):
+        images = kwargs.pop("images", None)
+        image_sizes = kwargs.pop("image_sizes", None)
+        inputs = super().prepare_inputs_for_generation(
+            input_ids, past_key_values=past_key_values, inputs_embeds=inputs_embeds, **kwargs
+        )
+        if images is not None:
+            inputs['images'] = images
+        if image_sizes is not None:
+            inputs['image_sizes'] = image_sizes
+        return inputs
+
+AutoConfig.register("llava_mistral", LlavaMistralConfig)
+AutoModelForCausalLM.register(LlavaMistralConfig, LlavaMistralForCausalLM)

llava/model/language_model/llava_mpt.py

@@ -0,0 +1,97 @@
+#    Copyright 2023 Haotian Liu
+#
+#    Licensed under the Apache License, Version 2.0 (the "License");
+#    you may not use this file except in compliance with the License.
+#    You may obtain a copy of the License at
+#
+#        http://www.apache.org/licenses/LICENSE-2.0
+#
+#    Unless required by applicable law or agreed to in writing, software
+#    distributed under the License is distributed on an "AS IS" BASIS,
+#    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#    See the License for the specific language governing permissions and
+#    limitations under the License.
+
+
+from typing import Optional, Tuple
+
+import torch
+
+from transformers import AutoConfig, AutoModelForCausalLM, \
+                         MptConfig, MptForCausalLM, MptModel
+from llava.model.llava_arch import LlavaMetaModel, LlavaMetaForCausalLM
+
+
+class LlavaMptConfig(MptConfig):
+    model_type = "llava_mpt"
+
+
+class LlavaMptModel(LlavaMetaModel, MptModel):
+    config_class = LlavaMptConfig
+
+    def __init__(self, config: MptConfig):
+        config.hidden_size = config.d_model
+        super(LlavaMptModel, self).__init__(config)
+    
+    def embed_tokens(self, x):
+        return self.wte(x)
+
+
+class LlavaMptForCausalLM(MptForCausalLM, LlavaMetaForCausalLM):
+    config_class = LlavaMptConfig
+    supports_gradient_checkpointing = True
+
+    def __init__(self, config):
+        super(MptForCausalLM, self).__init__(config)
+
+        self.transformer = LlavaMptModel(config)
+        self.lm_head = torch.nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_model(self):
+        return self.transformer
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, LlavaMptModel):
+            module.gradient_checkpointing = value
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        labels: Optional[torch.Tensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        images=None):
+
+        input_ids, attention_mask, past_key_values, inputs_embeds, labels = self.prepare_inputs_labels_for_multimodal(input_ids, attention_mask, past_key_values, labels, images)
+        
+        return super().forward(
+            input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            inputs_embeds=inputs_embeds,
+            labels=labels,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+    def prepare_inputs_for_generation(self, input_ids, past_key_values=None, inputs_embeds=None, **kwargs):
+        images = kwargs.pop("images", None)
+        _inputs = super().prepare_inputs_for_generation(
+            input_ids, past_key_values=past_key_values, inputs_embeds=inputs_embeds, **kwargs
+        )
+        _inputs['images'] = images
+        return _inputs
+
+
+AutoConfig.register("llava_mpt", LlavaMptConfig)
+AutoModelForCausalLM.register(LlavaMptConfig, LlavaMptForCausalLM)

llava/model/llava_arch.py

@@ -0,0 +1,367 @@
+#    Copyright 2023 Haotian Liu
+#
+#    Licensed under the Apache License, Version 2.0 (the "License");
+#    you may not use this file except in compliance with the License.
+#    You may obtain a copy of the License at
+#
+#        http://www.apache.org/licenses/LICENSE-2.0
+#
+#    Unless required by applicable law or agreed to in writing, software
+#    distributed under the License is distributed on an "AS IS" BASIS,
+#    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#    See the License for the specific language governing permissions and
+#    limitations under the License.
+
+
+from abc import ABC, abstractmethod
+
+import torch
+import torch.nn as nn
+
+from .multimodal_encoder.builder import build_vision_tower
+from .multimodal_projector.builder import build_vision_projector
+
+from llava.constants import IGNORE_INDEX, IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_PATCH_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
+
+from llava.mm_utils import get_anyres_image_grid_shape
+
+
+class LlavaMetaModel:
+
+    def __init__(self, config):
+        super(LlavaMetaModel, self).__init__(config)
+        if hasattr(config, "mm_vision_tower"):
+            self.vision_tower = build_vision_tower(config, delay_load=True)
+            self.mm_projector = build_vision_projector(config)
+
+            if 'unpad' in getattr(config, 'mm_patch_merge_type', ''):
+                self.image_newline = nn.Parameter(
+                    torch.empty(config.hidden_size, dtype=self.dtype)
+                )
+
+    def get_vision_tower(self):
+        vision_tower = getattr(self, 'vision_tower', None)
+        if type(vision_tower) is list:
+            vision_tower = vision_tower[0]
+        return vision_tower
+
+    def initialize_vision_modules(self, model_args, fsdp=None):
+        vision_tower = model_args.vision_tower
+        mm_vision_select_layer = model_args.mm_vision_select_layer
+        mm_vision_select_feature = model_args.mm_vision_select_feature
+        pretrain_mm_mlp_adapter = model_args.pretrain_mm_mlp_adapter
+        mm_patch_merge_type = model_args.mm_patch_merge_type
+
+        self.config.mm_vision_tower = vision_tower
+
+        if self.get_vision_tower() is None:
+            vision_tower = build_vision_tower(model_args)
+
+            if fsdp is not None and len(fsdp) > 0:
+                self.vision_tower = [vision_tower]
+            else:
+                self.vision_tower = vision_tower
+        else:
+            if fsdp is not None and len(fsdp) > 0:
+                vision_tower = self.vision_tower[0]
+            else:
+                vision_tower = self.vision_tower
+            vision_tower.load_model()
+
+        self.config.use_mm_proj = True
+        self.config.mm_projector_type = getattr(model_args, 'mm_projector_type', 'linear')
+        self.config.mm_hidden_size = vision_tower.hidden_size
+        self.config.mm_vision_select_layer = mm_vision_select_layer
+        self.config.mm_vision_select_feature = mm_vision_select_feature
+        self.config.mm_patch_merge_type = mm_patch_merge_type
+
+        if getattr(self, 'mm_projector', None) is None:
+            self.mm_projector = build_vision_projector(self.config)
+
+            if 'unpad' in mm_patch_merge_type:
+                embed_std = 1 / torch.sqrt(torch.tensor(self.config.hidden_size, dtype=self.dtype))
+                self.image_newline = nn.Parameter(
+                    torch.randn(self.config.hidden_size, dtype=self.dtype) * embed_std
+                )
+        else:
+            # In case it is frozen by LoRA
+            for p in self.mm_projector.parameters():
+                p.requires_grad = True
+
+        if pretrain_mm_mlp_adapter is not None:
+            mm_projector_weights = torch.load(pretrain_mm_mlp_adapter, map_location='cpu')
+            def get_w(weights, keyword):
+                return {k.split(keyword + '.')[1]: v for k, v in weights.items() if keyword in k}
+
+            self.mm_projector.load_state_dict(get_w(mm_projector_weights, 'mm_projector'))
+
+
+def unpad_image(tensor, original_size):
+    """
+    Unpads a PyTorch tensor of a padded and resized image.
+
+    Args:
+    tensor (torch.Tensor): The image tensor, assumed to be in CxHxW format.
+    original_size (tuple): The original size of PIL image (width, height).
+
+    Returns:
+    torch.Tensor: The unpadded image tensor.
+    """
+    original_width, original_height = original_size
+    current_height, current_width = tensor.shape[1:]
+
+    original_aspect_ratio = original_width / original_height
+    current_aspect_ratio = current_width / current_height
+
+    if original_aspect_ratio > current_aspect_ratio:
+        scale_factor = current_width / original_width
+        new_height = int(original_height * scale_factor)
+        padding = (current_height - new_height) // 2
+        unpadded_tensor = tensor[:, padding:current_height - padding, :]
+    else:
+        scale_factor = current_height / original_height
+        new_width = int(original_width * scale_factor)
+        padding = (current_width - new_width) // 2
+        unpadded_tensor = tensor[:, :, padding:current_width - padding]
+
+    return unpadded_tensor
+
+
+class LlavaMetaForCausalLM(ABC):
+
+    @abstractmethod
+    def get_model(self):
+        pass
+
+    def get_vision_tower(self):
+        return self.get_model().get_vision_tower()
+
+    def encode_images(self, images):
+        image_features = self.get_model().get_vision_tower()(images)
+        image_features = self.get_model().mm_projector(image_features)
+        return image_features
+
+    def prepare_inputs_labels_for_multimodal(
+        self, input_ids, position_ids, attention_mask, past_key_values, labels,
+        images, image_sizes=None
+    ):
+        vision_tower = self.get_vision_tower()
+        if vision_tower is None or images is None or input_ids.shape[1] == 1:
+            return input_ids, position_ids, attention_mask, past_key_values, None, labels
+
+        if type(images) is list or images.ndim == 5:
+            if type(images) is list:
+                images = [x.unsqueeze(0) if x.ndim == 3 else x for x in images]
+            concat_images = torch.cat([image for image in images], dim=0)
+            image_features = self.encode_images(concat_images)
+            split_sizes = [image.shape[0] for image in images]
+            image_features = torch.split(image_features, split_sizes, dim=0)
+            mm_patch_merge_type = getattr(self.config, 'mm_patch_merge_type', 'flat')
+            image_aspect_ratio = getattr(self.config, 'image_aspect_ratio', 'square')
+            if mm_patch_merge_type == 'flat':
+                image_features = [x.flatten(0, 1) for x in image_features]
+            elif mm_patch_merge_type.startswith('spatial'):
+                new_image_features = []
+                for image_idx, image_feature in enumerate(image_features):
+                    if image_feature.shape[0] > 1:
+                        base_image_feature = image_feature[0]
+                        image_feature = image_feature[1:]
+                        height = width = self.get_vision_tower().num_patches_per_side
+                        assert height * width == base_image_feature.shape[0]
+                        if image_aspect_ratio == 'anyres':
+                            num_patch_width, num_patch_height = get_anyres_image_grid_shape(image_sizes[image_idx], self.config.image_grid_pinpoints, self.get_vision_tower().config.image_size)
+                            image_feature = image_feature.view(num_patch_height, num_patch_width, height, width, -1)
+                        else:
+                            raise NotImplementedError
+                        if 'unpad' in mm_patch_merge_type:
+                            image_feature = image_feature.permute(4, 0, 2, 1, 3).contiguous()
+                            image_feature = image_feature.flatten(1, 2).flatten(2, 3)
+                            image_feature = unpad_image(image_feature, image_sizes[image_idx])
+                            image_feature = torch.cat((
+                                image_feature,
+                                self.model.image_newline[:, None, None].expand(*image_feature.shape[:-1], 1).to(image_feature.device)
+                            ), dim=-1)
+                            image_feature = image_feature.flatten(1, 2).transpose(0, 1)
+                        else:
+                            image_feature = image_feature.permute(0, 2, 1, 3, 4).contiguous()
+                            image_feature = image_feature.flatten(0, 3)
+                        image_feature = torch.cat((base_image_feature, image_feature), dim=0)
+                    else:
+                        image_feature = image_feature[0]
+                        if 'unpad' in mm_patch_merge_type:
+                            image_feature = torch.cat((
+                                image_feature,
+                                self.model.image_newline[None].to(image_feature.device)
+                            ), dim=0)
+                    new_image_features.append(image_feature)
+                image_features = new_image_features
+            else:
+                raise ValueError(f"Unexpected mm_patch_merge_type: {self.config.mm_patch_merge_type}")
+        else:
+            image_features = self.encode_images(images)
+
+        # TODO: image start / end is not implemented here to support pretraining.
+        if getattr(self.config, 'tune_mm_mlp_adapter', False) and getattr(self.config, 'mm_use_im_start_end', False):
+            raise NotImplementedError
+
+        # Let's just add dummy tensors if they do not exist,
+        # it is a headache to deal with None all the time.
+        # But it is not ideal, and if you have a better idea,
+        # please open an issue / submit a PR, thanks.
+        _labels = labels
+        _position_ids = position_ids
+        _attention_mask = attention_mask
+        if attention_mask is None:
+            attention_mask = torch.ones_like(input_ids, dtype=torch.bool)
+        else:
+            attention_mask = attention_mask.bool()
+        if position_ids is None:
+            position_ids = torch.arange(0, input_ids.shape[1], dtype=torch.long, device=input_ids.device)
+        if labels is None:
+            labels = torch.full_like(input_ids, IGNORE_INDEX)
+
+        # remove the padding using attention_mask -- FIXME
+        _input_ids = input_ids
+        input_ids = [cur_input_ids[cur_attention_mask] for cur_input_ids, cur_attention_mask in zip(input_ids, attention_mask)]
+        labels = [cur_labels[cur_attention_mask] for cur_labels, cur_attention_mask in zip(labels, attention_mask)]
+
+        new_input_embeds = []
+        new_labels = []
+        cur_image_idx = 0
+        for batch_idx, cur_input_ids in enumerate(input_ids):
+            num_images = (cur_input_ids == IMAGE_TOKEN_INDEX).sum()
+            if num_images == 0:
+                cur_image_features = image_features[cur_image_idx]
+                cur_input_embeds_1 = self.get_model().embed_tokens(cur_input_ids)
+                cur_input_embeds = torch.cat([cur_input_embeds_1, cur_image_features[0:0]], dim=0)
+                new_input_embeds.append(cur_input_embeds)
+                new_labels.append(labels[batch_idx])
+                cur_image_idx += 1
+                continue
+
+            image_token_indices = [-1] + torch.where(cur_input_ids == IMAGE_TOKEN_INDEX)[0].tolist() + [cur_input_ids.shape[0]]
+            cur_input_ids_noim = []
+            cur_labels = labels[batch_idx]
+            cur_labels_noim = []
+            for i in range(len(image_token_indices) - 1):
+                cur_input_ids_noim.append(cur_input_ids[image_token_indices[i]+1:image_token_indices[i+1]])
+                cur_labels_noim.append(cur_labels[image_token_indices[i]+1:image_token_indices[i+1]])
+            split_sizes = [x.shape[0] for x in cur_labels_noim]
+            cur_input_embeds = self.get_model().embed_tokens(torch.cat(cur_input_ids_noim))
+            cur_input_embeds_no_im = torch.split(cur_input_embeds, split_sizes, dim=0)
+            cur_new_input_embeds = []
+            cur_new_labels = []
+
+            for i in range(num_images + 1):
+                cur_new_input_embeds.append(cur_input_embeds_no_im[i])
+                cur_new_labels.append(cur_labels_noim[i])
+                if i < num_images:
+                    cur_image_features = image_features[cur_image_idx]
+                    cur_image_idx += 1
+                    cur_new_input_embeds.append(cur_image_features)
+                    cur_new_labels.append(torch.full((cur_image_features.shape[0],), IGNORE_INDEX, device=cur_labels.device, dtype=cur_labels.dtype))
+
+            cur_new_input_embeds = [x.to(self.device) for x in cur_new_input_embeds]
+
+            cur_new_input_embeds = torch.cat(cur_new_input_embeds)
+            cur_new_labels = torch.cat(cur_new_labels)
+
+            new_input_embeds.append(cur_new_input_embeds)
+            new_labels.append(cur_new_labels)
+
+        # Truncate sequences to max length as image embeddings can make the sequence longer
+        tokenizer_model_max_length = getattr(self.config, 'tokenizer_model_max_length', None)
+        if tokenizer_model_max_length is not None:
+            new_input_embeds = [x[:tokenizer_model_max_length] for x in new_input_embeds]
+            new_labels = [x[:tokenizer_model_max_length] for x in new_labels]
+
+        # Combine them
+        max_len = max(x.shape[0] for x in new_input_embeds)
+        batch_size = len(new_input_embeds)
+
+        new_input_embeds_padded = []
+        new_labels_padded = torch.full((batch_size, max_len), IGNORE_INDEX, dtype=new_labels[0].dtype, device=new_labels[0].device)
+        attention_mask = torch.zeros((batch_size, max_len), dtype=attention_mask.dtype, device=attention_mask.device)
+        position_ids = torch.zeros((batch_size, max_len), dtype=position_ids.dtype, device=position_ids.device)
+
+        for i, (cur_new_embed, cur_new_labels) in enumerate(zip(new_input_embeds, new_labels)):
+            cur_len = cur_new_embed.shape[0]
+            if getattr(self.config, 'tokenizer_padding_side', 'right') == "left":
+                new_input_embeds_padded.append(torch.cat((
+                    torch.zeros((max_len - cur_len, cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device),
+                    cur_new_embed
+                ), dim=0))
+                if cur_len > 0:
+                    new_labels_padded[i, -cur_len:] = cur_new_labels
+                    attention_mask[i, -cur_len:] = True
+                    position_ids[i, -cur_len:] = torch.arange(0, cur_len, dtype=position_ids.dtype, device=position_ids.device)
+            else:
+                new_input_embeds_padded.append(torch.cat((
+                    cur_new_embed,
+                    torch.zeros((max_len - cur_len, cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device)
+                ), dim=0))
+                if cur_len > 0:
+                    new_labels_padded[i, :cur_len] = cur_new_labels
+                    attention_mask[i, :cur_len] = True
+                    position_ids[i, :cur_len] = torch.arange(0, cur_len, dtype=position_ids.dtype, device=position_ids.device)
+
+        new_input_embeds = torch.stack(new_input_embeds_padded, dim=0)
+
+        if _labels is None:
+            new_labels = None
+        else:
+            new_labels = new_labels_padded
+
+        if _attention_mask is None:
+            attention_mask = None
+        else:
+            attention_mask = attention_mask.to(dtype=_attention_mask.dtype)
+
+        if _position_ids is None:
+            position_ids = None
+
+        return None, position_ids, attention_mask, past_key_values, new_input_embeds, new_labels
+
+    def initialize_vision_tokenizer(self, model_args, tokenizer):
+        if model_args.mm_use_im_patch_token:
+            tokenizer.add_tokens([DEFAULT_IMAGE_PATCH_TOKEN], special_tokens=True)
+            self.resize_token_embeddings(len(tokenizer))
+
+        if model_args.mm_use_im_start_end:
+            num_new_tokens = tokenizer.add_tokens([DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN], special_tokens=True)
+            self.resize_token_embeddings(len(tokenizer))
+
+            if num_new_tokens > 0:
+                input_embeddings = self.get_input_embeddings().weight.data
+                output_embeddings = self.get_output_embeddings().weight.data
+
+                input_embeddings_avg = input_embeddings[:-num_new_tokens].mean(
+                    dim=0, keepdim=True)
+                output_embeddings_avg = output_embeddings[:-num_new_tokens].mean(
+                    dim=0, keepdim=True)
+
+                input_embeddings[-num_new_tokens:] = input_embeddings_avg
+                output_embeddings[-num_new_tokens:] = output_embeddings_avg
+
+            if model_args.tune_mm_mlp_adapter:
+                for p in self.get_input_embeddings().parameters():
+                    p.requires_grad = True
+                for p in self.get_output_embeddings().parameters():
+                    p.requires_grad = False
+
+            if model_args.pretrain_mm_mlp_adapter:
+                mm_projector_weights = torch.load(model_args.pretrain_mm_mlp_adapter, map_location='cpu')
+                embed_tokens_weight = mm_projector_weights['model.embed_tokens.weight']
+                assert num_new_tokens == 2
+                if input_embeddings.shape == embed_tokens_weight.shape:
+                    input_embeddings[-num_new_tokens:] = embed_tokens_weight[-num_new_tokens:]
+                elif embed_tokens_weight.shape[0] == num_new_tokens:
+                    input_embeddings[-num_new_tokens:] = embed_tokens_weight
+                else:
+                    raise ValueError(f"Unexpected embed_tokens_weight shape. Pretrained: {embed_tokens_weight.shape}. Current: {input_embeddings.shape}. Numer of new tokens: {num_new_tokens}.")
+        elif model_args.mm_use_im_patch_token:
+            if model_args.tune_mm_mlp_adapter:
+                for p in self.get_input_embeddings().parameters():
+                    p.requires_grad = False
+                for p in self.get_output_embeddings().parameters():
+                    p.requires_grad = False

llava/model/make_delta.py

@@ -0,0 +1,52 @@
+"""
+Usage:
+python3 -m llava.model.make_delta --base ~/model_weights/llama-7b --target ~/model_weights/llava-7b --delta ~/model_weights/llava-7b-delta --hub-repo-id liuhaotian/llava-7b-delta
+"""
+import argparse
+
+import torch
+from tqdm import tqdm
+from transformers import AutoTokenizer, AutoModelForCausalLM
+from llava.model.utils import auto_upgrade
+
+
+def make_delta(base_model_path, target_model_path, delta_path, hub_repo_id):
+    print("Loading base model")
+    base = AutoModelForCausalLM.from_pretrained(
+        base_model_path, torch_dtype=torch.float16, low_cpu_mem_usage=True)
+
+    print("Loading target model")
+    auto_upgrade(target_model_path)
+    target = AutoModelForCausalLM.from_pretrained(target_model_path, torch_dtype=torch.float16, low_cpu_mem_usage=True)
+
+    print("Calculating delta")
+    for name, param in tqdm(target.state_dict().items(), desc="Calculating delta"):
+        if name not in base.state_dict():
+            assert name in ['model.mm_projector.weight', 'model.mm_projector.bias'], f'{name} not in base model'
+            continue
+        if param.data.shape == base.state_dict()[name].shape:
+            param.data -= base.state_dict()[name]
+        else:
+            assert name in ['model.embed_tokens.weight', 'lm_head.weight'], f'{name} dimension mismatch: {param.data.shape} vs {base.state_dict()[name].shape}'
+            bparam = base.state_dict()[name]
+            param.data[:bparam.shape[0], :bparam.shape[1]] -= bparam
+
+    print("Saving delta")
+    if hub_repo_id:
+        kwargs = {"push_to_hub": True, "repo_id": hub_repo_id}
+    else:
+        kwargs = {}
+    target.save_pretrained(delta_path, **kwargs)
+    target_tokenizer = AutoTokenizer.from_pretrained(target_model_path)
+    target_tokenizer.save_pretrained(delta_path, **kwargs)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--base-model-path", type=str, required=True)
+    parser.add_argument("--target-model-path", type=str, required=True)
+    parser.add_argument("--delta-path", type=str, required=True)
+    parser.add_argument("--hub-repo-id", type=str, default=None)
+    args = parser.parse_args()
+
+    make_delta(args.base_model_path, args.target_model_path, args.delta_path, args.hub_repo_id)

llava/model/multimodal_encoder/builder.py

@@ -0,0 +1,20 @@
+import os
+from .clip_encoder import CLIPVisionTower, CLIPVisionTowerS2
+from .siglip_encoder import SigLipVisionTower
+
+
+def build_vision_tower(vision_tower_cfg, **kwargs):
+    vision_tower = getattr(vision_tower_cfg, 'mm_vision_tower', getattr(vision_tower_cfg, 'vision_tower', None))
+    is_absolute_path_exists = os.path.exists(vision_tower)
+    use_s2 = getattr(vision_tower_cfg, 's2', False)
+    if 'siglip' not in vision_tower.lower():
+        if is_absolute_path_exists or vision_tower.startswith("openai") or vision_tower.startswith("laion") or "ShareGPT4V" in vision_tower:
+            if use_s2:
+                return CLIPVisionTowerS2(vision_tower, args=vision_tower_cfg, **kwargs)
+            else:
+                return CLIPVisionTower(vision_tower, args=vision_tower_cfg, **kwargs)
+    else:
+        if is_absolute_path_exists or vision_tower.startswith("google") or vision_tower.startswith('bczhou'):
+            return SigLipVisionTower(vision_tower, vision_tower_cfg, **kwargs)
+
+    raise ValueError(f'Unknown vision tower: {vision_tower}')

llava/model/multimodal_encoder/clip_encoder.py

@@ -0,0 +1,147 @@
+import torch
+import torch.nn as nn
+
+from transformers import CLIPVisionModel, CLIPImageProcessor, CLIPVisionConfig
+
+
+class CLIPVisionTower(nn.Module):
+    def __init__(self, vision_tower, args, delay_load=False):
+        super().__init__()
+
+        self.is_loaded = False
+
+        self.vision_tower_name = vision_tower
+        self.select_layer = args.mm_vision_select_layer
+        self.select_feature = getattr(args, 'mm_vision_select_feature', 'patch')
+
+        if not delay_load:
+            self.load_model()
+        elif getattr(args, 'unfreeze_mm_vision_tower', False):
+            self.load_model()
+        else:
+            self.cfg_only = CLIPVisionConfig.from_pretrained(self.vision_tower_name)
+
+    def load_model(self, device_map=None):
+        if self.is_loaded:
+            print('{} is already loaded, `load_model` called again, skipping.'.format(self.vision_tower_name))
+            return
+        print(f'Load vision tower from {self.vision_tower_name}')
+        self.image_processor = CLIPImageProcessor.from_pretrained(self.vision_tower_name)
+        self.vision_tower = CLIPVisionModel.from_pretrained(self.vision_tower_name, device_map=device_map)
+        self.vision_tower.requires_grad_(False)
+
+        self.is_loaded = True
+
+    def feature_select(self, image_forward_outs):
+        image_features = image_forward_outs.hidden_states[self.select_layer]
+        if self.select_feature == 'patch':
+            image_features = image_features[:, 1:]
+        elif self.select_feature == 'cls_patch':
+            image_features = image_features
+        else:
+            raise ValueError(f'Unexpected select feature: {self.select_feature}')
+        return image_features
+
+    # @torch.no_grad()
+    def forward(self, images):
+        if type(images) is list:
+            image_features = []
+            for image in images:
+                image_forward_out = self.vision_tower(image.to(device=self.device, dtype=self.dtype).unsqueeze(0), output_hidden_states=True)
+                image_feature = self.feature_select(image_forward_out).to(image.dtype)
+                image_features.append(image_feature)
+        else:
+            image_forward_outs = self.vision_tower(images.to(device=self.device, dtype=self.dtype), output_hidden_states=True)
+            image_features = self.feature_select(image_forward_outs).to(images.dtype)
+
+        return image_features
+
+    @property
+    def dummy_feature(self):
+        return torch.zeros(1, self.hidden_size, device=self.device, dtype=self.dtype)
+
+    @property
+    def dtype(self):
+        return self.vision_tower.dtype
+
+    @property
+    def device(self):
+        return self.vision_tower.device
+
+    @property
+    def config(self):
+        if self.is_loaded:
+            return self.vision_tower.config
+        else:
+            return self.cfg_only
+
+    @property
+    def hidden_size(self):
+        return self.config.hidden_size
+
+    @property
+    def num_patches_per_side(self):
+        return self.config.image_size // self.config.patch_size
+
+    @property
+    def num_patches(self):
+        return (self.config.image_size // self.config.patch_size) ** 2
+
+
+
+class CLIPVisionTowerS2(CLIPVisionTower):
+    def __init__(self, vision_tower, args, delay_load=False):
+        super().__init__(vision_tower, args, delay_load)
+
+        self.s2_scales = getattr(args, 's2_scales', '336,672,1008')
+        self.s2_scales = list(map(int, self.s2_scales.split(',')))
+        self.s2_scales.sort()
+        self.s2_split_size = self.s2_scales[0]
+        self.s2_image_size = self.s2_scales[-1]
+
+        try:
+            from s2wrapper import forward as multiscale_forward
+        except ImportError:
+            raise ImportError('Package s2wrapper not found! Please install by running: \npip install git+https://github.com/bfshi/scaling_on_scales.git')
+        self.multiscale_forward = multiscale_forward
+
+        # change resize/crop size in preprocessing to the largest image size in s2_scale
+        if not delay_load or getattr(args, 'unfreeze_mm_vision_tower', False):
+            self.image_processor.size['shortest_edge'] = self.s2_image_size
+            self.image_processor.crop_size['height'] = self.image_processor.crop_size['width'] = self.s2_image_size
+
+    def load_model(self, device_map=None):
+        if self.is_loaded:
+            print('{} is already loaded, `load_model` called again, skipping.'.format(self.vision_tower_name))
+            return
+
+        self.image_processor = CLIPImageProcessor.from_pretrained(self.vision_tower_name)
+        self.vision_tower = CLIPVisionModel.from_pretrained(self.vision_tower_name, device_map=device_map)
+        self.vision_tower.requires_grad_(False)
+
+        self.image_processor.size['shortest_edge'] = self.s2_image_size
+        self.image_processor.crop_size['height'] = self.image_processor.crop_size['width'] = self.s2_image_size
+
+        self.is_loaded = True
+
+    # @torch.no_grad()
+    def forward_feature(self, images):
+        image_forward_outs = self.vision_tower(images.to(device=self.device, dtype=self.dtype), output_hidden_states=True)
+        image_features = self.feature_select(image_forward_outs).to(images.dtype)
+        return image_features
+
+    # @torch.no_grad()
+    def forward(self, images):
+        if type(images) is list:
+            image_features = []
+            for image in images:
+                image_feature = self.multiscale_forward(self.forward_feature, image.unsqueeze(0), img_sizes=self.s2_scales, max_split_size=self.s2_split_size)
+                image_features.append(image_feature)
+        else:
+            image_features = self.multiscale_forward(self.forward_feature, images, img_sizes=self.s2_scales, max_split_size=self.s2_split_size)
+
+        return image_features
+
+    @property
+    def hidden_size(self):
+        return self.config.hidden_size * len(self.s2_scales)

llava/model/multimodal_encoder/siglip_encoder.py

@@ -0,0 +1,654 @@
+'''
+# Adapted from https://huggingface.co/MILVLG/imp-v1-3b/blob/main/vision_encoder.py
+'''
+
+from typing import Optional, Tuple, Union, Dict
+from dataclasses import dataclass
+from functools import partial, reduce
+from PIL import Image
+import torch
+import torch.utils.checkpoint
+from torch import nn
+import os
+from transformers.image_processing_utils import BatchFeature, get_size_dict
+from transformers.image_transforms import (
+    convert_to_rgb, normalize, rescale, resize, to_channel_dimension_format, )
+from transformers.image_utils import (
+    ChannelDimension, PILImageResampling, to_numpy_array, )
+from transformers.activations import ACT2FN
+from transformers.modeling_outputs import BaseModelOutput, BaseModelOutputWithPooling
+from transformers.modeling_utils import PreTrainedModel
+from transformers import PretrainedConfig
+from transformers.utils import ModelOutput
+from transformers.image_processing_utils import BaseImageProcessor
+
+
+class SigLipImageProcessor(BaseImageProcessor):
+    def __init__(self,
+                 image_mean=(0.5, 0.5, 0.5),
+                 image_std=(0.5, 0.5, 0.5),
+                 size=(384, 384),
+                 crop_size: Dict[str, int] = None,
+                 resample=PILImageResampling.BICUBIC,
+                 rescale_factor=1 / 255,
+                 data_format=ChannelDimension.FIRST):
+        crop_size = crop_size if crop_size is not None else {
+            "height": 384, "width": 384}
+        crop_size = get_size_dict(
+            crop_size, default_to_square=True, param_name="crop_size")
+
+        self.image_mean = image_mean
+        self.image_std = image_std
+        self.size = size
+        self.resample = resample
+        self.rescale_factor = rescale_factor
+        self.data_format = data_format
+        self.crop_size = crop_size
+
+    def preprocess(self, images, return_tensors):
+        if isinstance(images, Image.Image):
+            images = [images]
+        else:
+            assert isinstance(images, list)
+
+        transforms = [
+            convert_to_rgb,
+            to_numpy_array,
+            partial(resize, size=self.size, resample=self.resample,
+                    data_format=self.data_format),
+            partial(rescale, scale=self.rescale_factor,
+                    data_format=self.data_format),
+            partial(normalize, mean=self.image_mean,
+                    std=self.image_std, data_format=self.data_format),
+            partial(to_channel_dimension_format, channel_dim=self.data_format,
+                    input_channel_dim=self.data_format),
+        ]
+
+        images = reduce(lambda x, f: [*map(f, x)], transforms, images)
+        data = {"pixel_values": images}
+
+        return BatchFeature(data=data, tensor_type=return_tensors)
+
+
+class SigLipVisionConfig(PretrainedConfig):
+    model_type = "siglip_vision_model"
+
+    def __init__(
+            self,
+            hidden_size=1152,
+            image_mean=(0.5, 0.5, 0.5),
+            intermediate_size=4304,
+            num_hidden_layers=27,
+            num_attention_heads=16,
+            num_channels=3,
+            image_size=384,
+            patch_size=14,
+            hidden_act="gelu_pytorch_tanh",
+            layer_norm_eps=1e-6,
+            attention_dropout=0.0,
+            **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.num_channels = num_channels
+        self.patch_size = patch_size
+        self.image_size = image_size
+        self.attention_dropout = attention_dropout
+        self.layer_norm_eps = layer_norm_eps
+        self.hidden_act = hidden_act
+        self.image_mean = image_mean
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
+        cls._set_token_in_kwargs(kwargs)
+
+        config_dict, kwargs = cls.get_config_dict(
+            pretrained_model_name_or_path, **kwargs)
+
+        # get the vision config dict if we are loading from SigLipConfig
+        if config_dict.get("model_type") == "siglip":
+            config_dict = config_dict["vision_config"]
+
+        if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
+            logger.warning(
+                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+                f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
+            )
+
+        return cls.from_dict(config_dict, **kwargs)
+
+
+@dataclass
+# Copied from transformers.models.clip.modeling_clip.CLIPVisionModelOutput with CLIP->SigLip
+class SigLipVisionModelOutput(ModelOutput):
+    """
+    Base class for vision model's outputs that also contains image embeddings of the pooling of the last hidden states.
+
+    Args:
+        image_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim)` *optional* returned when model is initialized with `with_projection=True`):
+            The image embeddings obtained by applying the projection layer to the pooler_output.
+        last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
+            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
+        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+            sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    image_embeds: Optional[torch.FloatTensor] = None
+    last_hidden_state: torch.FloatTensor = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+class SigLipVisionEmbeddings(nn.Module):
+    def __init__(self, config: SigLipVisionConfig):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.image_size = config.image_size
+        self.patch_size = config.patch_size
+
+        self.patch_embedding = nn.Conv2d(
+            in_channels=config.num_channels,
+            out_channels=self.embed_dim,
+            kernel_size=self.patch_size,
+            stride=self.patch_size,
+            padding="valid",
+        )
+
+        self.num_patches = (self.image_size // self.patch_size) ** 2
+        self.num_positions = self.num_patches
+        self.position_embedding = nn.Embedding(
+            self.num_positions, self.embed_dim)
+        self.register_buffer("position_ids", torch.arange(
+            self.num_positions).expand((1, -1)), persistent=False)
+
+    def forward(self, pixel_values: torch.FloatTensor) -> torch.Tensor:
+        patch_embeds = self.patch_embedding(
+            pixel_values)  # shape = [*, width, grid, grid]
+        embeddings = patch_embeds.flatten(2).transpose(1, 2)
+
+        embeddings = embeddings + self.position_embedding(self.position_ids)
+        return embeddings
+
+
+class SigLipAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    # Copied from transformers.models.clip.modeling_clip.CLIPAttention.__init__
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.embed_dim // self.num_heads
+        if self.head_dim * self.num_heads != self.embed_dim:
+            raise ValueError(
+                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
+                f" {self.num_heads})."
+            )
+        self.scale = self.head_dim ** -0.5
+        self.dropout = config.attention_dropout
+
+        self.k_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.v_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.q_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.out_proj = nn.Linear(self.embed_dim, self.embed_dim)
+
+    def forward(
+            self,
+            hidden_states: torch.Tensor,
+            attention_mask: Optional[torch.Tensor] = None,
+            output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        """Input shape: Batch x Time x Channel"""
+
+        batch_size, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(
+            batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(
+            batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(
+            batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+
+        k_v_seq_len = key_states.shape[-2]
+        attn_weights = torch.matmul(
+            query_states, key_states.transpose(2, 3)) * self.scale
+
+        if attn_weights.size() != (batch_size, self.num_heads, q_len, k_v_seq_len):
+            raise ValueError(
+                f"Attention weights should be of size {(batch_size, self.num_heads, q_len, k_v_seq_len)}, but is"
+                f" {attn_weights.size()}"
+            )
+
+        if attention_mask is not None:
+            if attention_mask.size() != (batch_size, 1, q_len, k_v_seq_len):
+                raise ValueError(
+                    f"Attention mask should be of size {(batch_size, 1, q_len, k_v_seq_len)}, but is {attention_mask.size()}"
+                )
+            attn_weights = attn_weights + attention_mask
+
+        # upcast attention to fp32
+        attn_weights = nn.functional.softmax(
+            attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
+        attn_weights = nn.functional.dropout(
+            attn_weights, p=self.dropout, training=self.training)
+        attn_output = torch.matmul(attn_weights, value_states)
+
+        if attn_output.size() != (batch_size, self.num_heads, q_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(batch_size, self.num_heads, q_len, self.head_dim)}, but is"
+                f" {attn_output.size()}"
+            )
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.reshape(batch_size, q_len, self.embed_dim)
+
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output, attn_weights
+
+
+# Copied from transformers.models.clip.modeling_clip.CLIPMLP with CLIP->SigLip
+class SigLipMLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.activation_fn = ACT2FN[config.hidden_act]
+        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
+        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.fc1(hidden_states)
+        hidden_states = self.activation_fn(hidden_states)
+        hidden_states = self.fc2(hidden_states)
+        return hidden_states
+
+
+# Copied from transformers.models.clip.modeling_clip.CLIPEncoderLayer with CLIP->SigLip
+class SigLipEncoderLayer(nn.Module):
+    def __init__(self, config: SigLipVisionConfig):
+        super().__init__()
+        self.embed_dim = config.hidden_size
+        self.self_attn = SigLipAttention(config)
+        self.layer_norm1 = nn.LayerNorm(
+            self.embed_dim, eps=config.layer_norm_eps)
+        self.mlp = SigLipMLP(config)
+        self.layer_norm2 = nn.LayerNorm(
+            self.embed_dim, eps=config.layer_norm_eps)
+
+    # Ignore copy
+    def forward(
+            self,
+            hidden_states: torch.Tensor,
+            attention_mask: torch.Tensor,
+            output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.FloatTensor]:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`):
+                Input to the layer of shape `(batch, seq_len, embed_dim)`.
+            attention_mask (`torch.FloatTensor`):
+                Attention mask of shape `(batch, 1, q_len, k_v_seq_len)` where padding elements are indicated by very large negative values.
+            output_attentions (`bool`, *optional*, defaults to `False`):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+        """
+        residual = hidden_states
+
+        hidden_states = self.layer_norm1(hidden_states)
+        hidden_states, attn_weights = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            output_attentions=output_attentions,
+        )
+        hidden_states = residual + hidden_states
+
+        residual = hidden_states
+        hidden_states = self.layer_norm2(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (attn_weights,)
+
+        return outputs
+
+
+class SigLipPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = SigLipVisionConfig
+    base_model_prefix = "siglip"
+    supports_gradient_checkpointing = True
+
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        pass
+
+
+# Copied from transformers.models.clip.modeling_clip.CLIPEncoder with CLIP->SigLip
+class SigLipEncoder(nn.Module):
+    """
+    Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
+    [`SigLipEncoderLayer`].
+
+    Args:
+        config: SigLipVisionConfig
+    """
+
+    def __init__(self, config: SigLipVisionConfig):
+        super().__init__()
+        self.config = config
+        self.layers = nn.ModuleList(
+            [SigLipEncoderLayer(config) for _ in range(config.num_hidden_layers)])
+        self.gradient_checkpointing = False
+
+    # Ignore copy
+    def forward(
+            self,
+            inputs_embeds,
+            attention_mask: Optional[torch.Tensor] = None,
+            output_attentions: Optional[bool] = None,
+            output_hidden_states: Optional[bool] = None,
+            return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutput]:
+        r"""
+        Args:
+            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+                than the model's internal embedding lookup matrix.
+            attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+
+                [What are attention masks?](../glossary#attention-mask)
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            output_hidden_states (`bool`, *optional*):
+                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
+                for more detail.
+            return_dict (`bool`, *optional*):
+                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        encoder_states = () if output_hidden_states else None
+        all_attentions = () if output_attentions else None
+
+        hidden_states = inputs_embeds
+        for encoder_layer in self.layers:
+            if output_hidden_states:
+                encoder_states = encoder_states + (hidden_states,)
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = self._gradient_checkpointing_func(
+                    encoder_layer.__call__,
+                    hidden_states,
+                    attention_mask,
+                    output_attentions,
+                )
+            else:
+                layer_outputs = encoder_layer(
+                    hidden_states,
+                    attention_mask,
+                    output_attentions=output_attentions,
+                )
+
+            hidden_states = layer_outputs[0]
+
+            if output_attentions:
+                all_attentions = all_attentions + (layer_outputs[1],)
+
+        if output_hidden_states:
+            encoder_states = encoder_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
+        return BaseModelOutput(
+            last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
+        )
+
+
+class SigLipVisionTransformer(nn.Module):
+    def __init__(self, config: SigLipVisionConfig):
+        super().__init__()
+        self.config = config
+        embed_dim = config.hidden_size
+
+        self.embeddings = SigLipVisionEmbeddings(config)
+        self.encoder = SigLipEncoder(config)
+        self.post_layernorm = nn.LayerNorm(
+            embed_dim, eps=config.layer_norm_eps)
+        self.head = SigLipMultiheadAttentionPoolingHead(config)
+
+    def forward(
+            self,
+            pixel_values,
+            output_attentions: Optional[bool] = None,
+            output_hidden_states: Optional[bool] = None,
+            return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPooling]:
+        r"""
+        Returns:
+
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        hidden_states = self.embeddings(pixel_values)
+
+        encoder_outputs = self.encoder(
+            inputs_embeds=hidden_states,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        last_hidden_state = encoder_outputs[0]
+        last_hidden_state = self.post_layernorm(last_hidden_state)
+
+        pooled_output = self.head(last_hidden_state)
+
+        if not return_dict:
+            return (last_hidden_state, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPooling(
+            last_hidden_state=last_hidden_state,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
+
+
+class SigLipMultiheadAttentionPoolingHead(nn.Module):
+    """Multihead Attention Pooling."""
+
+    def __init__(self, config: SigLipVisionConfig):
+        super().__init__()
+
+        self.probe = nn.Parameter(torch.randn(1, 1, config.hidden_size))
+        self.attention = torch.nn.MultiheadAttention(
+            config.hidden_size, config.num_attention_heads, batch_first=True)
+        self.layernorm = nn.LayerNorm(
+            config.hidden_size, eps=config.layer_norm_eps)
+        self.mlp = SigLipMLP(config)
+
+    def forward(self, hidden_state):
+        batch_size = hidden_state.shape[0]
+        probe = self.probe.repeat(batch_size, 1, 1)
+
+        hidden_state = self.attention(probe, hidden_state, hidden_state)[0]
+
+        residual = hidden_state
+        hidden_state = self.layernorm(hidden_state)
+        hidden_state = residual + self.mlp(hidden_state)
+
+        return hidden_state[:, 0]
+
+
+class SigLipVisionModel(SigLipPreTrainedModel):
+    config_class = SigLipVisionConfig
+    main_input_name = "pixel_values"
+    _no_split_modules = ["SigLipEncoderLayer"]
+
+    def __init__(self, config: SigLipVisionConfig):
+        super().__init__(config)
+
+        self.vision_model = SigLipVisionTransformer(config)
+        del self.vision_model.encoder.layers[-1:]
+        self.vision_model.head = nn.Identity()
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.vision_model.embeddings.patch_embedding
+
+    def forward(
+            self,
+            pixel_values,
+            output_attentions: Optional[bool] = None,
+            output_hidden_states: Optional[bool] = None,
+            return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPooling]:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import AutoProcessor, SigLipVisionModel
+
+        >>> model = SigLipVisionModel.from_pretrained("google/siglip-base-patch16-224")
+        >>> processor = AutoProcessor.from_pretrained("google/siglip-base-patch16-224")
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> inputs = processor(images=image, return_tensors="pt")
+
+        >>> outputs = model(**inputs)
+        >>> last_hidden_state = outputs.last_hidden_state
+        >>> pooled_output = outputs.pooler_output  # pooled features
+        ```"""
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        return self.vision_model(
+            pixel_values=pixel_values,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+
+class SigLipVisionTower(nn.Module):
+    def __init__(self, vision_tower, vision_tower_cfg, delay_load=False):
+        super().__init__()
+
+        self.is_loaded = False
+
+        self.config = SigLipVisionConfig()
+
+        self.vision_tower_name = vision_tower
+
+        self.image_processor = SigLipImageProcessor()
+
+        if not delay_load:
+            self.load_model()
+        else:
+            self.cfg_only = self.config
+
+    def load_model(self, device_map=None):
+        if self.is_loaded:
+            print('{} is already loaded, `load_model` called again, skipping.'.format(
+                self.vision_tower_name))
+            return
+
+        self.vision_tower = SigLipVisionModel.from_pretrained(
+            self.vision_tower_name, device_map=device_map)
+
+        self.vision_tower.requires_grad_(False)
+        self.vision_tower.eval()
+
+        self.is_loaded = True
+
+    # @torch.no_grad()
+    def forward(self, images):
+        if type(images) is list:
+            image_features = []
+            for image in images:
+                image_forward_out = self.vision_tower(image.to(device=self.device, dtype=self.dtype).unsqueeze(0),
+                                                      output_hidden_states=True)
+                image_feature = image_forward_out.hidden_states[-1].to(
+                    image.dtype)
+                assert image_features.shape[-2] == 729
+                image_features.append(image_feature)
+        else:
+            image_forward_outs = self.vision_tower(images.to(device=self.device, dtype=self.dtype),
+                                                   output_hidden_states=True)
+            image_features = image_forward_outs.hidden_states[-1].to(
+                images.dtype)
+            assert image_features.shape[-2] == 729
+
+        return image_features
+
+    @property
+    def dummy_feature(self):
+        return torch.zeros(1, self.hidden_size, device=self.device, dtype=self.dtype)
+
+    @property
+    def dtype(self):
+        for p in self.vision_tower.parameters():
+            return p.dtype
+
+    @property
+    def device(self):
+        for p in self.vision_tower.parameters():
+            return p.device
+
+    @property
+    def hidden_size(self):
+        return self.config.hidden_size
+
+    @property
+    def num_patches_per_side(self):
+        return self.config.image_size // self.config.patch_size
+
+    @property
+    def num_patches(self):
+        return (self.config.image_size // self.config.patch_size) ** 2

llava/model/multimodal_projector/builder.py

@@ -0,0 +1,51 @@
+import torch
+import torch.nn as nn
+import re
+
+
+class IdentityMap(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x, *args, **kwargs):
+        return x
+
+    @property
+    def config(self):
+        return {"mm_projector_type": 'identity'}
+
+
+class SimpleResBlock(nn.Module):
+    def __init__(self, channels):
+        super().__init__()
+        self.pre_norm = nn.LayerNorm(channels)
+
+        self.proj = nn.Sequential(
+            nn.Linear(channels, channels),
+            nn.GELU(),
+            nn.Linear(channels, channels)
+        )
+    def forward(self, x):
+        x = self.pre_norm(x)
+        return x + self.proj(x)
+
+
+def build_vision_projector(config, delay_load=False, **kwargs):
+    projector_type = getattr(config, 'mm_projector_type', 'linear')
+
+    if projector_type == 'linear':
+        return nn.Linear(config.mm_hidden_size, config.hidden_size)
+
+    mlp_gelu_match = re.match(r'^mlp(\d+)x_gelu$', projector_type)
+    if mlp_gelu_match:
+        mlp_depth = int(mlp_gelu_match.group(1))
+        modules = [nn.Linear(config.mm_hidden_size, config.hidden_size)]
+        for _ in range(1, mlp_depth):
+            modules.append(nn.GELU())
+            modules.append(nn.Linear(config.hidden_size, config.hidden_size))
+        return nn.Sequential(*modules)
+
+    if projector_type == 'identity':
+        return IdentityMap()
+
+    raise ValueError(f'Unknown projector type: {projector_type}')

llava/model/utils.py

@@ -0,0 +1,20 @@
+from transformers import AutoConfig
+
+
+def auto_upgrade(config):
+    cfg = AutoConfig.from_pretrained(config)
+    if 'llava' in config and 'llava' not in cfg.model_type:
+        assert cfg.model_type == 'llama'
+        print("You are using newer LLaVA code base, while the checkpoint of v0 is from older code base.")
+        print("You must upgrade the checkpoint to the new code base (this can be done automatically).")
+        confirm = input("Please confirm that you want to upgrade the checkpoint. [Y/N]")
+        if confirm.lower() in ["y", "yes"]:
+            print("Upgrading checkpoint...")
+            assert len(cfg.architectures) == 1
+            setattr(cfg.__class__, "model_type", "llava")
+            cfg.architectures[0] = 'LlavaLlamaForCausalLM'
+            cfg.save_pretrained(config)
+            print("Checkpoint upgraded.")
+        else:
+            print("Checkpoint upgrade aborted.")
+            exit(1)

llava/serve/gradio_utils.py

@@ -0,0 +1,186 @@
+import numpy as np
+import torch
+from decord import VideoReader, cpu
+from PIL import Image
+
+from llava.constants import DEFAULT_IMAGE_TOKEN, IMAGE_TOKEN_INDEX
+from llava.conversation import conv_templates
+from llava.mm_utils import (KeywordsStoppingCriteria, get_model_name_from_path,
+                            process_images, tokenizer_image_token)
+from llava.model.builder import load_pretrained_model
+from llava.utils import disable_torch_init
+
+title_markdown = ("""
+<div style="display: flex; justify-content: flex-start; align-items: center; text-align: center;">
+  <div style="margin-right: 20px; display: flex; align-items: center;">
+    <a href="https://github.com/ShareGPT4Omni/ShareGPT4Video" style="text-decoration: none; display: flex; align-items: center;">
+      <img src="https://raw.githubusercontent.com/ShareGPT4V/ShareGPT4V-Resources/master/images/share4video_tight.png" alt="ShareGPT4Video🚀" style="max-width: 120px; height: auto;">
+    </a>
+  </div>
+  <div>
+    <h1>ShareGPT4Video: Improving Video Understanding and Generation with Better Captions</h1>
+    <h5 style="margin: 0;">If you like our project, please give us a star ✨ on Github for the latest update.</h5>
+    <h5 style="margin: 0;"> <a href="https://sharegpt4video.github.io/">[Project Page]</a> <a href="https://github.com/ShareGPT4Omni/ShareGPT4Video">[Code]</a> <a href="https://arxiv.org/abs/2406.04325v1">[Paper]</a>
+  </div>
+</div>
+""")
+
+block_css = """
+#buttons button {
+    min-width: min(120px,100%);
+}
+"""
+
+
+learn_more_markdown = ("""
+### License
+The service is a research preview intended for non-commercial use only, subject to the model [License](https://github.com/facebookresearch/llama/blob/main/MODEL_CARD.md) of LLaMA, [Terms of Use](https://openai.com/policies/terms-of-use) of the data generated by OpenAI, and [Privacy Practices](https://chrome.google.com/webstore/detail/sharegpt-share-your-chatg/daiacboceoaocpibfodeljbdfacokfjb) of ShareGPT. Please contact us if you find any potential violation.
+""")
+
+
+def create_frame_grid(img_array, interval_width=50):
+    n, h, w, c = img_array.shape
+    grid_size = int(np.ceil(np.sqrt(n)))
+
+    horizontal_band = np.ones((h, interval_width, c),
+                              dtype=img_array.dtype) * 255
+    vertical_band = np.ones((interval_width, w + (grid_size - 1)
+                            * (w + interval_width), c), dtype=img_array.dtype) * 255
+
+    rows = []
+    for i in range(grid_size):
+        row_frames = []
+        for j in range(grid_size):
+            idx = i * grid_size + j
+            if idx < n:
+                frame = img_array[idx]
+            else:
+                frame = np.ones_like(img_array[0]) * 255
+            if j > 0:
+                row_frames.append(horizontal_band)
+            row_frames.append(frame)
+        combined_row = np.concatenate(row_frames, axis=1)
+        if i > 0:
+            rows.append(vertical_band)
+        rows.append(combined_row)
+
+    final_grid = np.concatenate(rows, axis=0)
+    return final_grid
+
+
+def resize_image_grid(image, max_length=1920):
+    width, height = image.size
+    if max(width, height) > max_length:
+        if width > height:
+            scale = max_length / width
+        else:
+            scale = max_length / height
+
+        new_width = int(width * scale)
+        new_height = int(height * scale)
+
+        img_resized = image.resize((new_width, new_height), Image.BILINEAR)
+    else:
+        img_resized = image
+    return img_resized
+
+
+def get_index(num_frames, num_segments):
+    seg_size = float(num_frames - 1) / num_segments
+    start = int(seg_size / 2)
+    offsets = np.array([
+        start + int(np.round(seg_size * idx)) for idx in range(num_segments)
+    ])
+    return offsets
+
+
+def load_video(video_path, num_segments=8, return_msg=False, num_frames=4):
+    vr = VideoReader(video_path, ctx=cpu(0), num_threads=1)
+    num_frames = len(vr)
+    frame_indices = get_index(num_frames, num_segments)
+    img_array = vr.get_batch(frame_indices).asnumpy()
+    img_grid = create_frame_grid(img_array, 50)
+    img_grid = Image.fromarray(img_grid).convert("RGB")
+    img_grid = resize_image_grid(img_grid)
+    if return_msg:
+        fps = float(vr.get_avg_fps())
+        sec = ", ".join([str(round(f / fps, 1)) for f in frame_indices])
+        # " " should be added in the start and end
+        msg = f"The video contains {len(frame_indices)} frames sampled at {sec} seconds."
+        return img_grid, msg
+    else:
+        return img_grid
+
+
+def video_answer(prompt, model, processor, tokenizer, img_grid, do_sample=True,
+                 max_new_tokens=200, num_beams=1, top_p=0.9,
+                 temperature=1.0, print_res=False, **kwargs):
+    if not isinstance(img_grid, (list, tuple)):
+        img_grid = [img_grid]
+    image_size = img_grid[0].size
+    image_tensor = process_images(img_grid, processor, model.config)[0]
+    input_ids = tokenizer_image_token(
+        prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt')
+    input_ids = input_ids.unsqueeze(0).to(
+        device=model.device, non_blocking=True)
+    pad_token_id = tokenizer.pad_token_id if tokenizer.pad_token is not None else tokenizer.eos_token_id
+
+    with torch.inference_mode():
+        output_ids = model.generate(
+            input_ids,
+            images=image_tensor.to(
+                dtype=torch.float16, device=model.device, non_blocking=True),
+            image_sizes=[image_size],
+            do_sample=do_sample,
+            temperature=temperature,
+            top_p=top_p,
+            num_beams=num_beams,
+            max_new_tokens=max_new_tokens,
+            pad_token_id=pad_token_id,
+            use_cache=True,
+            **kwargs)
+        outputs = tokenizer.batch_decode(
+            output_ids, skip_special_tokens=True)[0].strip()
+    if print_res:  # debug usage
+        print('### PROMPTING LM WITH: ', prompt)
+        print('### LM OUTPUT TEXT:  ', outputs)
+
+    return outputs
+
+
+class Chat:
+    def __init__(self, model_path, conv_mode, model_base=None, load_8bit=False, load_4bit=False, device='cuda', num_frames=16):
+        disable_torch_init()
+        model_name = get_model_name_from_path(model_path)
+        self.tokenizer, self.model, self.processor, context_len = load_pretrained_model(
+            model_path, model_base, model_name,
+            load_8bit, load_4bit,
+            device=device)
+        self.model.eval()
+        self.conv_mode = conv_mode
+        self.device = self.model.device
+        self.num_frames = num_frames
+        self.pre_query_prompt = "The provided image arranges keyframes from a video in a grid view, keyframes are separated with white bands. Answer concisely with overall content and context of the video, highlighting any significant events, characters, or objects that appear throughout the frames."
+
+    def get_prompt(self, qs, state):
+        state.append_message(state.roles[0], qs)
+        state.append_message(state.roles[1], None)
+        return state
+
+    @torch.inference_mode()
+    def generate(self, vid_path: list, prompt: str, first_run: bool, state):
+        if self.num_frames != 0:
+            vid, msg = load_video(
+                vid_path, num_segments=self.num_frames, return_msg=True)
+        else:
+            vid, msg = None, 'num_frames is 0, not inputing image'
+        img_grid = vid
+        if self.pre_query_prompt is not None:
+            prompt = DEFAULT_IMAGE_TOKEN + '\n' + self.pre_query_prompt + prompt
+        else:
+            prompt = DEFAULT_IMAGE_TOKEN + '\n' + prompt
+        state = self.get_prompt(prompt, state)
+        prompt = state.get_prompt()
+        llm_response = video_answer(prompt, model=self.model, processor=self.processor, tokenizer=self.tokenizer,
+                                    do_sample=True, temperature=0.1, img_grid=img_grid, max_new_tokens=1024, print_res=True)
+        return llm_response, state

llava/train/llava_trainer.py

@@ -0,0 +1,301 @@
+import os
+import torch
+import torch.nn as nn
+
+from torch.utils.data import Sampler
+
+from transformers import Trainer
+from transformers.trainer import (
+    is_sagemaker_mp_enabled,
+    get_parameter_names,
+    has_length,
+    ALL_LAYERNORM_LAYERS,
+    logger,
+)
+from typing import List, Optional
+
+
+def maybe_zero_3(param, ignore_status=False, name=None):
+    from deepspeed import zero
+    from deepspeed.runtime.zero.partition_parameters import ZeroParamStatus
+    if hasattr(param, "ds_id"):
+        if param.ds_status == ZeroParamStatus.NOT_AVAILABLE:
+            if not ignore_status:
+                print(name, 'no ignore status')
+        with zero.GatheredParameters([param]):
+            param = param.data.detach().cpu().clone()
+    else:
+        param = param.detach().cpu().clone()
+    return param
+
+
+def get_mm_adapter_state_maybe_zero_3(named_params, keys_to_match):
+    to_return = {k: t for k, t in named_params if any(key_match in k for key_match in keys_to_match)}
+    to_return = {k: maybe_zero_3(v, ignore_status=True, name=k).cpu() for k, v in to_return.items()}
+    return to_return
+
+
+def split_to_even_chunks(indices, lengths, num_chunks):
+    """
+    Split a list of indices into `chunks` chunks of roughly equal lengths.
+    """
+
+    if len(indices) % num_chunks != 0:
+        return [indices[i::num_chunks] for i in range(num_chunks)]
+
+    num_indices_per_chunk = len(indices) // num_chunks
+
+    chunks = [[] for _ in range(num_chunks)]
+    chunks_lengths = [0 for _ in range(num_chunks)]
+    for index in indices:
+        shortest_chunk = chunks_lengths.index(min(chunks_lengths))
+        chunks[shortest_chunk].append(index)
+        chunks_lengths[shortest_chunk] += lengths[index]
+        if len(chunks[shortest_chunk]) == num_indices_per_chunk:
+            chunks_lengths[shortest_chunk] = float("inf")
+
+    return chunks
+
+
+def get_modality_length_grouped_indices(lengths, batch_size, world_size, generator=None):
+    # We need to use torch for the random part as a distributed sampler will set the random seed for torch.
+    assert all(l != 0 for l in lengths), "Should not have zero length."
+    if all(l > 0 for l in lengths) or all(l < 0 for l in lengths):
+        # all samples are in the same modality
+        return get_length_grouped_indices(lengths, batch_size, world_size, generator=generator)
+    mm_indices, mm_lengths = zip(*[(i, l) for i, l in enumerate(lengths) if l > 0])
+    lang_indices, lang_lengths = zip(*[(i, -l) for i, l in enumerate(lengths) if l < 0])
+
+    mm_shuffle = [mm_indices[i] for i in get_length_grouped_indices(mm_lengths, batch_size, world_size, generator=None)]
+    lang_shuffle = [lang_indices[i] for i in get_length_grouped_indices(lang_lengths, batch_size, world_size, generator=None)]
+    megabatch_size = world_size * batch_size
+    mm_megabatches = [mm_shuffle[i : i + megabatch_size] for i in range(0, len(mm_shuffle), megabatch_size)]
+    lang_megabatches = [lang_shuffle[i : i + megabatch_size] for i in range(0, len(lang_shuffle), megabatch_size)]
+
+    last_mm = mm_megabatches[-1]
+    last_lang = lang_megabatches[-1]
+    additional_batch = last_mm + last_lang
+    megabatches = mm_megabatches[:-1] + lang_megabatches[:-1]
+    megabatch_indices = torch.randperm(len(megabatches), generator=generator)
+    megabatches = [megabatches[i] for i in megabatch_indices]
+
+    if len(additional_batch) > 0:
+        megabatches.append(sorted(additional_batch))
+
+    return [i for megabatch in megabatches for i in megabatch]
+
+
+def get_length_grouped_indices(lengths, batch_size, world_size, generator=None, merge=True):
+    # We need to use torch for the random part as a distributed sampler will set the random seed for torch.
+    indices = torch.randperm(len(lengths), generator=generator)
+    megabatch_size = world_size * batch_size
+    megabatches = [indices[i : i + megabatch_size].tolist() for i in range(0, len(lengths), megabatch_size)]
+    megabatches = [sorted(megabatch, key=lambda i: lengths[i], reverse=True) for megabatch in megabatches]
+    megabatches = [split_to_even_chunks(megabatch, lengths, world_size) for megabatch in megabatches]
+
+    return [i for megabatch in megabatches for batch in megabatch for i in batch]
+
+
+class LengthGroupedSampler(Sampler):
+    r"""
+    Sampler that samples indices in a way that groups together features of the dataset of roughly the same length while
+    keeping a bit of randomness.
+    """
+
+    def __init__(
+        self,
+        batch_size: int,
+        world_size: int,
+        lengths: Optional[List[int]] = None,
+        generator=None,
+        group_by_modality: bool = False,
+    ):
+        if lengths is None:
+            raise ValueError("Lengths must be provided.")
+
+        self.batch_size = batch_size
+        self.world_size = world_size
+        self.lengths = lengths
+        self.generator = generator
+        self.group_by_modality = group_by_modality
+
+    def __len__(self):
+        return len(self.lengths)
+
+    def __iter__(self):
+        if self.group_by_modality:
+            indices = get_modality_length_grouped_indices(self.lengths, self.batch_size, self.world_size, generator=self.generator)
+        else:
+            indices = get_length_grouped_indices(self.lengths, self.batch_size, self.world_size, generator=self.generator)
+        return iter(indices)
+
+
+class LLaVATrainer(Trainer):
+
+    def _get_train_sampler(self) -> Optional[torch.utils.data.Sampler]:
+        if self.train_dataset is None or not has_length(self.train_dataset):
+            return None
+
+        if self.args.group_by_modality_length:
+            lengths = self.train_dataset.modality_lengths
+            return LengthGroupedSampler(
+                self.args.train_batch_size,
+                world_size=self.args.world_size * self.args.gradient_accumulation_steps,
+                lengths=lengths,
+                group_by_modality=True,
+            )
+        else:
+            return super()._get_train_sampler()
+
+    def create_optimizer(self):
+        """
+        Setup the optimizer.
+
+        We provide a reasonable default that works well. If you want to use something else, you can pass a tuple in the
+        Trainer's init through `optimizers`, or subclass and override this method in a subclass.
+        """
+        if is_sagemaker_mp_enabled():
+            return super().create_optimizer()
+
+        opt_model = self.model
+
+        if self.optimizer is None:
+            decay_parameters = get_parameter_names(opt_model, ALL_LAYERNORM_LAYERS)
+            decay_parameters = [name for name in decay_parameters if "bias" not in name]
+            if self.args.mm_projector_lr is not None:
+                projector_parameters = [name for name, _ in opt_model.named_parameters() if "mm_projector" in name]
+                if self.args.mm_vision_tower_lr is not None:
+                    vision_tower_parameters = [
+                        name for name, _ in opt_model.named_parameters() if "vision_tower" in name]
+                    optimizer_grouped_parameters = [
+                        {
+                            "params": [
+                                p for n, p in opt_model.named_parameters() if (n in decay_parameters and n not in projector_parameters and n not in vision_tower_parameters and p.requires_grad)
+                            ],
+                            "weight_decay": self.args.weight_decay,
+                        },
+                        {
+                            "params": [
+                                p for n, p in opt_model.named_parameters() if (n in decay_parameters and n not in projector_parameters and n in vision_tower_parameters and p.requires_grad)
+                            ],
+                            "weight_decay": self.args.weight_decay,
+                            "lr": self.args.mm_vision_tower_lr,
+                        },
+                        {
+                            "params": [
+                                p for n, p in opt_model.named_parameters() if (n not in decay_parameters and n not in projector_parameters and n not in vision_tower_parameters and p.requires_grad)
+                            ],
+                            "weight_decay": 0.0,
+                        },
+                        {
+                            "params": [
+                                p for n, p in opt_model.named_parameters() if (n not in decay_parameters and n not in projector_parameters and n in vision_tower_parameters and p.requires_grad)
+                            ],
+                            "weight_decay": 0.0,
+                            "lr": self.args.mm_vision_tower_lr,
+                        },
+                        {
+                            "params": [
+                                p for n, p in opt_model.named_parameters() if (n in decay_parameters and n in projector_parameters and p.requires_grad)
+                            ],
+                            "weight_decay": self.args.weight_decay,
+                            "lr": self.args.mm_projector_lr,
+                        },
+                        {
+                            "params": [
+                                p for n, p in opt_model.named_parameters() if (n not in decay_parameters and n in projector_parameters and p.requires_grad)
+                            ],
+                            "weight_decay": 0.0,
+                            "lr": self.args.mm_projector_lr,
+                        },
+                    ]
+                else:
+                    optimizer_grouped_parameters = [
+                        {
+                            "params": [
+                                p for n, p in opt_model.named_parameters() if (n in decay_parameters and n not in projector_parameters and p.requires_grad)
+                            ],
+                            "weight_decay": self.args.weight_decay,
+                        },
+                        {
+                            "params": [
+                                p for n, p in opt_model.named_parameters() if (n not in decay_parameters and n not in projector_parameters and p.requires_grad)
+                            ],
+                            "weight_decay": 0.0,
+                        },
+                        {
+                            "params": [
+                                p for n, p in opt_model.named_parameters() if (n in decay_parameters and n in projector_parameters and p.requires_grad)
+                            ],
+                            "weight_decay": self.args.weight_decay,
+                            "lr": self.args.mm_projector_lr,
+                        },
+                        {
+                            "params": [
+                                p for n, p in opt_model.named_parameters() if (n not in decay_parameters and n in projector_parameters and p.requires_grad)
+                            ],
+                            "weight_decay": 0.0,
+                            "lr": self.args.mm_projector_lr,
+                        },
+                    ]
+            else:
+                optimizer_grouped_parameters = [
+                    {
+                        "params": [
+                            p for n, p in opt_model.named_parameters() if (n in decay_parameters and p.requires_grad)
+                        ],
+                        "weight_decay": self.args.weight_decay,
+                    },
+                    {
+                        "params": [
+                            p for n, p in opt_model.named_parameters() if (n not in decay_parameters and p.requires_grad)
+                        ],
+                        "weight_decay": 0.0,
+                    },
+                ]
+
+            optimizer_cls, optimizer_kwargs = Trainer.get_optimizer_cls_and_kwargs(self.args)
+
+            self.optimizer = optimizer_cls(optimizer_grouped_parameters, **optimizer_kwargs)
+            if optimizer_cls.__name__ == "Adam8bit":
+                import bitsandbytes
+
+                manager = bitsandbytes.optim.GlobalOptimManager.get_instance()
+
+                skipped = 0
+                for module in opt_model.modules():
+                    if isinstance(module, nn.Embedding):
+                        skipped += sum({p.data_ptr(): p.numel() for p in module.parameters()}.values())
+                        logger.info(f"skipped {module}: {skipped/2**20}M params")
+                        manager.register_module_override(module, "weight", {"optim_bits": 32})
+                        logger.debug(f"bitsandbytes: will optimize {module} in fp32")
+                logger.info(f"skipped: {skipped/2**20}M params")
+
+        return self.optimizer
+
+    def _save_checkpoint(self, model, trial, metrics=None):
+        if getattr(self.args, 'tune_mm_mlp_adapter', False):
+            from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR
+            checkpoint_folder = f"{PREFIX_CHECKPOINT_DIR}-{self.state.global_step}"
+
+            run_dir = self._get_output_dir(trial=trial)
+            output_dir = os.path.join(run_dir, checkpoint_folder)
+
+            # Only save Adapter
+            keys_to_match = ['mm_projector', 'vision_resampler']
+            if getattr(self.args, "use_im_start_end", False):
+                keys_to_match.extend(['embed_tokens', 'embed_in'])
+
+            weight_to_save = get_mm_adapter_state_maybe_zero_3(self.model.named_parameters(), keys_to_match)
+
+            if self.args.local_rank == 0 or self.args.local_rank == -1:
+                self.model.config.save_pretrained(output_dir)
+                torch.save(weight_to_save, os.path.join(output_dir, f'mm_projector.bin'))
+        else:
+            super(LLaVATrainer, self)._save_checkpoint(model, trial, metrics)
+
+    def _save(self, output_dir: Optional[str] = None, state_dict=None):
+        if getattr(self.args, 'tune_mm_mlp_adapter', False):
+            pass
+        else:
+            super(LLaVATrainer, self)._save(output_dir, state_dict)

llava/train/train.py

@@ -0,0 +1,1323 @@
+# Adopted from https://github.com/lm-sys/FastChat. Below is the original copyright:
+# Adopted from tatsu-lab@stanford_alpaca. Below is the original copyright:
+#    Copyright 2023 Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li
+#
+#    Licensed under the Apache License, Version 2.0 (the "License");
+#    you may not use this file except in compliance with the License.
+#    You may obtain a copy of the License at
+#
+#        http://www.apache.org/licenses/LICENSE-2.0
+#
+#    Unless required by applicable law or agreed to in writing, software
+#    distributed under the License is distributed on an "AS IS" BASIS,
+#    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#    See the License for the specific language governing permissions and
+#    limitations under the License.
+import copy
+import json
+import logging
+import os
+import pathlib
+from dataclasses import dataclass, field
+from typing import Dict, List, Optional, Sequence
+
+import tokenizers
+import torch
+import transformers
+from packaging import version
+from PIL import Image
+from torch.utils.data import Dataset
+
+from llava import conversation as conversation_lib
+from llava.constants import (DEFAULT_IM_END_TOKEN, DEFAULT_IM_START_TOKEN,
+                             DEFAULT_IMAGE_TOKEN, IGNORE_INDEX,
+                             IMAGE_TOKEN_INDEX)
+from llava.mm_utils import process_anyres_image, tokenizer_image_token
+from llava.model import *
+from llava.train.llava_trainer import LLaVATrainer
+
+local_rank = None
+
+
+def rank0_print(*args):
+    if local_rank == 0:
+        print(*args)
+
+
+IS_TOKENIZER_GREATER_THAN_0_14 = version.parse(
+    tokenizers.__version__) >= version.parse('0.14')
+
+
+@dataclass
+class ModelArguments:
+    model_name_or_path: Optional[str] = field(default="facebook/opt-125m")
+    version: Optional[str] = field(default="v0")
+    freeze_backbone: bool = field(default=False)
+    tune_mm_mlp_adapter: bool = field(default=False)
+    vision_tower: Optional[str] = field(default=None)
+    mm_vision_select_layer: Optional[int] = field(
+        default=-1)   # default to the last layer
+    pretrain_mm_mlp_adapter: Optional[str] = field(default=None)
+    mm_projector_type: Optional[str] = field(default='linear')
+    mm_use_im_start_end: bool = field(default=False)
+    mm_use_im_patch_token: bool = field(default=True)
+    mm_patch_merge_type: Optional[str] = field(default='flat')
+    mm_vision_select_feature: Optional[str] = field(default="patch")
+
+
+@dataclass
+class DataArguments:
+    data_path: str = field(default=None,
+                           metadata={"help": "Path to the training data."})
+    lazy_preprocess: bool = False
+    is_multimodal: bool = False
+    image_folder: Optional[str] = field(default=None)
+    image_aspect_ratio: str = 'square'
+
+
+@dataclass
+class TrainingArguments(transformers.TrainingArguments):
+    cache_dir: Optional[str] = field(default=None)
+    optim: str = field(default="adamw_torch")
+    remove_unused_columns: bool = field(default=False)
+    freeze_mm_mlp_adapter: bool = field(default=False)
+    unfreeze_mm_vision_tower: bool = field(default=False)
+    mpt_attn_impl: Optional[str] = field(default="triton")
+    model_max_length: int = field(
+        default=512,
+        metadata={
+            "help":
+            "Maximum sequence length. Sequences will be right padded (and possibly truncated)."
+        },
+    )
+    double_quant: bool = field(
+        default=True,
+        metadata={
+            "help": "Compress the quantization statistics through double quantization."}
+    )
+    quant_type: str = field(
+        default="nf4",
+        metadata={
+            "help": "Quantization data type to use. Should be one of `fp4` or `nf4`."}
+    )
+    bits: int = field(
+        default=16,
+        metadata={"help": "How many bits to use."}
+    )
+    lora_enable: bool = False
+    lora_r: int = 64
+    lora_alpha: int = 16
+    lora_dropout: float = 0.05
+    lora_weight_path: str = ""
+    lora_bias: str = "none"
+    lora_qv_proj_only: bool = False
+    mm_projector_lr: Optional[float] = None
+    mm_vision_tower_lr: Optional[float] = None
+    group_by_modality_length: bool = field(default=False)
+
+
+def maybe_zero_3(param, ignore_status=False, name=None):
+    from deepspeed import zero
+    from deepspeed.runtime.zero.partition_parameters import ZeroParamStatus
+    if hasattr(param, "ds_id"):
+        if param.ds_status == ZeroParamStatus.NOT_AVAILABLE:
+            if not ignore_status:
+                logging.warning(
+                    f"{name}: param.ds_status != ZeroParamStatus.NOT_AVAILABLE: {param.ds_status}")
+        with zero.GatheredParameters([param]):
+            param = param.data.detach().cpu().clone()
+    else:
+        param = param.detach().cpu().clone()
+    return param
+
+
+# Borrowed from peft.utils.get_peft_model_state_dict
+def get_peft_state_maybe_zero_3(named_params, bias):
+    if bias == "none":
+        to_return = {k: t for k, t in named_params if "lora_" in k}
+    elif bias == "all":
+        to_return = {k: t for k,
+                     t in named_params if "lora_" in k or "bias" in k}
+    elif bias == "lora_only":
+        to_return = {}
+        maybe_lora_bias = {}
+        lora_bias_names = set()
+        for k, t in named_params:
+            if "lora_" in k:
+                to_return[k] = t
+                bias_name = k.split("lora_")[0] + "bias"
+                lora_bias_names.add(bias_name)
+            elif "bias" in k:
+                maybe_lora_bias[k] = t
+        for k, t in maybe_lora_bias:
+            if bias_name in lora_bias_names:
+                to_return[bias_name] = t
+    else:
+        raise NotImplementedError
+    to_return = {k: maybe_zero_3(v, ignore_status=True)
+                 for k, v in to_return.items()}
+    return to_return
+
+
+def get_peft_state_non_lora_maybe_zero_3(named_params, require_grad_only=True):
+    to_return = {k: t for k, t in named_params if "lora_" not in k}
+    if require_grad_only:
+        to_return = {k: t for k, t in to_return.items() if t.requires_grad}
+    to_return = {k: maybe_zero_3(v, ignore_status=True).cpu()
+                 for k, v in to_return.items()}
+    return to_return
+
+
+def get_mm_adapter_state_maybe_zero_3(named_params, keys_to_match):
+    to_return = {k: t for k, t in named_params if any(
+        key_match in k for key_match in keys_to_match)}
+    to_return = {k: maybe_zero_3(v, ignore_status=True).cpu()
+                 for k, v in to_return.items()}
+    return to_return
+
+
+def get_vision_tower_state_maybe_zero_3(named_params, keys_to_match=['']):
+    to_return = {k: t for k, t in named_params if any(
+        key_match in k for key_match in keys_to_match)}
+    to_return = {k: maybe_zero_3(v, ignore_status=True).cpu()
+                 for k, v in to_return.items()}
+    return to_return
+
+
+def find_all_linear_names(model, qv_proj_only=False):
+    if qv_proj_only:
+        rank0_print('Only add LoRA to QV proj')
+        return ['q_proj', 'v_proj']
+    cls = torch.nn.Linear
+    lora_module_names = set()
+    multimodal_keywords = ['mm_projector', 'vision_tower', 'vision_resampler']
+    for name, module in model.named_modules():
+        if any(mm_keyword in name for mm_keyword in multimodal_keywords):
+            continue
+        if isinstance(module, cls):
+            names = name.split('.')
+            lora_module_names.add(names[0] if len(names) == 1 else names[-1])
+
+    if 'lm_head' in lora_module_names:  # needed for 16-bit
+        lora_module_names.remove('lm_head')
+    return list(lora_module_names)
+
+
+def safe_save_model_for_hf_trainer(trainer: transformers.Trainer,
+                                   output_dir: str):
+    """Collects the state dict and dump to disk."""
+
+    if getattr(trainer.args, "tune_mm_mlp_adapter", False):
+        # Only save Adapter
+        keys_to_match = ['mm_projector']
+        if getattr(trainer.args, "use_im_start_end", False):
+            keys_to_match.extend(['embed_tokens', 'embed_in'])
+
+        weight_to_save = get_mm_adapter_state_maybe_zero_3(
+            trainer.model.named_parameters(), keys_to_match)
+        trainer.model.config.save_pretrained(output_dir)
+
+        current_folder = output_dir.split('/')[-1]
+        parent_folder = os.path.dirname(output_dir)
+        if trainer.args.local_rank == 0 or trainer.args.local_rank == -1:
+            if current_folder.startswith('checkpoint-'):
+                mm_projector_folder = os.path.join(
+                    parent_folder, "mm_projector")
+                os.makedirs(mm_projector_folder, exist_ok=True)
+                torch.save(weight_to_save, os.path.join(
+                    mm_projector_folder, f'{current_folder}.bin'))
+            else:
+                torch.save(weight_to_save, os.path.join(
+                    output_dir, f'mm_projector.bin'))
+        return
+
+    if getattr(trainer.args, "unfreeze_mm_vision_tower", False):
+        if trainer.deepspeed:
+            torch.cuda.synchronize()
+        mm_vision_tower_folder = os.path.join(output_dir, 'vision_tower')
+        os.makedirs(mm_vision_tower_folder, exist_ok=True)
+        trainer.model.get_vision_tower().image_processor.save_pretrained(mm_vision_tower_folder)
+        trainer.model.get_vision_tower().vision_tower.vision_model.config.save_pretrained(
+            mm_vision_tower_folder)
+        weight_to_save = get_vision_tower_state_maybe_zero_3(
+            trainer.model.get_vision_tower().vision_tower.named_parameters())
+        if trainer.args.local_rank == 0 or trainer.args.local_rank == -1:
+            torch.save(weight_to_save, os.path.join(
+                mm_vision_tower_folder, 'pytorch_model.bin'))
+
+    if getattr(trainer.model.model, 'vision_tower', None) is not None:
+        del trainer.model.model.vision_tower
+
+    if trainer.deepspeed:
+        torch.cuda.synchronize()
+        trainer.save_model(output_dir)
+        return
+
+    state_dict = trainer.model.state_dict()
+    if trainer.args.should_save:
+        cpu_state_dict = {
+            key: value.cpu()
+            for key, value in state_dict.items()
+        }
+        del state_dict
+        trainer._save(output_dir, state_dict=cpu_state_dict)  # noqa
+
+
+def smart_tokenizer_and_embedding_resize(
+    special_tokens_dict: Dict,
+    tokenizer: transformers.PreTrainedTokenizer,
+    model: transformers.PreTrainedModel,
+):
+    """Resize tokenizer and embedding.
+
+    Note: This is the unoptimized version that may make your embedding size not be divisible by 64.
+    """
+    num_new_tokens = tokenizer.add_special_tokens(special_tokens_dict)
+    model.resize_token_embeddings(len(tokenizer))
+
+    if num_new_tokens > 0:
+        input_embeddings = model.get_input_embeddings().weight.data
+        output_embeddings = model.get_output_embeddings().weight.data
+
+        input_embeddings_avg = input_embeddings[:-num_new_tokens].mean(
+            dim=0, keepdim=True)
+        output_embeddings_avg = output_embeddings[:-num_new_tokens].mean(
+            dim=0, keepdim=True)
+
+        input_embeddings[-num_new_tokens:] = input_embeddings_avg
+        output_embeddings[-num_new_tokens:] = output_embeddings_avg
+
+
+def _tokenize_fn(strings: Sequence[str],
+                 tokenizer: transformers.PreTrainedTokenizer) -> Dict:
+    """Tokenize a list of strings."""
+    tokenized_list = [
+        tokenizer(
+            text,
+            return_tensors="pt",
+            padding="longest",
+            max_length=tokenizer.model_max_length,
+            truncation=True,
+        ) for text in strings
+    ]
+    input_ids = labels = [
+        tokenized.input_ids[0] for tokenized in tokenized_list
+    ]
+    input_ids_lens = labels_lens = [
+        tokenized.input_ids.ne(tokenizer.pad_token_id).sum().item()
+        for tokenized in tokenized_list
+    ]
+    return dict(
+        input_ids=input_ids,
+        labels=labels,
+        input_ids_lens=input_ids_lens,
+        labels_lens=labels_lens,
+    )
+
+
+def _mask_targets(target, tokenized_lens, speakers):
+    # cur_idx = 0
+    cur_idx = tokenized_lens[0]
+    tokenized_lens = tokenized_lens[1:]
+    target[:cur_idx] = IGNORE_INDEX
+    for tokenized_len, speaker in zip(tokenized_lens, speakers):
+        if speaker == "human":
+            target[cur_idx+2:cur_idx + tokenized_len] = IGNORE_INDEX
+        cur_idx += tokenized_len
+
+
+def _add_speaker_and_signal(header, source, get_conversation=True):
+    """Add speaker and start/end signal on each round."""
+    BEGIN_SIGNAL = "### "
+    END_SIGNAL = "\n"
+    conversation = header
+    for sentence in source:
+        from_str = sentence["from"]
+        if from_str.lower() == "human":
+            from_str = conversation_lib.default_conversation.roles[0]
+        elif from_str.lower() == "gpt":
+            from_str = conversation_lib.default_conversation.roles[1]
+        else:
+            from_str = 'unknown'
+        sentence["value"] = (BEGIN_SIGNAL + from_str + ": " +
+                             sentence["value"] + END_SIGNAL)
+        if get_conversation:
+            conversation += sentence["value"]
+    conversation += BEGIN_SIGNAL
+    return conversation
+
+
+def preprocess_multimodal(
+    sources: Sequence[str],
+    data_args: DataArguments
+) -> Dict:
+    is_multimodal = data_args.is_multimodal
+    if not is_multimodal:
+        return sources
+
+    for source in sources:
+        for sentence in source:
+            if DEFAULT_IMAGE_TOKEN in sentence['value']:
+                sentence['value'] = sentence['value'].replace(
+                    DEFAULT_IMAGE_TOKEN, '').strip()
+                sentence['value'] = DEFAULT_IMAGE_TOKEN + \
+                    '\n' + sentence['value']
+                sentence['value'] = sentence['value'].strip()
+                if "mmtag" in conversation_lib.default_conversation.version:
+                    sentence['value'] = sentence['value'].replace(
+                        DEFAULT_IMAGE_TOKEN, '<Image>' + DEFAULT_IMAGE_TOKEN + '</Image>')
+            replace_token = DEFAULT_IMAGE_TOKEN
+            if data_args.mm_use_im_start_end:
+                replace_token = DEFAULT_IM_START_TOKEN + replace_token + DEFAULT_IM_END_TOKEN
+            sentence["value"] = sentence["value"].replace(
+                DEFAULT_IMAGE_TOKEN, replace_token)
+
+    return sources
+
+
+def preprocess_llama_2(
+    sources,
+    tokenizer: transformers.PreTrainedTokenizer,
+    has_image: bool = False
+) -> Dict:
+    conv = conversation_lib.default_conversation.copy()
+    roles = {"human": conv.roles[0], "gpt": conv.roles[1]}
+
+    # Apply prompt templates
+    conversations = []
+    for i, source in enumerate(sources):
+        if roles[source[0]["from"]] != conv.roles[0]:
+            # Skip the first one if it is not from human
+            source = source[1:]
+
+        conv.messages = []
+        for j, sentence in enumerate(source):
+            role = roles[sentence["from"]]
+            assert role == conv.roles[j % 2], f"{i}"
+            conv.append_message(role, sentence["value"])
+        conversations.append(conv.get_prompt())
+
+    # Tokenize conversations
+
+    if has_image:
+        input_ids = torch.stack([tokenizer_image_token(
+            prompt, tokenizer, return_tensors='pt') for prompt in conversations], dim=0)
+    else:
+        input_ids = tokenizer(
+            conversations,
+            return_tensors="pt",
+            padding="longest",
+            max_length=tokenizer.model_max_length,
+            truncation=True,
+        ).input_ids
+
+    targets = input_ids.clone()
+
+    assert conv.sep_style == conversation_lib.SeparatorStyle.LLAMA_2
+
+    # Mask targets
+    sep = "[/INST] "
+    for conversation, target in zip(conversations, targets):
+        total_len = int(target.ne(tokenizer.pad_token_id).sum())
+
+        rounds = conversation.split(conv.sep2)
+        cur_len = 1
+        target[:cur_len] = IGNORE_INDEX
+        for i, rou in enumerate(rounds):
+            if rou == "":
+                break
+
+            parts = rou.split(sep)
+            if len(parts) != 2:
+                break
+            parts[0] += sep
+
+            if has_image:
+                round_len = len(tokenizer_image_token(rou, tokenizer))
+                instruction_len = len(
+                    tokenizer_image_token(parts[0], tokenizer)) - 2
+            else:
+                round_len = len(tokenizer(rou).input_ids)
+                instruction_len = len(tokenizer(parts[0]).input_ids) - 2
+
+            target[cur_len: cur_len + instruction_len] = IGNORE_INDEX
+
+            cur_len += round_len
+        target[cur_len:] = IGNORE_INDEX
+
+        if cur_len < tokenizer.model_max_length:
+            if cur_len != total_len:
+                target[:] = IGNORE_INDEX
+                print(
+                    f"WARNING: tokenization mismatch: {cur_len} vs. {total_len}."
+                    f" (ignored)"
+                )
+
+    return dict(
+        input_ids=input_ids,
+        labels=targets,
+    )
+
+
+def preprocess_llama3(
+    sources,
+    tokenizer: transformers.PreTrainedTokenizer,
+    has_image: bool = False
+) -> Dict:
+    conv = conversation_lib.default_conversation.copy()
+    roles = {"human": conv.roles[0], "gpt": conv.roles[1]}
+
+    # Apply prompt templates
+    conversations = []
+    for i, source in enumerate(sources):
+        if roles[source[0]["from"]] != conv.roles[0]:
+            # Skip the first one if it is not from human
+            source = source[1:]
+
+        conv.messages = []
+        for j, sentence in enumerate(source):
+            role = roles[sentence["from"]]
+            assert role == conv.roles[j % 2], f"{i}"
+            conv.append_message(role, sentence["value"])
+        conversations.append(conv.get_prompt())
+
+    # Tokenize conversations
+
+    if has_image:
+        input_ids = torch.stack(
+            [tokenizer_image_token(prompt, tokenizer, return_tensors='pt') for prompt in conversations], dim=0)
+    else:
+        input_ids = tokenizer(
+            conversations,
+            return_tensors="pt",
+            padding="longest",
+            max_length=tokenizer.model_max_length,
+            truncation=True,
+        ).input_ids
+
+    targets = input_ids.clone()
+    assert conv.sep_style == conversation_lib.SeparatorStyle.MPT
+
+    # Mask targets
+    sep = conv.sep + conv.roles[1]
+    for conversation, target in zip(conversations, targets):
+        total_len = int(target.ne(tokenizer.pad_token_id).sum())
+
+        rounds = conversation.split(conv.sep)
+        re_rounds = [conv.sep.join(rounds[:3])]
+        for conv_idx in range(3, len(rounds), 2):
+            re_rounds.append(conv.sep.join(rounds[conv_idx:conv_idx + 2]))
+        cur_len = 0
+        target[:cur_len] = IGNORE_INDEX
+        for i, rou in enumerate(re_rounds):
+            if rou == "":
+                break
+
+            parts = rou.split(sep)
+            if len(parts) != 2:
+                break
+            parts[0] += sep
+
+            if has_image:
+                round_len = len(tokenizer_image_token(rou, tokenizer)) + 1
+                instruction_len = len(
+                    tokenizer_image_token(parts[0], tokenizer))
+            else:
+                round_len = len(tokenizer(rou).input_ids) + 1
+                instruction_len = len(tokenizer(parts[0]).input_ids)
+
+            if i > 0:
+                round_len -= 1
+                instruction_len -= 1
+
+            target[cur_len: cur_len + instruction_len] = IGNORE_INDEX
+
+            cur_len += round_len
+        target[cur_len:] = IGNORE_INDEX
+
+        if cur_len < tokenizer.model_max_length:
+            if cur_len != total_len:
+                target[:] = IGNORE_INDEX
+                print(
+                    f"WARNING: tokenization mismatch: {cur_len} vs. {total_len}."
+                    f" (ignored)"
+                )
+
+    return dict(
+        input_ids=input_ids,
+        labels=targets,
+    )
+
+
+def preprocess_yi(
+    sources,
+    tokenizer: transformers.PreTrainedTokenizer,
+    has_image: bool = False
+) -> Dict:
+    conv = conversation_lib.default_conversation.copy()
+    roles = {"human": conv.roles[0], "gpt": conv.roles[1]}
+
+    # Apply prompt templates
+    conversations = []
+    for i, source in enumerate(sources):
+        if roles[source[0]["from"]] != conv.roles[0]:
+            # Skip the first one if it is not from human
+            source = source[1:]
+
+        conv.messages = []
+        for j, sentence in enumerate(source):
+            role = roles[sentence["from"]]
+            assert role == conv.roles[j % 2], f"{i}"
+            conv.append_message(role, sentence["value"])
+        conversations.append(conv.get_prompt())
+
+    # Tokenize conversations
+
+    if has_image:
+        input_ids = torch.stack(
+            [tokenizer_image_token(prompt, tokenizer, return_tensors='pt') for prompt in conversations], dim=0)
+    else:
+        input_ids = tokenizer(
+            conversations,
+            return_tensors="pt",
+            padding="longest",
+            max_length=tokenizer.model_max_length,
+            truncation=True,
+        ).input_ids
+
+    targets = input_ids.clone()
+    assert conv.sep_style == conversation_lib.SeparatorStyle.MPT
+
+    # Mask targets
+    sep = conv.sep + conv.roles[1]
+    for conversation, target in zip(conversations, targets):
+        total_len = int(target.ne(tokenizer.pad_token_id).sum())
+
+        rounds = conversation.split(conv.sep)
+        re_rounds = [conv.sep.join(rounds[:3])]
+        for conv_idx in range(3, len(rounds), 2):
+            re_rounds.append(conv.sep.join(rounds[conv_idx:conv_idx + 2]))
+        cur_len = 0
+        target[:cur_len] = IGNORE_INDEX
+        for i, rou in enumerate(re_rounds):
+            if rou == "":
+                break
+
+            parts = rou.split(sep)
+            if len(parts) != 2:
+                break
+            parts[0] += sep
+
+            round_len = len(tokenizer_image_token(rou, tokenizer)) + len(tokenizer_image_token(conv.sep, tokenizer))
+            instruction_len = len(tokenizer_image_token(parts[0], tokenizer))
+            target[cur_len : cur_len + instruction_len] = IGNORE_INDEX
+
+            cur_len += round_len
+        target[cur_len:] = IGNORE_INDEX
+
+        if cur_len < tokenizer.model_max_length:
+            if cur_len != total_len:
+                target[:] = IGNORE_INDEX
+                print(
+                    f"WARNING: tokenization mismatch: {cur_len} vs. {total_len}."
+                    f" (ignored)"
+                )
+
+    return dict(
+        input_ids=input_ids,
+        labels=targets,
+    )
+
+
+def preprocess_v1(
+    sources,
+    tokenizer: transformers.PreTrainedTokenizer,
+    has_image: bool = False
+) -> Dict:
+    conv = conversation_lib.default_conversation.copy()
+    roles = {"human": conv.roles[0], "gpt": conv.roles[1]}
+
+    # Apply prompt templates
+    conversations = []
+    for i, source in enumerate(sources):
+        if roles[source[0]["from"]] != conv.roles[0]:
+            # Skip the first one if it is not from human
+            source = source[1:]
+
+        conv.messages = []
+        for j, sentence in enumerate(source):
+            role = roles[sentence["from"]]
+            assert role == conv.roles[j % 2], f"{i}"
+            conv.append_message(role, sentence["value"])
+        conversations.append(conv.get_prompt())
+
+    # Tokenize conversations
+
+    if has_image:
+        input_ids = torch.stack([tokenizer_image_token(
+            prompt, tokenizer, return_tensors='pt') for prompt in conversations], dim=0)
+    else:
+        input_ids = tokenizer(
+            conversations,
+            return_tensors="pt",
+            padding="longest",
+            max_length=tokenizer.model_max_length,
+            truncation=True,
+        ).input_ids
+
+    targets = input_ids.clone()
+
+    assert conv.sep_style == conversation_lib.SeparatorStyle.TWO
+
+    # Mask targets
+    sep = conv.sep + conv.roles[1] + ": "
+    for conversation, target in zip(conversations, targets):
+        total_len = int(target.ne(tokenizer.pad_token_id).sum())
+
+        rounds = conversation.split(conv.sep2)
+        cur_len = 1
+        target[:cur_len] = IGNORE_INDEX
+        for i, rou in enumerate(rounds):
+            if rou == "":
+                break
+
+            parts = rou.split(sep)
+            if len(parts) != 2:
+                break
+            parts[0] += sep
+
+            if has_image:
+                round_len = len(tokenizer_image_token(rou, tokenizer))
+                instruction_len = len(
+                    tokenizer_image_token(parts[0], tokenizer)) - 2
+            else:
+                round_len = len(tokenizer(rou).input_ids)
+                instruction_len = len(tokenizer(parts[0]).input_ids) - 2
+
+            if i != 0 and not tokenizer.legacy and IS_TOKENIZER_GREATER_THAN_0_14:
+                round_len -= 1
+                instruction_len -= 1
+
+            target[cur_len: cur_len + instruction_len] = IGNORE_INDEX
+
+            cur_len += round_len
+        target[cur_len:] = IGNORE_INDEX
+
+        if cur_len < tokenizer.model_max_length:
+            if cur_len != total_len:
+                target[:] = IGNORE_INDEX
+                print(
+                    f"WARNING: tokenization mismatch: {cur_len} vs. {total_len}."
+                    f" (ignored)"
+                )
+
+    return dict(
+        input_ids=input_ids,
+        labels=targets,
+    )
+
+
+def preprocess_mpt(
+    sources,
+    tokenizer: transformers.PreTrainedTokenizer,
+    has_image: bool = False
+) -> Dict:
+    conv = conversation_lib.default_conversation.copy()
+    roles = {"human": conv.roles[0], "gpt": conv.roles[1]}
+
+    # Apply prompt templates
+    conversations = []
+    for i, source in enumerate(sources):
+        if roles[source[0]["from"]] != conv.roles[0]:
+            # Skip the first one if it is not from human
+            source = source[1:]
+
+        conv.messages = []
+        for j, sentence in enumerate(source):
+            role = roles[sentence["from"]]
+            assert role == conv.roles[j % 2], f"{i}"
+            conv.append_message(role, sentence["value"])
+        conversations.append(conv.get_prompt())
+
+    # Tokenize conversations
+
+    if has_image:
+        input_ids = torch.stack([tokenizer_image_token(
+            prompt, tokenizer, return_tensors='pt') for prompt in conversations], dim=0)
+    else:
+        input_ids = tokenizer(
+            conversations,
+            return_tensors="pt",
+            padding="longest",
+            max_length=tokenizer.model_max_length,
+            truncation=True,
+        ).input_ids
+
+    targets = input_ids.clone()
+    assert conv.sep_style == conversation_lib.SeparatorStyle.MPT
+
+    # Mask targets
+    sep = conv.sep + conv.roles[1]
+    for conversation, target in zip(conversations, targets):
+        total_len = int(target.ne(tokenizer.pad_token_id).sum())
+
+        rounds = conversation.split(conv.sep)
+        re_rounds = [conv.sep.join(rounds[:3])]  # system + user + gpt
+        for conv_idx in range(3, len(rounds), 2):
+            re_rounds.append(conv.sep.join(
+                rounds[conv_idx:conv_idx+2]))    # user + gpt
+        cur_len = 0
+        target[:cur_len] = IGNORE_INDEX
+        for i, rou in enumerate(re_rounds):
+            if rou == "":
+                break
+
+            parts = rou.split(sep)
+            if len(parts) != 2:
+                break
+            parts[0] += sep
+            # not included <|im_end|>
+            if has_image:
+                round_len = len(tokenizer_image_token(rou, tokenizer))
+                instruction_len = len(
+                    tokenizer_image_token(parts[0], tokenizer)) - 1
+            else:
+                round_len = len(tokenizer(rou).input_ids)
+                instruction_len = len(tokenizer(parts[0]).input_ids) - 1
+
+            # include <|im_end|> for all rounds
+            # if i != 0 and getattr(tokenizer, 'legacy', False) and IS_TOKENIZER_GREATER_THAN_0_14:
+            if getattr(tokenizer, 'legacy', False) and IS_TOKENIZER_GREATER_THAN_0_14:
+                round_len += 1
+                instruction_len += 1
+
+            target[cur_len: cur_len + instruction_len] = IGNORE_INDEX
+
+            cur_len += round_len
+        target[cur_len:] = IGNORE_INDEX
+
+        if cur_len < tokenizer.model_max_length:
+            if cur_len != total_len:
+                target[:] = IGNORE_INDEX
+                print(
+                    f"WARNING: tokenization mismatch: {cur_len} vs. {total_len}."
+                    f" (ignored)"
+                )
+
+    return dict(
+        input_ids=input_ids,
+        labels=targets,
+    )
+
+
+def preprocess_plain(
+    sources: Sequence[str],
+    tokenizer: transformers.PreTrainedTokenizer,
+) -> Dict:
+    # add end signal and concatenate together
+    conversations = []
+    for source in sources:
+        assert len(source) == 2
+        assert DEFAULT_IMAGE_TOKEN in source[0]['value']
+        source[0]['value'] = DEFAULT_IMAGE_TOKEN
+        conversation = source[0]['value'] + source[1]['value'] + \
+            conversation_lib.default_conversation.sep
+        conversations.append(conversation)
+    # tokenize conversations
+    input_ids = [tokenizer_image_token(
+        prompt, tokenizer, return_tensors='pt') for prompt in conversations]
+    targets = copy.deepcopy(input_ids)
+    for target, source in zip(targets, sources):
+        tokenized_len = len(tokenizer_image_token(
+            source[0]['value'], tokenizer))
+        target[:tokenized_len] = IGNORE_INDEX
+
+    return dict(input_ids=input_ids, labels=targets)
+
+
+def preprocess(
+    sources: Sequence[str],
+    tokenizer: transformers.PreTrainedTokenizer,
+    has_image: bool = False
+) -> Dict:
+    """
+    Given a list of sources, each is a conversation list. This transform:
+    1. Add signal '### ' at the beginning each sentence, with end signal '\n';
+    2. Concatenate conversations together;
+    3. Tokenize the concatenated conversation;
+    4. Make a deepcopy as the target. Mask human words with IGNORE_INDEX.
+    """
+    if conversation_lib.default_conversation.sep_style == conversation_lib.SeparatorStyle.PLAIN:
+        return preprocess_plain(sources, tokenizer)
+    if conversation_lib.default_conversation.sep_style == conversation_lib.SeparatorStyle.LLAMA_2:
+        return preprocess_llama_2(sources, tokenizer, has_image=has_image)
+    if conversation_lib.default_conversation.version.startswith("v1"):
+        return preprocess_v1(sources, tokenizer, has_image=has_image)
+    if conversation_lib.default_conversation.version == "mpt":
+        return preprocess_mpt(sources, tokenizer, has_image=has_image)
+    if conversation_lib.default_conversation.version in ["llama3", "llava_llama_3"]:
+        return preprocess_llama3(sources, tokenizer, has_image=has_image)
+    if conversation_lib.default_conversation.version == "yi":
+        return preprocess_yi(sources, tokenizer, has_image=has_image)
+    # add end signal and concatenate together
+    conversations = []
+    for source in sources:
+        header = f"{conversation_lib.default_conversation.system}\n\n"
+        conversation = _add_speaker_and_signal(header, source)
+        conversations.append(conversation)
+    # tokenize conversations
+
+    def get_tokenize_len(prompts):
+        return [len(tokenizer_image_token(prompt, tokenizer)) for prompt in prompts]
+
+    if has_image:
+        input_ids = [tokenizer_image_token(
+            prompt, tokenizer, return_tensors='pt') for prompt in conversations]
+    else:
+        conversations_tokenized = _tokenize_fn(conversations, tokenizer)
+        input_ids = conversations_tokenized["input_ids"]
+
+    targets = copy.deepcopy(input_ids)
+    for target, source in zip(targets, sources):
+        if has_image:
+            tokenized_lens = get_tokenize_len(
+                [header] + [s["value"] for s in source])
+        else:
+            tokenized_lens = _tokenize_fn(
+                [header] + [s["value"] for s in source], tokenizer)["input_ids_lens"]
+        speakers = [sentence["from"] for sentence in source]
+        _mask_targets(target, tokenized_lens, speakers)
+
+    return dict(input_ids=input_ids, labels=targets)
+
+
+class LazySupervisedDataset(Dataset):
+    """Dataset for supervised fine-tuning."""
+
+    def __init__(self, data_path: str,
+                 tokenizer: transformers.PreTrainedTokenizer,
+                 data_args: DataArguments):
+        super(LazySupervisedDataset, self).__init__()
+        list_data_dict = json.load(open(data_path, "r"))
+
+        rank0_print("Formatting inputs...Skip in lazy mode")
+        self.tokenizer = tokenizer
+        self.list_data_dict = list_data_dict
+        self.data_args = data_args
+
+    def __len__(self):
+        return len(self.list_data_dict)
+
+    @property
+    def lengths(self):
+        length_list = []
+        for sample in self.list_data_dict:
+            img_tokens = 128 if 'image' in sample else 0
+            length_list.append(sum(len(conv['value'].split())
+                               for conv in sample['conversations']) + img_tokens)
+        return length_list
+
+    @property
+    def modality_lengths(self):
+        length_list = []
+        for sample in self.list_data_dict:
+            cur_len = sum(len(conv['value'].split())
+                          for conv in sample['conversations'])
+            cur_len = cur_len if 'image' in sample else -cur_len
+            length_list.append(cur_len)
+        return length_list
+
+    def __getitem__(self, i) -> Dict[str, torch.Tensor]:
+        sources = self.list_data_dict[i]
+        if isinstance(i, int):
+            sources = [sources]
+        assert len(
+            sources) == 1, "Don't know why it is wrapped to a list"  # FIXME
+        if 'image' in sources[0]:
+            image_file = self.list_data_dict[i]['image']
+            image_folder = self.data_args.image_folder
+            processor = self.data_args.image_processor
+            image = Image.open(os.path.join(
+                image_folder, image_file)).convert('RGB')
+            if self.data_args.image_aspect_ratio == 'pad':
+                def expand2square(pil_img, background_color):
+                    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
+
+                image = expand2square(image, tuple(int(x * 255)
+                                      for x in processor.image_mean))
+                image_size = image.size
+                image = processor.preprocess(image, return_tensors='pt')[
+                    'pixel_values'][0]
+            elif self.data_args.image_aspect_ratio == "anyres":
+                image_size = image.size
+                image = process_anyres_image(
+                    image, processor, self.data_args.image_grid_pinpoints)
+            else:
+                image_size = image.size
+                image = processor.preprocess(image, return_tensors='pt')[
+                    'pixel_values'][0]
+            sources = preprocess_multimodal(
+                copy.deepcopy([e["conversations"] for e in sources]),
+                self.data_args)
+        else:
+            sources = copy.deepcopy([e["conversations"] for e in sources])
+        data_dict = preprocess(
+            sources,
+            self.tokenizer,
+            has_image=('image' in self.list_data_dict[i]))
+        if isinstance(i, int):
+            data_dict = dict(input_ids=data_dict["input_ids"][0],
+                             labels=data_dict["labels"][0])
+
+        # image exist in the data
+        if 'image' in self.list_data_dict[i]:
+            data_dict['image'] = image
+            data_dict['image_size'] = image_size
+        elif self.data_args.is_multimodal:
+            # image does not exist in the data, but the model is multimodal
+            crop_size = self.data_args.image_processor.crop_size
+            data_dict['image'] = torch.zeros(
+                3, crop_size['height'], crop_size['width'])
+            data_dict['image_size'] = (crop_size['height'], crop_size['width'])
+        return data_dict
+
+
+@dataclass
+class DataCollatorForSupervisedDataset(object):
+    """Collate examples for supervised fine-tuning."""
+
+    tokenizer: transformers.PreTrainedTokenizer
+
+    def __call__(self, instances: Sequence[Dict]) -> Dict[str, torch.Tensor]:
+        input_ids, labels = tuple([instance[key] for instance in instances]
+                                  for key in ("input_ids", "labels"))
+        input_ids = torch.nn.utils.rnn.pad_sequence(
+            input_ids,
+            batch_first=True,
+            padding_value=self.tokenizer.pad_token_id)
+        labels = torch.nn.utils.rnn.pad_sequence(labels,
+                                                 batch_first=True,
+                                                 padding_value=IGNORE_INDEX)
+        input_ids = input_ids[:, :self.tokenizer.model_max_length]
+        labels = labels[:, :self.tokenizer.model_max_length]
+        batch = dict(
+            input_ids=input_ids,
+            labels=labels,
+            attention_mask=input_ids.ne(self.tokenizer.pad_token_id),
+        )
+
+        if 'image' in instances[0]:
+            images = [instance['image'] for instance in instances]
+            image_sizes = [instance['image_size'] for instance in instances]
+            if all(x is not None and x.shape == images[0].shape for x in images):
+                batch['images'] = torch.stack(images)
+            else:
+                batch['images'] = images
+            batch['image_sizes'] = image_sizes
+
+        return batch
+
+
+def make_supervised_data_module(tokenizer: transformers.PreTrainedTokenizer,
+                                data_args) -> Dict:
+    """Make dataset and collator for supervised fine-tuning."""
+    train_dataset = LazySupervisedDataset(tokenizer=tokenizer,
+                                          data_path=data_args.data_path,
+                                          data_args=data_args)
+    data_collator = DataCollatorForSupervisedDataset(tokenizer=tokenizer)
+    return dict(train_dataset=train_dataset,
+                eval_dataset=None,
+                data_collator=data_collator)
+
+
+def unfreeze_vit(vision_tower):
+    for _, p in vision_tower.named_parameters():
+        p.requires_grad = True
+
+
+def format_bytes(size):
+    billion = 10**9
+    million = 10**6
+
+    if size >= billion:
+        return f"{size / billion:.2f}B"
+    elif size >= million:
+        return f"{size / million:.2f}M"
+    else:
+        return f"{size} bytes"
+
+
+def train(attn_implementation=None):
+    global local_rank
+
+    parser = transformers.HfArgumentParser(
+        (ModelArguments, DataArguments, TrainingArguments))
+    model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+    local_rank = training_args.local_rank
+    compute_dtype = (torch.float16 if training_args.fp16 else (
+        torch.bfloat16 if training_args.bf16 else torch.float32))
+
+    bnb_model_from_pretrained_args = {}
+    if training_args.bits in [4, 8]:
+        from transformers import BitsAndBytesConfig
+        bnb_model_from_pretrained_args.update(dict(
+            device_map={"": training_args.device},
+            load_in_4bit=training_args.bits == 4,
+            load_in_8bit=training_args.bits == 8,
+            quantization_config=BitsAndBytesConfig(
+                load_in_4bit=training_args.bits == 4,
+                load_in_8bit=training_args.bits == 8,
+                llm_int8_skip_modules=["mm_projector"],
+                llm_int8_threshold=6.0,
+                llm_int8_has_fp16_weight=False,
+                bnb_4bit_compute_dtype=compute_dtype,
+                bnb_4bit_use_double_quant=training_args.double_quant,
+                bnb_4bit_quant_type=training_args.quant_type  # {'fp4', 'nf4'}
+            )
+        ))
+    model_max_length_args = {}
+    if 'llava-v1.6-8b' not in model_args.model_name_or_path:
+        config = transformers.AutoConfig.from_pretrained(
+            model_args.model_name_or_path, trust_remote_code=True)
+        if config.max_position_embeddings < training_args.model_max_length:
+            rank0_print(
+                f'Set the max_position_embeddings from {config.max_position_embeddings} to {training_args.model_max_length}')
+            model_max_length_args.update(
+                {'max_position_embeddings': training_args.model_max_length})
+    if model_args.vision_tower is not None:
+        if 'mpt' in model_args.model_name_or_path:
+            config = transformers.AutoConfig.from_pretrained(
+                model_args.model_name_or_path, trust_remote_code=True)
+            config.attn_config['attn_impl'] = training_args.mpt_attn_impl
+            model = LlavaMptForCausalLM.from_pretrained(
+                model_args.model_name_or_path,
+                config=config,
+                cache_dir=training_args.cache_dir,
+                **bnb_model_from_pretrained_args
+            )
+        else:
+            model = LlavaLlamaForCausalLM.from_pretrained(
+                model_args.model_name_or_path,
+                cache_dir=training_args.cache_dir,
+                attn_implementation=attn_implementation,
+                torch_dtype=(torch.bfloat16 if training_args.bf16 else None),
+                **bnb_model_from_pretrained_args,
+                **model_max_length_args
+            )
+    else:
+        model = transformers.LlamaForCausalLM.from_pretrained(
+            model_args.model_name_or_path,
+            cache_dir=training_args.cache_dir,
+            attn_implementation=attn_implementation,
+            torch_dtype=(torch.bfloat16 if training_args.bf16 else None),
+            **bnb_model_from_pretrained_args
+        )
+    model.config.use_cache = False
+
+    if model_args.freeze_backbone:
+        model.model.requires_grad_(False)
+
+    if training_args.bits in [4, 8]:
+        from peft import prepare_model_for_kbit_training
+        model.config.torch_dtype = (torch.float32 if training_args.fp16 else (
+            torch.bfloat16 if training_args.bf16 else torch.float32))
+        model = prepare_model_for_kbit_training(
+            model, use_gradient_checkpointing=training_args.gradient_checkpointing)
+
+    if training_args.gradient_checkpointing:
+        if hasattr(model, "enable_input_require_grads"):
+            model.enable_input_require_grads()
+        else:
+            def make_inputs_require_grad(module, input, output):
+                output.requires_grad_(True)
+            model.get_input_embeddings().register_forward_hook(make_inputs_require_grad)
+
+    if training_args.lora_enable:
+        from peft import LoraConfig, get_peft_model
+        lora_config = LoraConfig(
+            r=training_args.lora_r,
+            lora_alpha=training_args.lora_alpha,
+            target_modules=find_all_linear_names(model, training_args.lora_qv_proj_only),
+            lora_dropout=training_args.lora_dropout,
+            bias=training_args.lora_bias,
+            task_type="CAUSAL_LM",
+        )
+        if training_args.bits == 16:
+            if training_args.bf16:
+                model.to(torch.bfloat16)
+            if training_args.fp16:
+                model.to(torch.float16)
+        rank0_print("Adding LoRA adapters...")
+        model = get_peft_model(model, lora_config)
+
+    if 'mpt' in model_args.model_name_or_path:
+        tokenizer = transformers.AutoTokenizer.from_pretrained(
+            model_args.model_name_or_path,
+            cache_dir=training_args.cache_dir,
+            model_max_length=training_args.model_max_length,
+            padding_side="right"
+        )
+    else:
+        tokenizer = transformers.AutoTokenizer.from_pretrained(
+            model_args.model_name_or_path,
+            cache_dir=training_args.cache_dir,
+            model_max_length=training_args.model_max_length,
+            padding_side="right",
+            use_fast=False,
+        )
+
+    if model_args.version == "v0":
+        if tokenizer.pad_token is None:
+            smart_tokenizer_and_embedding_resize(
+                special_tokens_dict=dict(pad_token="[PAD]"),
+                tokenizer=tokenizer,
+                model=model,
+            )
+    elif model_args.version == "v0.5":
+        tokenizer.pad_token = tokenizer.unk_token
+    else:
+        if tokenizer.pad_token is None:
+            rank0_print("Adding pad token as '<pad>'")
+            smart_tokenizer_and_embedding_resize(
+                special_tokens_dict=dict(pad_token="<pad>"),
+                tokenizer=tokenizer,
+                model=model,
+            )
+        if model_args.version in conversation_lib.conv_templates:
+            conversation_lib.default_conversation = conversation_lib.conv_templates[
+                model_args.version]
+        else:
+            conversation_lib.default_conversation = conversation_lib.conv_templates["vicuna_v1"]
+
+    if model_args.vision_tower is not None:
+        model.get_model().initialize_vision_modules(
+            model_args=model_args,
+            fsdp=training_args.fsdp
+        )
+
+        vision_tower = model.get_vision_tower()
+        vision_tower.to(
+            dtype=torch.bfloat16 if training_args.bf16 else torch.float16, device=training_args.device)
+
+        data_args.image_processor = vision_tower.image_processor
+        data_args.is_multimodal = True
+
+        model.config.image_aspect_ratio = data_args.image_aspect_ratio
+        if data_args.image_aspect_ratio == 'anyres':
+            base_size = vision_tower.config.image_size
+            grids = [[1, 2], [2, 1], [2, 2], [3, 1], [1, 3]]
+            model.config.image_grid_pinpoints = data_args.image_grid_pinpoints = [
+                [g[0]*base_size, g[1]*base_size] for g in grids]
+        model.config.tokenizer_padding_side = tokenizer.padding_side
+        model.config.tokenizer_model_max_length = tokenizer.model_max_length
+
+        model.config.tune_mm_mlp_adapter = training_args.tune_mm_mlp_adapter = model_args.tune_mm_mlp_adapter
+        if model_args.tune_mm_mlp_adapter:
+            model.requires_grad_(False)
+            for p in model.get_model().mm_projector.parameters():
+                p.requires_grad = True
+
+        model.config.freeze_mm_mlp_adapter = training_args.freeze_mm_mlp_adapter
+        if training_args.freeze_mm_mlp_adapter:
+            for p in model.get_model().mm_projector.parameters():
+                p.requires_grad = False
+
+        model.config.unfreeze_mm_vision_tower = training_args.unfreeze_mm_vision_tower
+        if training_args.unfreeze_mm_vision_tower:
+            lr_of_vit = training_args.mm_vision_tower_lr if training_args.mm_vision_tower_lr is not None else training_args.learning_rate
+            lr_of_mlp = training_args.mm_projector_lr if training_args.mm_projector_lr is not None else training_args.learning_rate
+            training_args.mm_projector_lr = lr_of_mlp
+            unfreeze_vit(vision_tower)
+            rank0_print(
+                f'Tune the entire model! The LR of ViT is {lr_of_vit}. The LR of MLP is {lr_of_mlp}. The LR of LLM is {training_args.learning_rate}')
+
+        # Calculate total parameters and trainable parameters
+        total_params = sum(p.numel() for p in model.get_model().parameters())
+        trainable_params = sum(
+            p.numel() for p in model.get_model().parameters() if p.requires_grad)
+
+        rank0_print(f"Total parameters: {format_bytes(total_params)}")
+        rank0_print(f"Trainable parameters: {format_bytes(trainable_params)}")
+
+        if training_args.bits in [4, 8]:
+            model.get_model().mm_projector.to(dtype=compute_dtype, device=training_args.device)
+
+        model.config.mm_use_im_start_end = data_args.mm_use_im_start_end = model_args.mm_use_im_start_end
+        model.config.mm_projector_lr = training_args.mm_projector_lr
+        model.config.mm_vision_tower_lr = training_args.mm_vision_tower_lr
+        training_args.use_im_start_end = model_args.mm_use_im_start_end
+        model.config.mm_use_im_patch_token = model_args.mm_use_im_patch_token
+        model.initialize_vision_tokenizer(model_args, tokenizer=tokenizer)
+        model.config.pad_token_id = tokenizer.pad_token_id
+
+    if training_args.bits in [4, 8]:
+        from peft.tuners.lora import LoraLayer
+        for name, module in model.named_modules():
+            if isinstance(module, LoraLayer):
+                if training_args.bf16:
+                    module = module.to(torch.bfloat16)
+            if 'norm' in name:
+                module = module.to(torch.float32)
+            if 'lm_head' in name or 'embed_tokens' in name:
+                if hasattr(module, 'weight'):
+                    if training_args.bf16 and module.weight.dtype == torch.float32:
+                        module = module.to(torch.bfloat16)
+
+    data_module = make_supervised_data_module(tokenizer=tokenizer,
+                                              data_args=data_args)
+    trainer = LLaVATrainer(model=model,
+                           tokenizer=tokenizer,
+                           args=training_args,
+                           **data_module)
+
+    if list(pathlib.Path(training_args.output_dir).glob("checkpoint-*")):
+        trainer.train(resume_from_checkpoint=True)
+    else:
+        trainer.train()
+    trainer.save_state()
+
+    model.config.use_cache = True
+
+    if training_args.lora_enable:
+        state_dict = get_peft_state_maybe_zero_3(
+            model.named_parameters(), training_args.lora_bias
+        )
+        non_lora_state_dict = get_peft_state_non_lora_maybe_zero_3(
+            model.named_parameters()
+        )
+        if training_args.local_rank == 0 or training_args.local_rank == -1:
+            model.config.save_pretrained(training_args.output_dir)
+            model.save_pretrained(
+                training_args.output_dir, state_dict=state_dict)
+            torch.save(non_lora_state_dict, os.path.join(
+                training_args.output_dir, 'non_lora_trainables.bin'))
+            if training_args.unfreeze_mm_vision_tower:
+                if trainer.deepspeed:
+                    torch.cuda.synchronize()
+                trainer.model.get_vision_tower().image_processor.save_pretrained(
+                    os.path.join(training_args.output_dir, 'vision_tower'))
+                trainer.model.get_vision_tower().vision_tower.vision_model.config.save_pretrained(
+                    os.path.join(training_args.output_dir, 'vision_tower'))
+                weight_to_save = get_vision_tower_state_maybe_zero_3(
+                    trainer.model.get_vision_tower().vision_tower.named_parameters())
+                torch.save(weight_to_save, os.path.join(
+                    training_args.output_dir, 'vision_tower/pytorch_model.bin'))
+    else:
+        safe_save_model_for_hf_trainer(trainer=trainer,
+                                       output_dir=training_args.output_dir)
+
+
+if __name__ == "__main__":
+    train()

llava/train/train_mem.py

@@ -0,0 +1,4 @@
+from llava.train.train import train
+
+if __name__ == "__main__":
+    train(attn_implementation="flash_attention_2")

llava/utils.py

@@ -0,0 +1,126 @@
+import datetime
+import logging
+import logging.handlers
+import os
+import sys
+
+import requests
+
+from llava.constants import LOGDIR
+
+server_error_msg = "**NETWORK ERROR DUE TO HIGH TRAFFIC. PLEASE REGENERATE OR REFRESH THIS PAGE.**"
+moderation_msg = "YOUR INPUT VIOLATES OUR CONTENT MODERATION GUIDELINES. PLEASE TRY AGAIN."
+
+handler = None
+
+
+def build_logger(logger_name, logger_filename):
+    global handler
+
+    formatter = logging.Formatter(
+        fmt="%(asctime)s | %(levelname)s | %(name)s | %(message)s",
+        datefmt="%Y-%m-%d %H:%M:%S",
+    )
+
+    # Set the format of root handlers
+    if not logging.getLogger().handlers:
+        logging.basicConfig(level=logging.INFO)
+    logging.getLogger().handlers[0].setFormatter(formatter)
+
+    # Redirect stdout and stderr to loggers
+    stdout_logger = logging.getLogger("stdout")
+    stdout_logger.setLevel(logging.INFO)
+    sl = StreamToLogger(stdout_logger, logging.INFO)
+    sys.stdout = sl
+
+    stderr_logger = logging.getLogger("stderr")
+    stderr_logger.setLevel(logging.ERROR)
+    sl = StreamToLogger(stderr_logger, logging.ERROR)
+    sys.stderr = sl
+
+    # Get logger
+    logger = logging.getLogger(logger_name)
+    logger.setLevel(logging.INFO)
+
+    # Add a file handler for all loggers
+    if handler is None:
+        os.makedirs(LOGDIR, exist_ok=True)
+        filename = os.path.join(LOGDIR, logger_filename)
+        handler = logging.handlers.TimedRotatingFileHandler(
+            filename, when='D', utc=True, encoding='UTF-8')
+        handler.setFormatter(formatter)
+
+        for name, item in logging.root.manager.loggerDict.items():
+            if isinstance(item, logging.Logger):
+                item.addHandler(handler)
+
+    return logger
+
+
+class StreamToLogger(object):
+    """
+    Fake file-like stream object that redirects writes to a logger instance.
+    """
+    def __init__(self, logger, log_level=logging.INFO):
+        self.terminal = sys.stdout
+        self.logger = logger
+        self.log_level = log_level
+        self.linebuf = ''
+
+    def __getattr__(self, attr):
+        return getattr(self.terminal, attr)
+
+    def write(self, buf):
+        temp_linebuf = self.linebuf + buf
+        self.linebuf = ''
+        for line in temp_linebuf.splitlines(True):
+            # From the io.TextIOWrapper docs:
+            #   On output, if newline is None, any '\n' characters written
+            #   are translated to the system default line separator.
+            # By default sys.stdout.write() expects '\n' newlines and then
+            # translates them so this is still cross platform.
+            if line[-1] == '\n':
+                self.logger.log(self.log_level, line.rstrip())
+            else:
+                self.linebuf += line
+
+    def flush(self):
+        if self.linebuf != '':
+            self.logger.log(self.log_level, self.linebuf.rstrip())
+        self.linebuf = ''
+
+
+def disable_torch_init():
+    """
+    Disable the redundant torch default initialization to accelerate model creation.
+    """
+    import torch
+    setattr(torch.nn.Linear, "reset_parameters", lambda self: None)
+    setattr(torch.nn.LayerNorm, "reset_parameters", lambda self: None)
+
+
+def violates_moderation(text):
+    """
+    Check whether the text violates OpenAI moderation API.
+    """
+    url = "https://api.openai.com/v1/moderations"
+    headers = {"Content-Type": "application/json",
+               "Authorization": "Bearer " + os.environ["OPENAI_API_KEY"]}
+    text = text.replace("\n", "")
+    data = "{" + '"input": ' + f'"{text}"' + "}"
+    data = data.encode("utf-8")
+    try:
+        ret = requests.post(url, headers=headers, data=data, timeout=5)
+        flagged = ret.json()["results"][0]["flagged"]
+    except requests.exceptions.RequestException as e:
+        flagged = False
+    except KeyError as e:
+        flagged = False
+
+    return flagged
+
+
+def pretty_print_semaphore(semaphore):
+    if semaphore is None:
+        return "None"
+    return f"Semaphore(value={semaphore._value}, locked={semaphore.locked()})"

llava/video_utils.py

@@ -0,0 +1,151 @@
+import argparse
+import io
+import json
+import os
+import random
+import tempfile
+from multiprocessing import Manager, Pool, cpu_count
+
+import cv2
+import imageio
+import numpy as np
+from decord import VideoReader
+from PIL import Image
+
+
+def get_frame_indices(num_frames, vlen, sample='rand', fix_start=None, input_fps=1, max_num_frames=-1):
+    if sample in ["rand", "middle"]:  # uniform sampling
+        acc_samples = min(num_frames, vlen)
+        # split the video into `acc_samples` intervals, and sample from each interval.
+        intervals = np.linspace(
+            start=0, stop=vlen, num=acc_samples + 1).astype(int)
+        ranges = []
+        for idx, interv in enumerate(intervals[:-1]):
+            ranges.append((interv, intervals[idx + 1] - 1))
+        if sample == 'rand':
+            try:
+                frame_indices = [random.choice(
+                    range(x[0], x[1])) for x in ranges]
+            except Exception:
+                frame_indices = np.random.permutation(vlen)[:acc_samples]
+                frame_indices.sort()
+                frame_indices = list(frame_indices)
+        elif fix_start is not None:
+            frame_indices = [x[0] + fix_start for x in ranges]
+        elif sample == 'middle':
+            frame_indices = [(x[0] + x[1]) // 2 for x in ranges]
+        else:
+            raise NotImplementedError
+
+        if len(frame_indices) < num_frames:  # padded with last frame
+            padded_frame_indices = [frame_indices[-1]] * num_frames
+            padded_frame_indices[:len(frame_indices)] = frame_indices
+            frame_indices = padded_frame_indices
+    elif "fps" in sample:  # fps0.5, sequentially sample frames at 0.5 fps
+        output_fps = float(sample[3:])
+        duration = float(vlen) / input_fps
+        # gap between frames, this is also the clip length each frame represents
+        delta = 1 / output_fps
+        frame_seconds = np.arange(0 + delta / 2, duration + delta / 2, delta)
+        frame_indices = np.around(frame_seconds * input_fps).astype(int)
+        frame_indices = [e for e in frame_indices if e < vlen]
+        if max_num_frames > 0 and len(frame_indices) > max_num_frames:
+            frame_indices = frame_indices[:max_num_frames]
+    else:
+        raise ValueError
+    return frame_indices
+
+
+def get_index(num_frames, bound, fps, max_frame, first_idx=0):
+    if bound:
+        start, end = bound[0], bound[1]
+    else:
+        start, end = -100000, 100000
+    start_idx = max(first_idx, round(start * fps))
+    end_idx = min(round(end * fps), max_frame)
+    seg_size = float(end_idx - start_idx) / num_frames
+    frame_indices = np.array([
+        int(start_idx + (seg_size / 2) + np.round(seg_size * idx))
+        for idx in range(num_frames)
+    ])
+    return frame_indices
+
+
+def read_frames_gif(
+    video_path, num_frames, sample='rand', fix_start=None,
+    max_num_frames=-1, client=None, clip=None,
+):
+    if video_path.startswith('s3') or video_path.startswith('p2'):
+        video_bytes = client.get(video_path)
+        gif = imageio.get_reader(io.BytesIO(video_bytes))
+    else:
+        gif = imageio.get_reader(video_path)
+    vlen = len(gif)
+    frame_indices = get_frame_indices(
+        num_frames, vlen, sample=sample, fix_start=fix_start,
+        max_num_frames=max_num_frames
+    )
+    frames = []
+    reference_size = None
+    for index, frame in enumerate(gif):
+        # for index in frame_idxs:
+        if index in frame_indices:
+            if frame.ndim == 2:
+                frame = cv2.cvtColor(frame, cv2.COLOR_GRAY2RGB)
+            elif frame.shape[2] == 4:
+                frame = cv2.cvtColor(frame, cv2.COLOR_RGBA2RGB)
+            if reference_size is None:
+                reference_size = (frame.shape[1], frame.shape[0])
+            frame = cv2.resize(frame, reference_size,
+                               interpolation=cv2.INTER_LINEAR)
+            frames.append(frame)
+    frames = np.stack(frames, axis=0)  # .float() / 255
+
+    return frames
+
+
+def read_frames_decord(
+    video_path, num_frames, sample='rand', fix_start=None,
+    max_num_frames=-1, client=None, clip=None
+):
+    if video_path.startswith('s3') or video_path.startswith('p2') or video_path.startswith('p_hdd') or video_path.startswith('cluster1'):
+        video_bytes = client.get(video_path)
+        video_reader = VideoReader(io.BytesIO(video_bytes), num_threads=1)
+    else:
+        video_reader = VideoReader(video_path, num_threads=1)
+    vlen = len(video_reader)
+    fps = video_reader.get_avg_fps()
+    duration = vlen / float(fps)
+
+    if clip:
+        vlen = int(duration * fps)
+        frame_indices = get_index(num_frames, clip, fps, vlen)
+    else:
+        frame_indices = get_frame_indices(
+            num_frames, vlen, sample=sample, fix_start=fix_start,
+            input_fps=fps, max_num_frames=max_num_frames
+        )
+    # if clip:
+    #     frame_indices = [f + start_index for f in frame_indices]
+
+    frames = video_reader.get_batch(frame_indices).asnumpy()  # (T, H, W, C)
+    return frames
+
+
+def read_diff_frames_decord(
+    video_path, clip, client=None
+):
+    if video_path.startswith('s3') or video_path.startswith('p2') or video_path.startswith('p_hdd') or video_path.startswith('cluster1') or video_path.startswith('s_hdd'):
+        video_bytes = client.get(video_path)
+        video_reader = VideoReader(io.BytesIO(video_bytes), num_threads=1)
+    else:
+        video_reader = VideoReader(video_path, num_threads=1)
+    vlen = len(video_reader)
+    fps = video_reader.get_avg_fps()
+
+    start_idx = round(clip[0]*fps)
+    end_idx = min(round(clip[1]*fps), vlen)
+    frame_indices = [start_idx, end_idx]
+
+    frames = video_reader.get_batch(frame_indices).asnumpy()  # (T, H, W, C)
+    return frames

pyproject.toml

@@ -0,0 +1,37 @@
+[build-system]
+requires = ["setuptools>=61.0"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "llava"
+version = "1.2.2.post1"
+description = "Towards GPT-4 like large language and visual assistant."
+readme = "README.md"
+requires-python = ">=3.8"
+classifiers = [
+    "Programming Language :: Python :: 3",
+    "License :: OSI Approved :: Apache Software License",
+]
+dependencies = [
+    "torch==2.1.2", "torchvision==0.16.2",
+    "transformers==4.37.2", "tokenizers==0.15.1", "sentencepiece==0.1.99", "shortuuid",
+    "accelerate==0.21.0", "peft", "bitsandbytes",
+    "pydantic", "markdown2[all]", "numpy", "scikit-learn==1.2.2",
+    "gradio==4.16.0", "gradio_client==0.8.1", "openai", "spaces",
+    "requests", "httpx==0.24.0", "uvicorn", "fastapi", "decord",
+    "einops==0.6.1", "einops-exts==0.0.4", "timm==0.6.13",
+]
+
+[project.optional-dependencies]
+train = ["deepspeed==0.12.6", "ninja", "wandb"]
+build = ["build", "twine"]
+
+[project.urls]
+"Homepage" = "https://llava-vl.github.io"
+"Bug Tracker" = "https://github.com/haotian-liu/LLaVA/issues"
+
+[tool.setuptools.packages.find]
+exclude = ["assets*", "benchmark*", "docs", "dist*", "playground*", "scripts*", "tests*"]
+
+[tool.wheel]
+exclude = ["assets*", "benchmark*", "docs", "dist*", "playground*", "scripts*", "tests*"]