Initial commit

deepseek_vl2/serve/app_modules/gradio_utils.py

+# Copyright (c) 2023-2024 DeepSeek.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy of
+# this software and associated documentation files (the "Software"), to deal in
+# the Software without restriction, including without limitation the rights to
+# use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+# the Software, and to permit persons to whom the Software is furnished to do so,
+# subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
+# FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+# COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+# IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+# CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+from functools import wraps
+
+import gradio as gr
+
+
+def wrap_gen_fn(gen_fn):
+    @wraps(gen_fn)
+    def wrapped_gen_fn(prompt, *args, **kwargs):
+        try:
+            yield from gen_fn(prompt, *args, **kwargs)
+        except gr.Error as g_err:
+            raise g_err
+        except Exception as e:
+            raise gr.Error(f"Failed to generate text: {e}") from e
+
+    return wrapped_gen_fn
+
+
+def delete_last_conversation(chatbot, history):
+    if len(history) % 2 != 0:
+        gr.Error("history length is not even")
+        return (
+            chatbot,
+            history,
+            "Delete Done",
+        )
+
+    if len(chatbot) > 0:
+        chatbot.pop()
+
+    if len(history) > 0 and len(history) % 2 == 0:
+        history.pop()
+        history.pop()
+
+    return (
+        chatbot,
+        history,
+        "Delete Done",
+    )
+
+
+def reset_state():
+    return [], [], None, "Reset Done"
+
+
+def reset_textbox():
+    return gr.update(value=""), ""
+
+
+def cancel_outputing():
+    return "Stop Done"
+
+
+class State:
+    interrupted = False
+
+    def interrupt(self):
+        self.interrupted = True
+
+    def recover(self):
+        self.interrupted = False
+
+
+shared_state = State()

deepseek_vl2/serve/app_modules/overwrites.py

+# Copyright (c) 2023-2024 DeepSeek.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy of
+# this software and associated documentation files (the "Software"), to deal in
+# the Software without restriction, including without limitation the rights to
+# use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+# the Software, and to permit persons to whom the Software is furnished to do so,
+# subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
+# FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+# COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+# IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+# CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+from __future__ import annotations
+
+import logging
+from typing import List, Tuple
+
+from deepseek_vl2.serve.app_modules.presets import gr
+from deepseek_vl2.serve.app_modules.utils import convert_asis, convert_mdtext, detect_converted_mark
+
+
+def compact_text_chunks(self, prompt, text_chunks: List[str]) -> List[str]:
+    logging.debug("Compacting text chunks...🚀🚀🚀")
+    combined_str = [c.strip() for c in text_chunks if c.strip()]
+    combined_str = [f"[{index+1}] {c}" for index, c in enumerate(combined_str)]
+    combined_str = "\n\n".join(combined_str)
+    # resplit based on self.max_chunk_overlap
+    text_splitter = self.get_text_splitter_given_prompt(prompt, 1, padding=1)
+    return text_splitter.split_text(combined_str)
+
+
+def postprocess(
+    self, y: List[Tuple[str | None, str | None]]
+) -> List[Tuple[str | None, str | None]]:
+    """
+    Parameters:
+        y: List of tuples representing the message and response pairs. Each message and response should be a string, which may be in Markdown format.
+    Returns:
+        List of tuples representing the message and response. Each message and response will be a string of HTML.
+    """
+    if y is None or y == []:
+        return []
+    temp = []
+    for x in y:
+        user, bot = x
+        if not detect_converted_mark(user):
+            user = convert_asis(user)
+        if not detect_converted_mark(bot):
+            bot = convert_mdtext(bot)
+        temp.append((user, bot))
+    return temp
+
+
+with open("deepseek_vl2/serve/assets/custom.js", "r", encoding="utf-8") as f, open(
+    "deepseek_vl2/serve/assets/Kelpy-Codos.js", "r", encoding="utf-8"
+) as f2:
+    customJS = f.read()
+    kelpyCodos = f2.read()
+
+
+def reload_javascript():
+    print("Reloading javascript...")
+    js = f"<script>{customJS}</script><script>{kelpyCodos}</script>"
+
+    def template_response(*args, **kwargs):
+        res = GradioTemplateResponseOriginal(*args, **kwargs)
+        res.body = res.body.replace(b"</html>", f"{js}</html>".encode("utf8"))
+        res.init_headers()
+        return res
+
+    gr.routes.templates.TemplateResponse = template_response
+
+
+GradioTemplateResponseOriginal = gr.routes.templates.TemplateResponse

deepseek_vl2/serve/app_modules/presets.py

+# Copyright (c) 2023-2024 DeepSeek.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy of
+# this software and associated documentation files (the "Software"), to deal in
+# the Software without restriction, including without limitation the rights to
+# use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+# the Software, and to permit persons to whom the Software is furnished to do so,
+# subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
+# FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+# COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+# IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+# CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+# -*- coding:utf-8 -*-
+import gradio as gr
+
+title = """<h1 align="left" style="min-width:200px; margin-top:0;">Chat with DeepSeek-VL2 </h1>"""
+description_top = """Special Tokens: `<image>`,     Visual Grounding: `<|ref|>{query}<|/ref|>`,    Grounding Conversation: `<|grounding|>{question}`"""
+description = """"""
+CONCURRENT_COUNT = 1
+MAX_EVENTS = 10
+MAX_IMAGE_SIZE = 800
+MIN_IMAGE_SIZE = 400
+
+BOX2COLOR = {
+    0: (255, 0, 0),
+    1: (0, 255, 0),
+    2: (0, 0, 255),
+    3: (0, 255, 255),
+    4: (255, 255, 0),
+    5: (255, 0, 255),
+    6: (127, 127, 127),
+    7: (255, 255, 127),
+    8: (255, 127, 255),
+    9: (127, 255, 255),
+    10: (127, 127, 255),
+    11: (127, 255, 127),
+    12: (255, 127, 127),
+}
+
+
+ALREADY_CONVERTED_MARK = "<!-- ALREADY CONVERTED BY PARSER. -->"
+
+small_and_beautiful_theme = gr.themes.Soft(
+    primary_hue=gr.themes.Color(
+        c50="#EBFAF2",
+        c100="#CFF3E1",
+        c200="#A8EAC8",
+        c300="#77DEA9",
+        c400="#3FD086",
+        c500="#02C160",
+        c600="#06AE56",
+        c700="#05974E",
+        c800="#057F45",
+        c900="#04673D",
+        c950="#2E5541",
+        name="small_and_beautiful",
+    ),
+    secondary_hue=gr.themes.Color(
+        c50="#576b95",
+        c100="#576b95",
+        c200="#576b95",
+        c300="#576b95",
+        c400="#576b95",
+        c500="#576b95",
+        c600="#576b95",
+        c700="#576b95",
+        c800="#576b95",
+        c900="#576b95",
+        c950="#576b95",
+    ),
+    neutral_hue=gr.themes.Color(
+        name="gray",
+        c50="#f6f7f8",
+        # c100="#f3f4f6",
+        c100="#F2F2F2",
+        c200="#e5e7eb",
+        c300="#d1d5db",
+        c400="#B2B2B2",
+        c500="#808080",
+        c600="#636363",
+        c700="#515151",
+        c800="#393939",
+        # c900="#272727",
+        c900="#2B2B2B",
+        c950="#171717",
+    ),
+    radius_size=gr.themes.sizes.radius_sm,
+).set(
+    # button_primary_background_fill="*primary_500",
+    button_primary_background_fill_dark="*primary_600",
+    # button_primary_background_fill_hover="*primary_400",
+    # button_primary_border_color="*primary_500",
+    button_primary_border_color_dark="*primary_600",
+    button_primary_text_color="white",
+    button_primary_text_color_dark="white",
+    button_secondary_background_fill="*neutral_100",
+    button_secondary_background_fill_hover="*neutral_50",
+    button_secondary_background_fill_dark="*neutral_900",
+    button_secondary_text_color="*neutral_800",
+    button_secondary_text_color_dark="white",
+    # background_fill_primary="#F7F7F7",
+    # background_fill_primary_dark="#1F1F1F",
+    # block_title_text_color="*primary_500",
+    block_title_background_fill_dark="*primary_900",
+    block_label_background_fill_dark="*primary_900",
+    input_background_fill="#F6F6F6",
+    # chatbot_code_background_color_dark="*neutral_950",
+)

deepseek_vl2/serve/app_modules/utils.py

+# Copyright (c) 2023-2024 DeepSeek.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy of
+# this software and associated documentation files (the "Software"), to deal in
+# the Software without restriction, including without limitation the rights to
+# use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+# the Software, and to permit persons to whom the Software is furnished to do so,
+# subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
+# FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+# COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+# IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+# CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+# -*- coding:utf-8 -*-
+from __future__ import annotations
+
+import html
+import logging
+import io
+import os
+import re
+import base64
+import time
+from PIL import Image, ImageDraw, ImageFont
+
+import mdtex2html
+from markdown import markdown
+from pygments import highlight
+from pygments.formatters import HtmlFormatter
+from pygments.lexers import ClassNotFound, get_lexer_by_name, guess_lexer
+
+from deepseek_vl2.serve.app_modules.presets import (
+    ALREADY_CONVERTED_MARK,
+    BOX2COLOR,
+    MAX_IMAGE_SIZE,
+    MIN_IMAGE_SIZE
+)
+
+logger = logging.getLogger("gradio_logger")
+
+
+def configure_logger():
+    logger = logging.getLogger("gradio_logger")
+    logger.setLevel(logging.DEBUG)
+
+    timestr = time.strftime("%Y%m%d-%H%M%S")
+    os.makedirs("deepseek_vl2/serve/logs", exist_ok=True)
+    file_handler = logging.FileHandler(
+        f"deepseek_vl2/serve/logs/{timestr}_gradio_log.log"
+    )
+    console_handler = logging.StreamHandler()
+
+    formatter = logging.Formatter(
+        "%(asctime)s - %(name)s - %(levelname)s - %(message)s"
+    )
+    console_handler.setFormatter(formatter)
+    file_handler.setFormatter(formatter)
+
+    console_handler.setLevel(logging.INFO)
+    file_handler.setLevel(logging.INFO)
+
+    logger.addHandler(console_handler)
+    logger.addHandler(file_handler)
+
+    return logger
+
+
+def strip_stop_words(x, stop_words):
+    for w in stop_words:
+        if w in x:
+            return x[: x.index(w)].strip()
+    return x.strip()
+
+
+def format_output(history, text, x):
+    updated_history = history + [[text, x]]
+    a = [[y[0], convert_to_markdown(y[1])] for y in updated_history]
+    return a, updated_history
+
+
+def markdown_to_html_with_syntax_highlight(md_str):  # deprecated
+    def replacer(match):
+        lang = match.group(1) or "text"
+        code = match.group(2)
+
+        try:
+            lexer = get_lexer_by_name(lang, stripall=True)
+        except ValueError:
+            lexer = get_lexer_by_name("text", stripall=True)
+
+        formatter = HtmlFormatter()
+        highlighted_code = highlight(code, lexer, formatter)
+
+        return f'<pre><code class="{lang}">{highlighted_code}</code></pre>'
+
+    code_block_pattern = r"```(\w+)?\n([\s\S]+?)\n```"
+    md_str = re.sub(code_block_pattern, replacer, md_str, flags=re.MULTILINE)
+
+    html_str = markdown(md_str)
+    return html_str
+
+
+def normalize_markdown(md_text: str) -> str:  # deprecated
+    lines = md_text.split("\n")
+    normalized_lines = []
+    inside_list = False
+
+    for i, line in enumerate(lines):
+        if re.match(r"^(\d+\.|-|\*|\+)\s", line.strip()):
+            if not inside_list and i > 0 and lines[i - 1].strip() != "":
+                normalized_lines.append("")
+            inside_list = True
+            normalized_lines.append(line)
+        elif inside_list and line.strip() == "":
+            if i < len(lines) - 1 and not re.match(
+                r"^(\d+\.|-|\*|\+)\s", lines[i + 1].strip()
+            ):
+                normalized_lines.append(line)
+            continue
+        else:
+            inside_list = False
+            normalized_lines.append(line)
+
+    return "\n".join(normalized_lines)
+
+
+def convert_mdtext(md_text):
+    code_block_pattern = re.compile(r"```(.*?)(?:```|$)", re.DOTALL)
+    inline_code_pattern = re.compile(r"`(.*?)`", re.DOTALL)
+    code_blocks = code_block_pattern.findall(md_text)
+    non_code_parts = code_block_pattern.split(md_text)[::2]
+
+    result = []
+    for non_code, code in zip(non_code_parts, code_blocks + [""]):
+        if non_code.strip():
+            non_code = normalize_markdown(non_code)
+            if inline_code_pattern.search(non_code):
+                result.append(markdown(non_code, extensions=["tables"]))
+            else:
+                result.append(mdtex2html.convert(non_code, extensions=["tables"]))
+        if code.strip():
+            code = f"\n```{code}\n\n```"
+            code = markdown_to_html_with_syntax_highlight(code)
+            result.append(code)
+    result = "".join(result)
+    result += ALREADY_CONVERTED_MARK
+    return result
+
+
+def convert_asis(userinput):
+    return f'<p style="white-space:pre-wrap;">{html.escape(userinput)}</p>{ALREADY_CONVERTED_MARK}'
+
+
+def is_stop_word_or_prefix(s: str, stop_words: list) -> bool:
+    return any(s.endswith(stop_word) for stop_word in stop_words)
+
+
+def detect_converted_mark(userinput):
+    return bool(userinput.endswith(ALREADY_CONVERTED_MARK))
+
+
+def detect_language(code):
+    first_line = "" if code.startswith("\n") else code.strip().split("\n", 1)[0]
+    language = first_line.lower() if first_line else ""
+    code_without_language = code[len(first_line) :].lstrip() if first_line else code
+    return language, code_without_language
+
+
+def convert_to_markdown(text):
+    text = text.replace("$", "&#36;")
+    text = text.replace("\r\n", "\n")
+
+    def replace_leading_tabs_and_spaces(line):
+        new_line = []
+
+        for char in line:
+            if char == "\t":
+                new_line.append("&#9;")
+            elif char == " ":
+                new_line.append("&nbsp;")
+            else:
+                break
+        return "".join(new_line) + line[len(new_line) :]
+
+    markdown_text = ""
+    lines = text.split("\n")
+    in_code_block = False
+
+    for line in lines:
+        if in_code_block is False and line.startswith("```"):
+            in_code_block = True
+            markdown_text += f"{line}\n"
+        elif in_code_block is True and line.startswith("```"):
+            in_code_block = False
+            markdown_text += f"{line}\n"
+        elif in_code_block:
+            markdown_text += f"{line}\n"
+        else:
+            line = replace_leading_tabs_and_spaces(line)
+            line = re.sub(r"^(#)", r"\\\1", line)
+            markdown_text += f"{line}  \n"
+
+    return markdown_text
+
+
+def add_language_tag(text):
+    def detect_language(code_block):
+        try:
+            lexer = guess_lexer(code_block)
+            return lexer.name.lower()
+        except ClassNotFound:
+            return ""
+
+    code_block_pattern = re.compile(r"(```)(\w*\n[^`]+```)", re.MULTILINE)
+
+    def replacement(match):
+        code_block = match.group(2)
+        if match.group(2).startswith("\n"):
+            language = detect_language(code_block)
+            return (
+                f"```{language}{code_block}```" if language else f"```\n{code_block}```"
+            )
+        else:
+            return match.group(1) + code_block + "```"
+
+    text2 = code_block_pattern.sub(replacement, text)
+    return text2
+
+
+def is_variable_assigned(var_name: str) -> bool:
+    return var_name in locals()
+
+
+def pil_to_base64(
+    image: Image.Image,
+    alt: str = "user upload image",
+    resize: bool = True,
+    max_size: int = MAX_IMAGE_SIZE,
+    min_size: int = MIN_IMAGE_SIZE,
+    format: str = "JPEG",
+    quality: int = 95
+) -> str:
+
+    if resize:
+        max_hw, min_hw = max(image.size), min(image.size)
+        aspect_ratio = max_hw / min_hw
+        shortest_edge = int(min(max_size / aspect_ratio, min_size, 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))
+
+    buffered = io.BytesIO()
+    image.save(buffered, format=format, quality=quality)
+    img_b64_str = base64.b64encode(buffered.getvalue()).decode()
+    img_str = f'<img src="data:image/png;base64,{img_b64_str}" alt="{alt}" />'
+
+    return img_str
+
+
+def parse_ref_bbox(response, image: Image.Image):
+    try:
+        image = image.copy()
+        image_h, image_w = image.size
+        draw = ImageDraw.Draw(image)
+
+        ref = re.findall(r'<\|ref\|>.*?<\|/ref\|>', response)
+        bbox = re.findall(r'<\|det\|>.*?<\|/det\|>', response)
+        assert len(ref) == len(bbox)
+
+        if len(ref) == 0:
+            return None
+
+        boxes, labels = [], []
+        for box, label in zip(bbox, ref):
+            box = box.replace('<|det|>', '').replace('<|/det|>', '')
+            label = label.replace('<|ref|>', '').replace('<|/ref|>', '')
+            box = box[1:-1]
+            for onebox in re.findall(r'\[.*?\]', box):
+                boxes.append(eval(onebox))
+                labels.append(label)
+
+        for indice, (box, label) in enumerate(zip(boxes, labels)):
+            box = (
+                int(box[0] / 999 * image_h),
+                int(box[1] / 999 * image_w),
+                int(box[2] / 999 * image_h),
+                int(box[3] / 999 * image_w),
+            )
+
+            box_color = BOX2COLOR[indice % len(BOX2COLOR.keys())]
+            box_width = 3
+            draw.rectangle(box, outline=box_color, width=box_width)
+
+            text_x = box[0]
+            text_y = box[1] - 20
+            text_color = box_color
+            font = ImageFont.truetype("deepseek_vl2/serve/assets/simsun.ttc", size=20)
+            draw.text((text_x, text_y), label, font=font, fill=text_color)
+
+        # print(f"boxes = {boxes}, labels = {labels}, re-render = {image}")
+        return image
+    except:
+        return None
+
+
+def display_example(image_list):
+    images_html = ""
+    for i, img_path in enumerate(image_list):
+        image = Image.open(img_path)
+        buffered = io.BytesIO()
+        image.save(buffered, format="PNG", quality=100)
+        img_b64_str = base64.b64encode(buffered.getvalue()).decode()
+        img_str = f'<img src="data:image/png;base64,{img_b64_str}" alt="{img_path}" style="height:80px; margin-right: 10px;" />'
+        images_html += img_str
+
+    result_html = f"""
+    <div style="display: flex; align-items: center; margin-bottom: 10px;">
+        <div style="flex: 1; margin-right: 10px;">{images_html}</div>
+    </div>
+    """
+
+    return result_html
+

deepseek_vl2/serve/assets/Kelpy-Codos.js

+/**
+ * Copyright (c) 2023-2024 DeepSeek.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy of
+ * this software and associated documentation files (the "Software"), to deal in
+ * the Software without restriction, including without limitation the rights to
+ * use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+ * the Software, and to permit persons to whom the Software is furnished to do so,
+ * subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
+ * FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+ * COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+ * IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+ */
+
+// ==UserScript==
+// @name         Kelpy Codos
+// @namespace    https://github.com/Keldos-Li/Kelpy-Codos
+// @version      1.0.5
+// @author       Keldos; https://keldos.me/
+// @description  Add copy button to PRE tags before CODE tag, for Chuanhu ChatGPT especially.
+//               Based on Chuanhu ChatGPT version: ac04408 (2023-3-22)
+// @license      GPL-3.0
+// @grant        none
+// ==/UserScript==
+
+(function () {
+  "use strict";
+
+  function addCopyButton(pre) {
+    var code = pre.querySelector("code");
+    if (!code) {
+      return; // 如果没有找到 <code> 元素，则不添加按钮
+    }
+    var firstChild = code.firstChild;
+    if (!firstChild) {
+      return; // 如果 <code> 元素没有子节点，则不添加按钮
+    }
+    var button = document.createElement("button");
+    button.textContent = "\uD83D\uDCCE"; // 使用 📎 符号作为“复制”按钮的文本
+    button.style.position = "relative";
+    button.style.float = "right";
+    button.style.fontSize = "1em"; // 可选：调整按钮大小
+    button.style.background = "none"; // 可选：去掉背景颜色
+    button.style.border = "none"; // 可选：去掉边框
+    button.style.cursor = "pointer"; // 可选：显示指针样式
+    button.addEventListener("click", function () {
+      var range = document.createRange();
+      range.selectNodeContents(code);
+      range.setStartBefore(firstChild); // 将范围设置为第一个子节点之前
+      var selection = window.getSelection();
+      selection.removeAllRanges();
+      selection.addRange(range);
+
+      try {
+        var success = document.execCommand("copy");
+        if (success) {
+          button.textContent = "\u2714";
+          setTimeout(function () {
+            button.textContent = "\uD83D\uDCCE"; // 恢复按钮为“复制”
+          }, 2000);
+        } else {
+          button.textContent = "\u2716";
+        }
+      } catch (e) {
+        console.error(e);
+        button.textContent = "\u2716";
+      }
+
+      selection.removeAllRanges();
+    });
+    code.insertBefore(button, firstChild); // 将按钮插入到第一个子元素之前
+  }
+
+  function handleNewElements(mutationsList, observer) {
+    for (var mutation of mutationsList) {
+      if (mutation.type === "childList") {
+        for (var node of mutation.addedNodes) {
+          if (node.nodeName === "PRE") {
+            addCopyButton(node);
+          }
+        }
+      }
+    }
+  }
+
+  var observer = new MutationObserver(handleNewElements);
+  observer.observe(document.documentElement, {
+    childList: true,
+    subtree: true,
+  });
+
+  document.querySelectorAll("pre").forEach(addCopyButton);
+})();

deepseek_vl2/serve/assets/custom.css

+/**
+ * Copyright (c) 2023-2024 DeepSeek.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy of
+ * this software and associated documentation files (the "Software"), to deal in
+ * the Software without restriction, including without limitation the rights to
+ * use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+ * the Software, and to permit persons to whom the Software is furnished to do so,
+ * subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
+ * FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+ * COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+ * IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+ */
+
+:root {
+  --chatbot-color-light: #f3f3f3;
+  --chatbot-color-dark: #121111;
+}
+
+/* status_display */
+#status_display {
+  display: flex;
+  min-height: 2.5em;
+  align-items: flex-end;
+  justify-content: flex-end;
+}
+#status_display p {
+  font-size: 0.85em;
+  font-family: monospace;
+  color: var(--body-text-color-subdued);
+}
+
+/* usage_display */
+#usage_display {
+  height: 1em;
+}
+#usage_display p {
+  padding: 0 1em;
+  font-size: 0.85em;
+  font-family: monospace;
+  color: var(--body-text-color-subdued);
+}
+/* list */
+ol:not(.options),
+ul:not(.options) {
+  padding-inline-start: 2em !important;
+}
+
+/* Thank @Keldos-Li for fixing it */
+/* Light mode (default) */
+#deepseek_chatbot {
+  background-color: var(--chatbot-color-light) !important;
+  color: #000000 !important;
+}
+[data-testid="bot"] {
+  background-color: #ffffff !important;
+}
+[data-testid="user"] {
+  background-color: #95ec69 !important;
+}
+
+/* Dark mode */
+.dark #deepseek_chatbot {
+  background-color: var(--chatbot-color-dark) !important;
+  color: #ffffff !important;
+}
+.dark [data-testid="bot"] {
+  background-color: #2c2c2c !important;
+}
+.dark [data-testid="user"] {
+  background-color: #26b561 !important;
+}
+
+#deepseek_chatbot {
+  height: 100%;
+  min-height: 800px;
+  flex-grow: 1;
+  overflow: auto;
+}
+
+[class*="message"] {
+  border-radius: var(--radius-xl) !important;
+  border: none;
+  padding: var(--spacing-xl) !important;
+  font-size: var(--text-md) !important;
+  line-height: var(--line-md) !important;
+  min-height: calc(var(--text-md) * var(--line-md) + 2 * var(--spacing-xl));
+  min-width: calc(var(--text-md) * var(--line-md) + 2 * var(--spacing-xl));
+}
+[data-testid="bot"] {
+  max-width: 85%;
+  border-bottom-left-radius: 0 !important;
+}
+[data-testid="user"] {
+  max-width: 85%;
+  width: auto !important;
+  border-bottom-right-radius: 0 !important;
+}
+/* Table */
+table {
+  margin: 1em 0;
+  border-collapse: collapse;
+  empty-cells: show;
+}
+td,
+th {
+  border: 1.2px solid var(--border-color-primary) !important;
+  padding: 0.2em;
+}
+thead {
+  background-color: rgba(175, 184, 193, 0.2);
+}
+thead th {
+  padding: 0.5em 0.2em;
+}
+/* Inline code */
+#deepseek_chatbot code {
+  display: inline;
+  white-space: break-spaces;
+  border-radius: 6px;
+  margin: 0 2px 0 2px;
+  padding: 0.2em 0.4em 0.1em 0.4em;
+  background-color: rgba(175, 184, 193, 0.2);
+}
+/* Code block */
+#deepseek_chatbot pre code {
+  display: block;
+  overflow: auto;
+  white-space: pre;
+  background-color: #1c1d1e !important;
+  border-radius: 10px;
+  padding: 1.4em 1.2em 0em 1.4em;
+  margin: 1.2em 2em 1.2em 0.5em;
+  color: #fdf8f8;
+  box-shadow: 6px 6px 16px hsla(0, 0%, 0%, 0.2);
+}
+/* Hightlight */
+#deepseek_chatbot .highlight {
+  background-color: transparent;
+}
+#deepseek_chatbot .highlight .hll {
+  background-color: #49483e;
+}
+#deepseek_chatbot .highlight .c {
+  color: #75715e;
+} /* Comment */
+#deepseek_chatbot .highlight .err {
+  color: #960050;
+  background-color: #1e0010;
+} /* Error */
+#deepseek_chatbot .highlight .k {
+  color: #66d9ef;
+} /* Keyword */
+#deepseek_chatbot .highlight .l {
+  color: #ae81ff;
+} /* Literal */
+#deepseek_chatbot .highlight .n {
+  color: #f8f8f2;
+} /* Name */
+#deepseek_chatbot .highlight .o {
+  color: #f92672;
+} /* Operator */
+#deepseek_chatbot .highlight .p {
+  color: #f8f8f2;
+} /* Punctuation */
+#deepseek_chatbot .highlight .ch {
+  color: #75715e;
+} /* Comment.Hashbang */
+#deepseek_chatbot .highlight .cm {
+  color: #75715e;
+} /* Comment.Multiline */
+#deepseek_chatbot .highlight .cp {
+  color: #75715e;
+} /* Comment.Preproc */
+#deepseek_chatbot .highlight .cpf {
+  color: #75715e;
+} /* Comment.PreprocFile */
+#deepseek_chatbot .highlight .c1 {
+  color: #75715e;
+} /* Comment.Single */
+#deepseek_chatbot .highlight .cs {
+  color: #75715e;
+} /* Comment.Special */
+#deepseek_chatbot .highlight .gd {
+  color: #f92672;
+} /* Generic.Deleted */
+#deepseek_chatbot .highlight .ge {
+  font-style: italic;
+} /* Generic.Emph */
+#deepseek_chatbot .highlight .gi {
+  color: #a6e22e;
+} /* Generic.Inserted */
+#deepseek_chatbot .highlight .gs {
+  font-weight: bold;
+} /* Generic.Strong */
+#deepseek_chatbot .highlight .gu {
+  color: #75715e;
+} /* Generic.Subheading */
+#deepseek_chatbot .highlight .kc {
+  color: #66d9ef;
+} /* Keyword.Constant */
+#deepseek_chatbot .highlight .kd {
+  color: #66d9ef;
+} /* Keyword.Declaration */
+#deepseek_chatbot .highlight .kn {
+  color: #f92672;
+} /* Keyword.Namespace */
+#deepseek_chatbot .highlight .kp {
+  color: #66d9ef;
+} /* Keyword.Pseudo */
+#deepseek_chatbot .highlight .kr {
+  color: #66d9ef;
+} /* Keyword.Reserved */
+#deepseek_chatbot .highlight .kt {
+  color: #66d9ef;
+} /* Keyword.Type */
+#deepseek_chatbot .highlight .ld {
+  color: #e6db74;
+} /* Literal.Date */
+#deepseek_chatbot .highlight .m {
+  color: #ae81ff;
+} /* Literal.Number */
+#deepseek_chatbot .highlight .s {
+  color: #e6db74;
+} /* Literal.String */
+#deepseek_chatbot .highlight .na {
+  color: #a6e22e;
+} /* Name.Attribute */
+#deepseek_chatbot .highlight .nb {
+  color: #f8f8f2;
+} /* Name.Builtin */
+#deepseek_chatbot .highlight .nc {
+  color: #a6e22e;
+} /* Name.Class */
+#deepseek_chatbot .highlight .no {
+  color: #66d9ef;
+} /* Name.Constant */
+#deepseek_chatbot .highlight .nd {
+  color: #a6e22e;
+} /* Name.Decorator */
+#deepseek_chatbot .highlight .ni {
+  color: #f8f8f2;
+} /* Name.Entity */
+#deepseek_chatbot .highlight .ne {
+  color: #a6e22e;
+} /* Name.Exception */
+#deepseek_chatbot .highlight .nf {
+  color: #a6e22e;
+} /* Name.Function */
+#deepseek_chatbot .highlight .nl {
+  color: #f8f8f2;
+} /* Name.Label */
+#deepseek_chatbot .highlight .nn {
+  color: #f8f8f2;
+} /* Name.Namespace */
+#deepseek_chatbot .highlight .nx {
+  color: #a6e22e;
+} /* Name.Other */
+#deepseek_chatbot .highlight .py {
+  color: #f8f8f2;
+} /* Name.Property */
+#deepseek_chatbot .highlight .nt {
+  color: #f92672;
+} /* Name.Tag */
+#deepseek_chatbot .highlight .nv {
+  color: #f8f8f2;
+} /* Name.Variable */
+#deepseek_chatbot .highlight .ow {
+  color: #f92672;
+} /* Operator.Word */
+#deepseek_chatbot .highlight .w {
+  color: #f8f8f2;
+} /* Text.Whitespace */
+#deepseek_chatbot .highlight .mb {
+  color: #ae81ff;
+} /* Literal.Number.Bin */
+#deepseek_chatbot .highlight .mf {
+  color: #ae81ff;
+} /* Literal.Number.Float */
+#deepseek_chatbot .highlight .mh {
+  color: #ae81ff;
+} /* Literal.Number.Hex */
+#deepseek_chatbot .highlight .mi {
+  color: #ae81ff;
+} /* Literal.Number.Integer */
+#deepseek_chatbot .highlight .mo {
+  color: #ae81ff;
+} /* Literal.Number.Oct */
+#deepseek_chatbot .highlight .sa {
+  color: #e6db74;
+} /* Literal.String.Affix */
+#deepseek_chatbot .highlight .sb {
+  color: #e6db74;
+} /* Literal.String.Backtick */
+#deepseek_chatbot .highlight .sc {
+  color: #e6db74;
+} /* Literal.String.Char */
+#deepseek_chatbot .highlight .dl {
+  color: #e6db74;
+} /* Literal.String.Delimiter */
+#deepseek_chatbot .highlight .sd {
+  color: #e6db74;
+} /* Literal.String.Doc */
+#deepseek_chatbot .highlight .s2 {
+  color: #e6db74;
+} /* Literal.String.Double */
+#deepseek_chatbot .highlight .se {
+  color: #ae81ff;
+} /* Literal.String.Escape */
+#deepseek_chatbot .highlight .sh {
+  color: #e6db74;
+} /* Literal.String.Heredoc */
+#deepseek_chatbot .highlight .si {
+  color: #e6db74;
+} /* Literal.String.Interpol */
+#deepseek_chatbot .highlight .sx {
+  color: #e6db74;
+} /* Literal.String.Other */
+#deepseek_chatbot .highlight .sr {
+  color: #e6db74;
+} /* Literal.String.Regex */
+#deepseek_chatbot .highlight .s1 {
+  color: #e6db74;
+} /* Literal.String.Single */
+#deepseek_chatbot .highlight .ss {
+  color: #e6db74;
+} /* Literal.String.Symbol */
+#deepseek_chatbot .highlight .bp {
+  color: #f8f8f2;
+} /* Name.Builtin.Pseudo */
+#deepseek_chatbot .highlight .fm {
+  color: #a6e22e;
+} /* Name.Function.Magic */
+#deepseek_chatbot .highlight .vc {
+  color: #f8f8f2;
+} /* Name.Variable.Class */
+#deepseek_chatbot .highlight .vg {
+  color: #f8f8f2;
+} /* Name.Variable.Global */
+#deepseek_chatbot .highlight .vi {
+  color: #f8f8f2;
+} /* Name.Variable.Instance */
+#deepseek_chatbot .highlight .vm {
+  color: #f8f8f2;
+} /* Name.Variable.Magic */
+#deepseek_chatbot .highlight .il {
+  color: #ae81ff;
+} /* Literal.Number.Integer.Long */

deepseek_vl2/serve/assets/custom.js

+/**
+ * Copyright (c) 2023-2024 DeepSeek.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy of
+ * this software and associated documentation files (the "Software"), to deal in
+ * the Software without restriction, including without limitation the rights to
+ * use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+ * the Software, and to permit persons to whom the Software is furnished to do so,
+ * subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
+ * FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+ * COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+ * IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+ */
+
+// custom javascript here

deepseek_vl2/serve/inference.py

+# Copyright (c) 2023-2024 DeepSeek.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy of
+# this software and associated documentation files (the "Software"), to deal in
+# the Software without restriction, including without limitation the rights to
+# use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+# the Software, and to permit persons to whom the Software is furnished to do so,
+# subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
+# FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+# COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+# IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+# CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+from threading import Thread
+from typing import List
+
+import torch
+import transformers
+from transformers import (
+    AutoModelForCausalLM,
+    StoppingCriteria,
+    StoppingCriteriaList,
+    TextIteratorStreamer,
+)
+
+from deepseek_vl2.models import DeepseekVLV2Processor, DeepseekVLV2ForCausalLM
+from deepseek_vl2.models.conversation import Conversation
+
+
+def load_model(model_path, dtype=torch.bfloat16):
+    vl_chat_processor = DeepseekVLV2Processor.from_pretrained(model_path)
+    tokenizer = vl_chat_processor.tokenizer
+
+    vl_gpt: DeepseekVLV2ForCausalLM = AutoModelForCausalLM.from_pretrained(
+        model_path, trust_remote_code=True, torch_dtype=dtype
+    )
+    vl_gpt = vl_gpt.cuda().eval()
+    return tokenizer, vl_gpt, vl_chat_processor
+
+
+def convert_conversation_to_prompts(conversation: Conversation):
+    conv_prompts = []
+
+    last_image = None
+
+    messages = conversation.messages
+    for i in range(0, len(messages), 2):
+
+        if isinstance(messages[i][1], tuple):
+            text, images = messages[i][1]
+            last_image = images[-1]
+        else:
+            text, images = messages[i][1], []
+
+        prompt = {
+            "role": messages[i][0],
+            "content": text,
+            "images": images
+        }
+        response = {"role": messages[i + 1][0], "content": messages[i + 1][1]}
+        conv_prompts.extend([prompt, response])
+
+    return conv_prompts, last_image
+
+
+class StoppingCriteriaSub(StoppingCriteria):
+    def __init__(self, stops=[], encounters=1):
+        super().__init__()
+        self.stops = [stop.to("cuda") for stop in stops]
+
+    def __call__(
+        self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs
+    ):
+        for stop in self.stops:
+            if input_ids.shape[-1] < len(stop):
+                continue
+            if torch.all((stop == input_ids[0][-len(stop) :])).item():
+                return True
+
+        return False
+
+
+@torch.inference_mode()
+def deepseek_generate(
+    conversations: list,
+    vl_gpt: torch.nn.Module,
+    vl_chat_processor: DeepseekVLV2Processor,
+    tokenizer: transformers.PreTrainedTokenizer,
+    stop_words: list,
+    max_length: int = 256,
+    temperature: float = 1.0,
+    top_p: float = 1.0,
+    repetition_penalty: float = 1.1,
+    chunk_size: int = -1
+):
+    pil_images = []
+    for message in conversations:
+        if "images" not in message:
+            continue
+        pil_images.extend(message["images"])
+
+    prepare_inputs = vl_chat_processor.__call__(
+        conversations=conversations,
+        images=pil_images,
+        inference_mode=True,
+        force_batchify=True,
+        system_prompt=""
+    ).to(vl_gpt.device)
+
+    return generate(
+        vl_gpt,
+        tokenizer,
+        prepare_inputs,
+        max_gen_len=max_length,
+        temperature=temperature,
+        repetition_penalty=repetition_penalty,
+        top_p=top_p,
+        stop_words=stop_words,
+        chunk_size=chunk_size
+    )
+
+
+@torch.inference_mode()
+def generate(
+    vl_gpt,
+    tokenizer,
+    prepare_inputs,
+    max_gen_len: int = 256,
+    temperature: float = 0,
+    repetition_penalty=1.1,
+    top_p: float = 0.95,
+    stop_words: List[str] = [],
+    chunk_size: int = -1
+):
+    """Stream the text output from the multimodality model with prompt and image inputs."""
+    streamer = TextIteratorStreamer(tokenizer, skip_prompt=True)
+
+    stop_words_ids = [
+        torch.tensor(tokenizer.encode(stop_word)) for stop_word in stop_words
+    ]
+    stopping_criteria = StoppingCriteriaList(
+        [StoppingCriteriaSub(stops=stop_words_ids)]
+    )
+
+    if chunk_size != -1:
+        inputs_embeds, past_key_values = vl_gpt.incremental_prefilling(
+            input_ids=prepare_inputs.input_ids,
+            images=prepare_inputs.images,
+            images_seq_mask=prepare_inputs.images_seq_mask,
+            images_spatial_crop=prepare_inputs.images_spatial_crop,
+            attention_mask=prepare_inputs.attention_mask,
+            chunk_size=chunk_size
+        )
+    else:
+        inputs_embeds = vl_gpt.prepare_inputs_embeds(**prepare_inputs)
+        past_key_values = None
+
+    generation_config = dict(
+        inputs_embeds=inputs_embeds,
+        input_ids=prepare_inputs.input_ids,
+        images=prepare_inputs.images,
+        images_seq_mask=prepare_inputs.images_seq_mask,
+        images_spatial_crop=prepare_inputs.images_spatial_crop,
+        attention_mask=prepare_inputs.attention_mask,
+        past_key_values=past_key_values,
+        pad_token_id=tokenizer.eos_token_id,
+        bos_token_id=tokenizer.bos_token_id,
+        eos_token_id=tokenizer.eos_token_id,
+        max_new_tokens=max_gen_len,
+        do_sample=True,
+        use_cache=True,
+        streamer=streamer,
+        stopping_criteria=stopping_criteria,
+    )
+
+    if temperature > 0:
+        generation_config.update(
+            {
+                "do_sample": True,
+                "top_p": top_p,
+                "temperature": temperature,
+                "repetition_penalty": repetition_penalty,
+            }
+        )
+    else:
+        generation_config["do_sample"] = False
+
+    thread = Thread(target=vl_gpt.generate, kwargs=generation_config)
+    thread.start()
+
+    yield from streamer

deepseek_vl2/utils/__init__.py

+# Copyright (c) 2023-2024 DeepSeek.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy of
+# this software and associated documentation files (the "Software"), to deal in
+# the Software without restriction, including without limitation the rights to
+# use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+# the Software, and to permit persons to whom the Software is furnished to do so,
+# subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
+# FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+# COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+# IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+# CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

deepseek_vl2/utils/io.py

+# Copyright (c) 2023-2024 DeepSeek.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy of
+# this software and associated documentation files (the "Software"), to deal in
+# the Software without restriction, including without limitation the rights to
+# use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+# the Software, and to permit persons to whom the Software is furnished to do so,
+# subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
+# FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+# COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+# IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+# CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+import json
+from typing import Dict, List
+
+import PIL.Image
+import torch
+from transformers import AutoModelForCausalLM
+
+
+def load_pretrained_model(model_path: str):
+
+    from deepseek_vl2.models.processing_deepseek_vl_v2 import DeepseekVLV2Processor
+    from deepseek_vl2.models.modeling_deepseek_vl_v2 import DeepseekVLV2ForCausalLM
+
+    vl_chat_processor = DeepseekVLV2Processor.from_pretrained(model_path)
+    tokenizer = vl_chat_processor.tokenizer
+
+    vl_gpt: DeepseekVLV2ForCausalLM = AutoModelForCausalLM.from_pretrained(
+        model_path, trust_remote_code=True
+    )
+    vl_gpt = vl_gpt.to(torch.bfloat16).cuda().eval()
+
+    return tokenizer, vl_chat_processor, vl_gpt
+
+
+def load_pil_images(conversations: List[Dict[str, str]]) -> List[PIL.Image.Image]:
+    """
+
+    Args:
+        conversations (List[Dict[str, str]]): the conversations with a list of messages. An example is :
+            [
+                {
+                    "role": "User",
+                    "content": "<image>\nExtract all information from this image and convert them into markdown format.",
+                    "images": ["./examples/table_datasets.png"]
+                },
+                {"role": "Assistant", "content": ""},
+            ]
+
+    Returns:
+        pil_images (List[PIL.Image.Image]): the list of PIL images.
+
+    """
+
+    pil_images = []
+
+    for message in conversations:
+        if "images" not in message:
+            continue
+
+        for image_path in message["images"]:
+            pil_img = PIL.Image.open(image_path)
+            pil_img = pil_img.convert("RGB")
+            pil_images.append(pil_img)
+
+    return pil_images
+
+
+def load_json(filepath):
+    with open(filepath, "r") as f:
+        data = json.load(f)
+        return data

images/badge.svg

+<svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" width="151" height="20" role="img" aria-label="DeepSeek: Homepage"><title>DeepSeek: Homepage</title><linearGradient id="s" x2="0" y2="100%"><stop offset="0" stop-color="#bbb" stop-opacity=".1"/><stop offset="1" stop-opacity=".1"/></linearGradient><clipPath id="r"><rect width="151" height="20" rx="3" fill="#fff"/></clipPath><g clip-path="url(#r)"><rect width="82" height="20" fill="#555"/><rect x="82" width="69" height="20" fill="#536af5"/><rect width="151" height="20" fill="url(#s)"/></g><g fill="#fff" text-anchor="middle" font-family="Verdana,Geneva,DejaVu Sans,sans-serif" text-rendering="geometricPrecision" font-size="110"><image x="5" y="3" width="14" height="14" xlink:href="data:image/svg+xml;base64,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"/><text aria-hidden="true" x="505" y="150" fill="#010101" fill-opacity=".3" transform="scale(.1)" textLength="550">DeepSeek</text><text x="505" y="140" transform="scale(.1)" fill="#fff" textLength="550">DeepSeek</text><text aria-hidden="true" x="1155" y="150" fill="#010101" fill-opacity=".3" transform="scale(.1)" textLength="590">Homepage</text><text x="1155" y="140" transform="scale(.1)" fill="#fff" textLength="590">Homepage</text></g></svg>