Spaces:
Sleeping
Sleeping
Upload folder using huggingface_hub
Browse files
app.py
CHANGED
|
@@ -1,4 +1,5 @@
|
|
| 1 |
import gradio as gr
|
|
|
|
| 2 |
import os
|
| 3 |
import torch
|
| 4 |
from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
|
|
@@ -83,18 +84,19 @@ class ModelHandler:
|
|
| 83 |
|
| 84 |
|
| 85 |
def predict(self, messages, temperature, max_tokens):
|
| 86 |
-
#
|
| 87 |
-
# 不要在函数内部重新解析 history,而是直接接收处理好的标准 messages 列表
|
| 88 |
|
| 89 |
-
#
|
| 90 |
-
|
| 91 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 92 |
sys_prompt_fmt = "\n" + " ".join(sys_prompt.split())
|
| 93 |
|
| 94 |
-
|
| 95 |
-
|
| 96 |
-
elif isinstance(content, list):
|
| 97 |
-
# 如果是多模态列表,找到 text 部分追加
|
| 98 |
text_found = False
|
| 99 |
for item in content:
|
| 100 |
if item.get("type") == "text":
|
|
@@ -103,14 +105,14 @@ class ModelHandler:
|
|
| 103 |
break
|
| 104 |
if not text_found:
|
| 105 |
content.append({"type": "text", "text": sys_prompt_fmt})
|
|
|
|
|
|
|
| 106 |
|
| 107 |
-
|
| 108 |
-
|
| 109 |
text_prompt = self.processor.apply_chat_template(
|
| 110 |
-
|
| 111 |
)
|
| 112 |
-
|
| 113 |
-
image_inputs, video_inputs = process_vision_info(messages)
|
| 114 |
|
| 115 |
inputs = self.processor(
|
| 116 |
text=[text_prompt],
|
|
@@ -141,7 +143,7 @@ class ModelHandler:
|
|
| 141 |
generated_ids,
|
| 142 |
skip_special_tokens=True
|
| 143 |
)
|
| 144 |
-
|
| 145 |
if generated_text:
|
| 146 |
yield generated_text
|
| 147 |
else:
|
|
@@ -164,106 +166,59 @@ def get_model_handler():
|
|
| 164 |
model_handler = ModelHandler(MODEL_PATH)
|
| 165 |
return model_handler
|
| 166 |
|
| 167 |
-
def _history_to_messages(history):
|
| 168 |
-
"""
|
| 169 |
-
【新增辅助函数】学习参考代码的逻辑:
|
| 170 |
-
将 Gradio 的 Tuple 历史 [[user_msg, bot_msg], ...]
|
| 171 |
-
转换为 Model 需要的 OpenAI 格式 [{'role': 'user', ...}, ...]
|
| 172 |
-
"""
|
| 173 |
-
messages = []
|
| 174 |
-
|
| 175 |
-
for pair in history:
|
| 176 |
-
user_msg, bot_msg = pair
|
| 177 |
-
|
| 178 |
-
# --- 1. 处理用户消息 ---
|
| 179 |
-
if user_msg:
|
| 180 |
-
# 判断是否为图片路径(简单判断)
|
| 181 |
-
is_image = False
|
| 182 |
-
if isinstance(user_msg, str):
|
| 183 |
-
if os.path.exists(user_msg) or any(user_msg.lower().endswith(ext) for ext in ['.jpg', '.jpeg', '.png', '.webp']):
|
| 184 |
-
is_image = True
|
| 185 |
-
|
| 186 |
-
if is_image:
|
| 187 |
-
# 图片消息
|
| 188 |
-
messages.append({
|
| 189 |
-
"role": "user",
|
| 190 |
-
"content": [{"type": "image", "image": user_msg}]
|
| 191 |
-
})
|
| 192 |
-
else:
|
| 193 |
-
# 文本消息
|
| 194 |
-
messages.append({
|
| 195 |
-
"role": "user",
|
| 196 |
-
"content": [{"type": "text", "text": str(user_msg)}]
|
| 197 |
-
})
|
| 198 |
-
|
| 199 |
-
# --- 2. 处理机器人消息 ---
|
| 200 |
-
if bot_msg:
|
| 201 |
-
messages.append({
|
| 202 |
-
"role": "assistant",
|
| 203 |
-
"content": [{"type": "text", "text": str(bot_msg)}]
|
| 204 |
-
})
|
| 205 |
-
|
| 206 |
-
return messages
|
| 207 |
-
|
| 208 |
@gpu_decorator
|
| 209 |
def respond(user_msg, history, temp, tokens):
|
| 210 |
"""
|
| 211 |
-
|
| 212 |
-
|
| 213 |
"""
|
| 214 |
|
| 215 |
-
# 1.
|
| 216 |
-
|
| 217 |
-
text = user_msg.get("text", "")
|
| 218 |
|
| 219 |
-
#
|
| 220 |
-
|
| 221 |
-
# 先把图片加进历史
|
| 222 |
for f in files:
|
| 223 |
-
|
| 224 |
-
|
| 225 |
-
|
|
|
|
|
|
|
| 226 |
if text:
|
| 227 |
-
|
| 228 |
-
elif not text and files:
|
| 229 |
-
# 如果只有图没字,也需要触发回复,不做操作,history 已经有了图片项
|
| 230 |
-
pass
|
| 231 |
|
| 232 |
-
#
|
| 233 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 234 |
|
| 235 |
-
#
|
| 236 |
-
|
|
|
|
|
|
|
| 237 |
try:
|
| 238 |
handler = get_model_handler()
|
| 239 |
|
| 240 |
-
|
| 241 |
-
messages = _history_to_messages(history)
|
| 242 |
|
| 243 |
-
# 在 UI 历史上预留一个空位给机器人回复
|
| 244 |
-
# 如果最后一条只有 User 内容 (例如 [text, None]),我们填充它
|
| 245 |
-
if history and history[-1][1] is None:
|
| 246 |
-
history[-1][1] = ""
|
| 247 |
-
else:
|
| 248 |
-
history.append([None, ""]) # 万一没对应上,追加一个空回复
|
| 249 |
-
|
| 250 |
-
# 4. 调用模型流式生成
|
| 251 |
full_response = ""
|
| 252 |
-
|
|
|
|
| 253 |
full_response += chunk
|
| 254 |
-
|
| 255 |
-
history[-1][1] = full_response
|
| 256 |
yield history, gr.MultimodalTextbox(interactive=False)
|
| 257 |
|
| 258 |
except Exception as e:
|
| 259 |
import traceback
|
| 260 |
traceback.print_exc()
|
| 261 |
-
|
| 262 |
-
err_msg = f"❌ Error: {str(e)}"
|
| 263 |
-
if history and history[-1][1] is None:
|
| 264 |
-
history[-1][1] = err_msg
|
| 265 |
-
else:
|
| 266 |
-
history.append([None, err_msg])
|
| 267 |
yield history, gr.MultimodalTextbox(interactive=True)
|
| 268 |
|
| 269 |
# 恢复输入框
|
|
@@ -275,18 +230,20 @@ def create_chat_ui():
|
|
| 275 |
#chatbot { height: 650px !important; overflow-y: auto; }
|
| 276 |
"""
|
| 277 |
|
| 278 |
-
with gr.Blocks(title="Robust-R1") as demo:
|
| 279 |
|
| 280 |
with gr.Row():
|
| 281 |
-
gr.Markdown("# 🤖Robust-R1:Degradation-Aware Reasoning
|
| 282 |
|
| 283 |
with gr.Row():
|
| 284 |
with gr.Column(scale=4):
|
|
|
|
| 285 |
chatbot = gr.Chatbot(
|
| 286 |
elem_id="chatbot",
|
| 287 |
label="Chat",
|
| 288 |
avatar_images=(None, "https://api.dicebear.com/7.x/bottts/svg?seed=Qwen"),
|
| 289 |
-
height=650
|
|
|
|
| 290 |
)
|
| 291 |
|
| 292 |
chat_input = gr.MultimodalTextbox(
|
|
@@ -342,8 +299,8 @@ def create_chat_ui():
|
|
| 342 |
outputs=[chatbot, chat_input]
|
| 343 |
)
|
| 344 |
|
| 345 |
-
|
| 346 |
-
clear_btn.click(
|
| 347 |
|
| 348 |
return demo
|
| 349 |
|
|
|
|
| 1 |
import gradio as gr
|
| 2 |
+
print(f"当前使用的 Gradio 版本是: {gr.__version__}")
|
| 3 |
import os
|
| 4 |
import torch
|
| 5 |
from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
|
|
|
|
| 84 |
|
| 85 |
|
| 86 |
def predict(self, messages, temperature, max_tokens):
|
| 87 |
+
# 注意:这里接收到的 messages 已经是标准的 [{'role': 'user', 'content': [...]}, ...]
|
|
|
|
| 88 |
|
| 89 |
+
# 我们需要做一个深拷贝,避免修改 UI 上的 history 显示 System Prompt
|
| 90 |
+
import copy
|
| 91 |
+
messages_payload = copy.deepcopy(messages)
|
| 92 |
+
|
| 93 |
+
# 拼接 System Prompt
|
| 94 |
+
if messages_payload and messages_payload[-1]["role"] == "user":
|
| 95 |
+
content = messages_payload[-1]["content"]
|
| 96 |
sys_prompt_fmt = "\n" + " ".join(sys_prompt.split())
|
| 97 |
|
| 98 |
+
# 现在的 content 肯定是 list (因为我们上面的 respond 函数构建的是 list)
|
| 99 |
+
if isinstance(content, list):
|
|
|
|
|
|
|
| 100 |
text_found = False
|
| 101 |
for item in content:
|
| 102 |
if item.get("type") == "text":
|
|
|
|
| 105 |
break
|
| 106 |
if not text_found:
|
| 107 |
content.append({"type": "text", "text": sys_prompt_fmt})
|
| 108 |
+
elif isinstance(content, str):
|
| 109 |
+
messages_payload[-1]["content"] += sys_prompt_fmt
|
| 110 |
|
| 111 |
+
# 后续逻辑保持不变 ...
|
|
|
|
| 112 |
text_prompt = self.processor.apply_chat_template(
|
| 113 |
+
messages_payload, tokenize=False, add_generation_prompt=True
|
| 114 |
)
|
| 115 |
+
image_inputs, video_inputs = process_vision_info(messages_payload)
|
|
|
|
| 116 |
|
| 117 |
inputs = self.processor(
|
| 118 |
text=[text_prompt],
|
|
|
|
| 143 |
generated_ids,
|
| 144 |
skip_special_tokens=True
|
| 145 |
)
|
| 146 |
+
print(f"Generated text: {generated_text}")
|
| 147 |
if generated_text:
|
| 148 |
yield generated_text
|
| 149 |
else:
|
|
|
|
| 166 |
model_handler = ModelHandler(MODEL_PATH)
|
| 167 |
return model_handler
|
| 168 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 169 |
@gpu_decorator
|
| 170 |
def respond(user_msg, history, temp, tokens):
|
| 171 |
"""
|
| 172 |
+
针对 type="messages" 的 Chatbot 重写的响应函数
|
| 173 |
+
history 现在的格式直接是: [{'role': 'user', 'content': ...}, {'role': 'assistant', ...}]
|
| 174 |
"""
|
| 175 |
|
| 176 |
+
# 1. 构建当前用户的消息内容 (OpenAI 多模态格式)
|
| 177 |
+
user_content = []
|
|
|
|
| 178 |
|
| 179 |
+
# 处理图片/文件
|
| 180 |
+
files = user_msg.get("files", [])
|
|
|
|
| 181 |
for f in files:
|
| 182 |
+
# qwen_vl_utils 识别 "image" 字段作为本地路径
|
| 183 |
+
user_content.append({"type": "image", "image": f})
|
| 184 |
+
|
| 185 |
+
# 处理文本
|
| 186 |
+
text = user_msg.get("text", "")
|
| 187 |
if text:
|
| 188 |
+
user_content.append({"type": "text", "text": text})
|
|
|
|
|
|
|
|
|
|
| 189 |
|
| 190 |
+
# 如果既没图也没字,直接返回
|
| 191 |
+
if not user_content:
|
| 192 |
+
yield history, gr.MultimodalTextbox(value=None, interactive=True)
|
| 193 |
+
return
|
| 194 |
+
|
| 195 |
+
# 2. 将用户消息加入历史
|
| 196 |
+
# 注意:这里直接 append 一个 dict,而不是 tuple
|
| 197 |
+
history.append({
|
| 198 |
+
"role": "user",
|
| 199 |
+
"content": user_content
|
| 200 |
+
})
|
| 201 |
|
| 202 |
+
# 立即更新 UI,让用户看到自己的输入(图文会在同一个气泡里)
|
| 203 |
+
yield history, gr.MultimodalTextbox(value=None, interactive=False)
|
| 204 |
+
|
| 205 |
+
# 3. 调用模型
|
| 206 |
try:
|
| 207 |
handler = get_model_handler()
|
| 208 |
|
| 209 |
+
history.append({"role": "assistant", "content": ""})
|
|
|
|
| 210 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 211 |
full_response = ""
|
| 212 |
+
# 调用你的 handler.predict (注意:你需要稍微调整 handler.predict 里的 sys_prompt 处理逻辑,见下文建议)
|
| 213 |
+
for chunk in handler.predict(history[:-1], temp, tokens): # 传入除最后一条空回复外的历史
|
| 214 |
full_response += chunk
|
| 215 |
+
history[-1]["content"] = full_response
|
|
|
|
| 216 |
yield history, gr.MultimodalTextbox(interactive=False)
|
| 217 |
|
| 218 |
except Exception as e:
|
| 219 |
import traceback
|
| 220 |
traceback.print_exc()
|
| 221 |
+
history.append({"role": "assistant", "content": f"❌ Error: {str(e)}"})
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 222 |
yield history, gr.MultimodalTextbox(interactive=True)
|
| 223 |
|
| 224 |
# 恢复输入框
|
|
|
|
| 230 |
#chatbot { height: 650px !important; overflow-y: auto; }
|
| 231 |
"""
|
| 232 |
|
| 233 |
+
with gr.Blocks(title="Robust-R1", css=custom_css) as demo:
|
| 234 |
|
| 235 |
with gr.Row():
|
| 236 |
+
gr.Markdown("# 🤖 Robust-R1: Degradation-Aware Reasoning")
|
| 237 |
|
| 238 |
with gr.Row():
|
| 239 |
with gr.Column(scale=4):
|
| 240 |
+
# 【关键修改】添加 type="messages"
|
| 241 |
chatbot = gr.Chatbot(
|
| 242 |
elem_id="chatbot",
|
| 243 |
label="Chat",
|
| 244 |
avatar_images=(None, "https://api.dicebear.com/7.x/bottts/svg?seed=Qwen"),
|
| 245 |
+
height=650,
|
| 246 |
+
type="messages" # <--- 这里是重点!
|
| 247 |
)
|
| 248 |
|
| 249 |
chat_input = gr.MultimodalTextbox(
|
|
|
|
| 299 |
outputs=[chatbot, chat_input]
|
| 300 |
)
|
| 301 |
|
| 302 |
+
# 清空历史只需要返回空列表 []
|
| 303 |
+
clear_btn.click(lambda: ([], None), outputs=[chatbot, chat_input])
|
| 304 |
|
| 305 |
return demo
|
| 306 |
|