Spaces:
Build error
Build error
voidDescriptor
commited on
Commit
•
183d24b
1
Parent(s):
041e17c
Upload app.py
Browse files
app.py
ADDED
@@ -0,0 +1,309 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
import torch
|
3 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
4 |
+
import time
|
5 |
+
import bitsandbytes as bnb
|
6 |
+
|
7 |
+
print(f"bitsandbytes version: {bnb.__version__}")
|
8 |
+
print(f"CUDA is available: {torch.cuda.is_available()}")
|
9 |
+
print(f"CUDA device count: {torch.cuda.device_count()}")
|
10 |
+
if torch.cuda.is_available():
|
11 |
+
print(f"Current CUDA device: {torch.cuda.current_device()}")
|
12 |
+
print(f"CUDA device name: {torch.cuda.get_device_name(0)}")
|
13 |
+
|
14 |
+
class ConversationManager:
|
15 |
+
def __init__(self):
|
16 |
+
self.models = {}
|
17 |
+
self.conversation = []
|
18 |
+
self.delay = 3
|
19 |
+
self.is_paused = False
|
20 |
+
self.current_model = None
|
21 |
+
self.initial_prompt = ""
|
22 |
+
self.task_complete = False
|
23 |
+
|
24 |
+
def load_model(self, model_name):
|
25 |
+
if not model_name:
|
26 |
+
print("Error: Empty model name provided")
|
27 |
+
return None
|
28 |
+
|
29 |
+
if model_name in self.models:
|
30 |
+
return self.models[model_name]
|
31 |
+
|
32 |
+
try:
|
33 |
+
print(f"Attempting to load model: {model_name}")
|
34 |
+
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
35 |
+
# Try to load the model with 8-bit quantization
|
36 |
+
model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto", load_in_8bit=True)
|
37 |
+
except RuntimeError as e:
|
38 |
+
print(f"8-bit quantization not available, falling back to full precision: {e}")
|
39 |
+
if torch.cuda.is_available():
|
40 |
+
model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto")
|
41 |
+
else:
|
42 |
+
model = AutoModelForCausalLM.from_pretrained(model_name)
|
43 |
+
except Exception as e:
|
44 |
+
print(f"Failed to load model {model_name}: {e}")
|
45 |
+
print(f"Error type: {type(e).__name__}")
|
46 |
+
print(f"Error details: {str(e)}")
|
47 |
+
return None
|
48 |
+
|
49 |
+
self.models[model_name] = (model, tokenizer)
|
50 |
+
print(f"Successfully loaded model: {model_name}")
|
51 |
+
return self.models[model_name]
|
52 |
+
|
53 |
+
def generate_response(self, model_name, prompt):
|
54 |
+
model, tokenizer = self.load_model(model_name)
|
55 |
+
|
56 |
+
formatted_prompt = f"Human: {prompt.strip()}\n\nAssistant:"
|
57 |
+
|
58 |
+
inputs = tokenizer(formatted_prompt, return_tensors="pt", max_length=1024, truncation=True)
|
59 |
+
with torch.no_grad():
|
60 |
+
outputs = model.generate(**inputs, max_length=200, num_return_sequences=1, do_sample=True)
|
61 |
+
return tokenizer.decode(outputs[0], skip_special_tokens=True)
|
62 |
+
|
63 |
+
def add_to_conversation(self, model_name, response):
|
64 |
+
self.conversation.append((model_name, response))
|
65 |
+
if "task complete?" in response.lower():
|
66 |
+
self.task_complete = True
|
67 |
+
|
68 |
+
def get_conversation_history(self):
|
69 |
+
return "\n".join([f"{model}: {msg}" for model, msg in self.conversation])
|
70 |
+
|
71 |
+
def clear_conversation(self):
|
72 |
+
self.conversation = []
|
73 |
+
self.initial_prompt = ""
|
74 |
+
self.models = {}
|
75 |
+
self.current_model = None
|
76 |
+
self.task_complete = False
|
77 |
+
|
78 |
+
def rewind_conversation(self, steps):
|
79 |
+
self.conversation = self.conversation[:-steps]
|
80 |
+
self.task_complete = False
|
81 |
+
|
82 |
+
def rewind_and_insert(self, steps, inserted_response):
|
83 |
+
if steps > 0:
|
84 |
+
self.conversation = self.conversation[:-steps]
|
85 |
+
if inserted_response.strip():
|
86 |
+
last_model = self.conversation[-1][0] if self.conversation else "User"
|
87 |
+
next_model = "Model 1" if last_model == "Model 2" or last_model == "User" else "Model 2"
|
88 |
+
self.conversation.append((next_model, inserted_response))
|
89 |
+
self.current_model = last_model
|
90 |
+
self.task_complete = False
|
91 |
+
|
92 |
+
manager = ConversationManager()
|
93 |
+
|
94 |
+
def get_model(dropdown, custom):
|
95 |
+
return custom if custom and custom.strip() else dropdown
|
96 |
+
|
97 |
+
def chat(model1, model2, user_input, history, inserted_response=""):
|
98 |
+
try:
|
99 |
+
print(f"Starting chat with models: {model1}, {model2}")
|
100 |
+
print(f"User input: {user_input}")
|
101 |
+
|
102 |
+
model1 = get_model(model1, model1_custom.value)
|
103 |
+
model2 = get_model(model2, model2_custom.value)
|
104 |
+
|
105 |
+
print(f"Selected models: {model1}, {model2}")
|
106 |
+
|
107 |
+
if not manager.load_model(model1) or not manager.load_model(model2):
|
108 |
+
return "Error: Failed to load one or both models. Please check the model names and try again.", ""
|
109 |
+
|
110 |
+
if not manager.conversation:
|
111 |
+
manager.initial_prompt = user_input
|
112 |
+
manager.clear_conversation()
|
113 |
+
manager.add_to_conversation("User", user_input)
|
114 |
+
|
115 |
+
models = [model1, model2]
|
116 |
+
current_model_index = 0 if manager.current_model in ["User", "Model 2"] else 1
|
117 |
+
|
118 |
+
while not manager.task_complete:
|
119 |
+
if manager.is_paused:
|
120 |
+
yield history, "Conversation paused."
|
121 |
+
return
|
122 |
+
|
123 |
+
model = models[current_model_index]
|
124 |
+
manager.current_model = model
|
125 |
+
|
126 |
+
if inserted_response and current_model_index == 0:
|
127 |
+
response = inserted_response
|
128 |
+
inserted_response = ""
|
129 |
+
else:
|
130 |
+
conversation_history = manager.get_conversation_history()
|
131 |
+
prompt = f"{conversation_history}\n\nPlease continue the conversation. If you believe the task is complete, end your response with 'Task complete?'"
|
132 |
+
response = manager.generate_response(model, prompt)
|
133 |
+
|
134 |
+
manager.add_to_conversation(model, response)
|
135 |
+
history = manager.get_conversation_history()
|
136 |
+
|
137 |
+
for i in range(manager.delay, 0, -1):
|
138 |
+
yield history, f"{model} is writing... {i}"
|
139 |
+
time.sleep(1)
|
140 |
+
|
141 |
+
yield history, ""
|
142 |
+
|
143 |
+
if manager.task_complete:
|
144 |
+
yield history, "Models believe the task is complete. Are you satisfied with the result? (Yes/No)"
|
145 |
+
return
|
146 |
+
|
147 |
+
current_model_index = (current_model_index + 1) % 2
|
148 |
+
|
149 |
+
return history, "Conversation completed."
|
150 |
+
except Exception as e:
|
151 |
+
print(f"Error in chat function: {str(e)}")
|
152 |
+
print(f"Error type: {type(e).__name__}")
|
153 |
+
print(f"Error details: {str(e)}")
|
154 |
+
return f"An error occurred: {str(e)}", ""
|
155 |
+
|
156 |
+
def user_satisfaction(satisfied, history):
|
157 |
+
if satisfied.lower() == 'yes':
|
158 |
+
return history, "Task completed successfully."
|
159 |
+
else:
|
160 |
+
manager.task_complete = False
|
161 |
+
return history, "Continuing the conversation..."
|
162 |
+
|
163 |
+
def pause_conversation():
|
164 |
+
manager.is_paused = True
|
165 |
+
return "Conversation paused. Press Resume to continue."
|
166 |
+
|
167 |
+
def resume_conversation():
|
168 |
+
manager.is_paused = False
|
169 |
+
return "Conversation resumed."
|
170 |
+
|
171 |
+
def edit_response(edited_text):
|
172 |
+
if manager.conversation:
|
173 |
+
manager.conversation[-1] = (manager.current_model, edited_text)
|
174 |
+
manager.task_complete = False
|
175 |
+
return manager.get_conversation_history()
|
176 |
+
|
177 |
+
def restart_conversation(model1, model2, user_input):
|
178 |
+
manager.clear_conversation()
|
179 |
+
return chat(model1, model2, user_input, "")
|
180 |
+
|
181 |
+
def rewind_and_insert(steps, inserted_response, history):
|
182 |
+
manager.rewind_and_insert(int(steps), inserted_response)
|
183 |
+
return manager.get_conversation_history(), ""
|
184 |
+
|
185 |
+
open_source_models = [
|
186 |
+
"mistralai/Mixtral-8x7B-Instruct-v0.1",
|
187 |
+
"bigcode/starcoder2-15b",
|
188 |
+
"bigcode/starcoder2-3b",
|
189 |
+
"tiiuae/falcon-7b",
|
190 |
+
"EleutherAI/gpt-neox-20b",
|
191 |
+
"google/flan-ul2",
|
192 |
+
"stabilityai/stablelm-zephyr-3b",
|
193 |
+
"HuggingFaceH4/zephyr-7b-beta",
|
194 |
+
"microsoft/phi-2",
|
195 |
+
"google/gemma-7b-it",
|
196 |
+
"OpenAssistant/oasst-sft-4-pythia-12b-epoch-3.5",
|
197 |
+
"mosaicml/mpt-7b-chat",
|
198 |
+
"databricks/dolly-v2-12b",
|
199 |
+
"thebloke/Wizard-Vicuna-13B-Uncensored-HF",
|
200 |
+
"bigscience/bloom-560m"
|
201 |
+
]
|
202 |
+
|
203 |
+
with gr.Blocks() as demo:
|
204 |
+
gr.Markdown("# ConversAI Playground")
|
205 |
+
|
206 |
+
with gr.Row():
|
207 |
+
with gr.Column(scale=1):
|
208 |
+
model1_dropdown = gr.Dropdown(choices=open_source_models, label="Model 1")
|
209 |
+
model1_custom = gr.Textbox(label="Custom Model 1")
|
210 |
+
with gr.Column(scale=1):
|
211 |
+
model2_dropdown = gr.Dropdown(choices=open_source_models, label="Model 2")
|
212 |
+
model2_custom = gr.Textbox(label="Custom Model 2")
|
213 |
+
|
214 |
+
user_input = gr.Textbox(label="Initial prompt", lines=2)
|
215 |
+
chat_history = gr.Textbox(label="Conversation", lines=20)
|
216 |
+
current_response = gr.Textbox(label="Current model response", lines=3)
|
217 |
+
|
218 |
+
with gr.Row():
|
219 |
+
pause_btn = gr.Button("Pause")
|
220 |
+
edit_btn = gr.Button("Edit")
|
221 |
+
rewind_btn = gr.Button("Rewind")
|
222 |
+
resume_btn = gr.Button("Resume")
|
223 |
+
restart_btn = gr.Button("Restart")
|
224 |
+
clear_btn = gr.Button("Clear")
|
225 |
+
|
226 |
+
with gr.Row():
|
227 |
+
rewind_steps = gr.Slider(0, 10, 1, label="Steps to rewind")
|
228 |
+
inserted_response = gr.Textbox(label="Insert response after rewind", lines=2)
|
229 |
+
|
230 |
+
delay_slider = gr.Slider(0, 10, 3, label="Response Delay (seconds)")
|
231 |
+
|
232 |
+
user_satisfaction_input = gr.Textbox(label="Are you satisfied with the result? (Yes/No)", visible=False)
|
233 |
+
|
234 |
+
gr.Markdown("""
|
235 |
+
## Button Descriptions
|
236 |
+
- **Pause**: Temporarily stops the conversation. The current model will finish its response.
|
237 |
+
- **Edit**: Allows you to modify the last response in the conversation.
|
238 |
+
- **Rewind**: Removes the specified number of last responses from the conversation.
|
239 |
+
- **Resume**: Continues the conversation from where it was paused.
|
240 |
+
- **Restart**: Begins a new conversation with the same or different models, keeping the initial prompt.
|
241 |
+
- **Clear**: Resets everything, including loaded models, conversation history, and initial prompt.
|
242 |
+
""")
|
243 |
+
|
244 |
+
def on_chat_update(history, response):
|
245 |
+
if response and "Models believe the task is complete" in response:
|
246 |
+
return gr.update(visible=True), gr.update(visible=False)
|
247 |
+
return gr.update(visible=False), gr.update(visible=True)
|
248 |
+
|
249 |
+
start_btn = gr.Button("Start Conversation")
|
250 |
+
chat_output = start_btn.click(
|
251 |
+
chat,
|
252 |
+
inputs=[
|
253 |
+
model1_dropdown,
|
254 |
+
model2_dropdown,
|
255 |
+
user_input,
|
256 |
+
chat_history
|
257 |
+
],
|
258 |
+
outputs=[chat_history, current_response]
|
259 |
+
)
|
260 |
+
|
261 |
+
chat_output.then(
|
262 |
+
on_chat_update,
|
263 |
+
inputs=[chat_history, current_response],
|
264 |
+
outputs=[user_satisfaction_input, start_btn]
|
265 |
+
)
|
266 |
+
|
267 |
+
user_satisfaction_input.submit(
|
268 |
+
user_satisfaction,
|
269 |
+
inputs=[user_satisfaction_input, chat_history],
|
270 |
+
outputs=[chat_history, current_response]
|
271 |
+
).then(
|
272 |
+
chat,
|
273 |
+
inputs=[
|
274 |
+
model1_dropdown,
|
275 |
+
model2_dropdown,
|
276 |
+
user_input,
|
277 |
+
chat_history
|
278 |
+
],
|
279 |
+
outputs=[chat_history, current_response]
|
280 |
+
)
|
281 |
+
|
282 |
+
pause_btn.click(pause_conversation, outputs=[current_response])
|
283 |
+
resume_btn.click(
|
284 |
+
chat,
|
285 |
+
inputs=[
|
286 |
+
model1_dropdown,
|
287 |
+
model2_dropdown,
|
288 |
+
user_input,
|
289 |
+
chat_history,
|
290 |
+
inserted_response
|
291 |
+
],
|
292 |
+
outputs=[chat_history, current_response]
|
293 |
+
)
|
294 |
+
edit_btn.click(edit_response, inputs=[current_response], outputs=[chat_history])
|
295 |
+
rewind_btn.click(rewind_and_insert, inputs=[rewind_steps, inserted_response, chat_history], outputs=[chat_history, current_response])
|
296 |
+
restart_btn.click(
|
297 |
+
restart_conversation,
|
298 |
+
inputs=[
|
299 |
+
model1_dropdown,
|
300 |
+
model2_dropdown,
|
301 |
+
user_input
|
302 |
+
],
|
303 |
+
outputs=[chat_history, current_response]
|
304 |
+
)
|
305 |
+
clear_btn.click(manager.clear_conversation, outputs=[chat_history, current_response, user_input])
|
306 |
+
delay_slider.change(lambda x: setattr(manager, 'delay', x), inputs=[delay_slider])
|
307 |
+
|
308 |
+
if __name__ == "__main__":
|
309 |
+
demo.launch()
|