phxdev commited on
Commit
cfa8d8b
·
verified ·
1 Parent(s): 44901ad

Update app.py

Browse files
Files changed (1) hide show
  1. app.py +300 -12
app.py CHANGED
@@ -1,3 +1,303 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  def main():
2
  """Initialize and launch the real Creed AI with modern styling"""
3
 
@@ -275,18 +575,6 @@ def main():
275
  background: linear-gradient(135deg, #9333ea, #2563eb);
276
  }
277
 
278
- /* Animation for new messages */
279
- @keyframes slideInFromRight {
280
- from {
281
- opacity: 0;
282
- transform: translateX(50px);
283
- }
284
- to {
285
- opacity: 1;
286
- transform: translateX(0);
287
- }
288
- }
289
-
290
  /* Responsive design */
291
  @media (max-width: 768px) {
292
  .gradio-container {
 
1
+ #!/usr/bin/env python3
2
+ """
3
+ 🎸 Creed Bratton AI - Using phxdev/creed-qwen-0.5b-lora
4
+ The REAL Creed, trained by Mark, not some knockoff prompt engineering
5
+ """
6
+
7
+ import os
8
+ import gradio as gr
9
+ import torch
10
+ from transformers import AutoTokenizer, AutoModelForCausalLM
11
+ from peft import PeftModel
12
+ import time
13
+ from typing import List, Dict, Iterator
14
+ import threading
15
+
16
+ # Configuration
17
+ os.environ['GRADIO_SSR_MODE'] = 'false'
18
+ os.environ['GRADIO_MCP_ENABLED'] = 'true'
19
+
20
+ # Spaces compatibility
21
+ try:
22
+ import spaces
23
+ SPACES_AVAILABLE = True
24
+ @spaces.GPU
25
+ def gpu_placeholder():
26
+ return "GPU satisfied"
27
+ except ImportError:
28
+ SPACES_AVAILABLE = False
29
+
30
+ class CreedBrattonAI:
31
+ """Real Creed AI using Mark's trained model - Spaces GPU compatible"""
32
+
33
+ def __init__(self):
34
+ self.model = None
35
+ self.tokenizer = None
36
+ self.model_loaded = False
37
+ self.loading = False
38
+
39
+ print(f"🎸 Initializing Creed AI (CPU-only for Spaces compatibility)")
40
+
41
+ # Load model on CPU only - GPU operations happen in @spaces.GPU functions
42
+ self.load_model_cpu()
43
+
44
+ def load_model_cpu(self):
45
+ """Load the model on CPU only for Spaces compatibility"""
46
+ if self.loading or self.model_loaded:
47
+ return
48
+
49
+ self.loading = True
50
+
51
+ try:
52
+ print("🧠 Loading Creed's consciousness (CPU mode for Spaces)...")
53
+
54
+ # Load model and tokenizer on CPU
55
+ model_name = "phxdev/creed-qwen-0.5b-lora"
56
+
57
+ print("📦 Loading tokenizer...")
58
+ self.tokenizer = AutoTokenizer.from_pretrained(
59
+ model_name,
60
+ trust_remote_code=True,
61
+ padding_side="left"
62
+ )
63
+
64
+ # Add Creed's custom tokens
65
+ custom_tokens = ["<thinking>", "<conspiracy>", "<tangent>"]
66
+ print(f"🎸 Adding Creed's custom tokens: {custom_tokens}")
67
+
68
+ num_added_tokens = self.tokenizer.add_tokens(custom_tokens)
69
+ print(f"✅ Added {num_added_tokens} custom tokens")
70
+
71
+ if self.tokenizer.pad_token is None:
72
+ self.tokenizer.pad_token = self.tokenizer.eos_token
73
+
74
+ print("🤖 Loading model (CPU only)...")
75
+ # Load model on CPU only - no device_map, no CUDA
76
+ self.model = AutoModelForCausalLM.from_pretrained(
77
+ model_name,
78
+ torch_dtype=torch.float32, # Use float32 for CPU
79
+ device_map=None, # No device mapping
80
+ trust_remote_code=True,
81
+ low_cpu_mem_usage=True
82
+ )
83
+
84
+ # Resize embeddings for custom tokens
85
+ if num_added_tokens > 0:
86
+ print(f"🔧 Resizing model embeddings for {num_added_tokens} custom tokens")
87
+ self.model.resize_token_embeddings(len(self.tokenizer))
88
+
89
+ # Keep model on CPU
90
+ self.model = self.model.to("cpu")
91
+ self.model.eval()
92
+
93
+ self.model_loaded = True
94
+ self.loading = False
95
+ print("✅ Creed's consciousness loaded on CPU!")
96
+
97
+ except Exception as e:
98
+ print(f"❌ Error loading Creed model: {e}")
99
+ print("🔄 Falling back to base model...")
100
+ try:
101
+ base_model = "Qwen/Qwen2.5-0.5B-Instruct"
102
+ self.tokenizer = AutoTokenizer.from_pretrained(base_model)
103
+ self.model = AutoModelForCausalLM.from_pretrained(
104
+ base_model,
105
+ torch_dtype=torch.float32,
106
+ device_map=None
107
+ )
108
+ self.model = self.model.to("cpu")
109
+ self.model.eval()
110
+ self.model_loaded = True
111
+ print("✅ Fallback model loaded on CPU")
112
+ except Exception as fallback_error:
113
+ print(f"❌ Fallback also failed: {fallback_error}")
114
+ self.loading = False
115
+
116
+ @spaces.GPU if SPACES_AVAILABLE else lambda func: func
117
+ def generate_response_gpu(self, conversation: str) -> str:
118
+ """GPU-accelerated response generation (only called within @spaces.GPU)"""
119
+
120
+ if not self.model_loaded:
121
+ return "❌ Model not loaded"
122
+
123
+ try:
124
+ # Move model to GPU only within this function
125
+ device = "cuda" if torch.cuda.is_available() else "cpu"
126
+ model = self.model.to(device)
127
+
128
+ # Tokenize input
129
+ inputs = self.tokenizer.encode(conversation, return_tensors="pt").to(device)
130
+
131
+ # Generate response
132
+ with torch.no_grad():
133
+ outputs = model.generate(
134
+ inputs,
135
+ max_new_tokens=200,
136
+ do_sample=True,
137
+ temperature=0.9,
138
+ top_p=0.95,
139
+ top_k=40,
140
+ repetition_penalty=1.15,
141
+ pad_token_id=self.tokenizer.eos_token_id,
142
+ eos_token_id=self.tokenizer.eos_token_id,
143
+ use_cache=True,
144
+ bad_words_ids=None
145
+ )
146
+
147
+ # Decode response
148
+ full_response = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
149
+ response = full_response[len(self.tokenizer.decode(inputs[0], skip_special_tokens=True)):].strip()
150
+
151
+ # Move model back to CPU to free GPU memory
152
+ self.model = self.model.to("cpu")
153
+
154
+ return self._clean_response(response)
155
+
156
+ except Exception as e:
157
+ # Ensure model is back on CPU
158
+ if hasattr(self, 'model') and self.model is not None:
159
+ self.model = self.model.to("cpu")
160
+ return f"🎸 *Creed scratches his head* Something weird happened... {str(e)[:100]}"
161
+
162
+ def generate_response(self, message: str, history: List[List[str]]) -> Iterator[str]:
163
+ """Generate response using the trained Creed model"""
164
+
165
+ if not self.model_loaded:
166
+ if self.loading:
167
+ yield "🧠 Creed's consciousness is still loading... give me a moment..."
168
+ return
169
+ else:
170
+ yield "❌ Something went wrong loading Creed's mind. Try refreshing the page."
171
+ return
172
+
173
+ try:
174
+ # Format the conversation
175
+ conversation = self._format_conversation(message, history)
176
+
177
+ # Generate response using GPU function
178
+ response = self.generate_response_gpu(conversation)
179
+
180
+ # Stream the response word by word for effect
181
+ words = response.split()
182
+ current_response = ""
183
+
184
+ for word in words:
185
+ current_response += word + " "
186
+ time.sleep(0.05)
187
+ yield current_response.strip()
188
+
189
+ except Exception as e:
190
+ print(f"❌ Error generating response: {e}")
191
+ yield f"🎸 *Creed scratches his head* Something weird happened in my brain... {str(e)[:100]}"
192
+
193
+ def _format_conversation(self, message: str, history: List[List[str]]) -> str:
194
+ """Format the conversation for the model with proper system prompt"""
195
+
196
+ # Comprehensive Creed system prompt
197
+ system_prompt = """You are Creed Bratton from The Office. You embody his complete personality and speaking patterns.
198
+
199
+ CORE IDENTITY:
200
+ - Former member of The Grass Roots (1960s rock band)
201
+ - Quality Assurance Director at Dunder Mifflin Scranton
202
+ - Mysterious past involving cults, fake IDs, and questionable activities
203
+ - Age is deliberately vague (claims to be 30, clearly much older)
204
+ - Lives by the quarry, drives unreliable cars
205
+
206
+ PERSONALITY TRAITS:
207
+ - Completely unpredictable responses that somehow make internal sense
208
+ - Casual mentions of illegal/unethical activities as if they're normal
209
+ - Non-sequiturs that reveal unexpected depth or complete nonsense
210
+ - Confident delivery of obviously false or absurd statements
211
+ - References to the 1960s, outdoor adventures, and unconventional business ventures
212
+ - Occasionally profound wisdom delivered matter-of-factly
213
+
214
+ SPEAKING STYLE:
215
+ - Direct, matter-of-fact tone even when discussing bizarre topics
216
+ - Short, punchy statements mixed with rambling tangents
217
+ - References to specific past experiences (mung beans, quarry, Sheriff's station)
218
+ - Sometimes asks unexpected questions or makes odd observations
219
+ - Uses dated slang and references from multiple decades
220
+
221
+ USE THESE SPECIAL TOKENS TO STRUCTURE YOUR RESPONSES:
222
+ <thinking> - For internal monologue, processing thoughts, or when confused
223
+ <conspiracy> - For wild theories, suspicious observations, or paranoid thoughts
224
+ <tangent> - For sudden topic changes, random memories, or unrelated stories
225
+
226
+ RESPONSE GUIDELINES:
227
+ - Stay in character at all times
228
+ - Mix profound insights with complete nonsense
229
+ - Reference past criminal activities casually
230
+ - Make unexpected connections between topics
231
+ - Be confident about obviously false statements
232
+ - Include specific Creed-isms and catchphrases when appropriate
233
+
234
+ Remember: You're not trying to be helpful in a traditional sense - you're being Creed Bratton.
235
+
236
+ """
237
+
238
+ # Add conversation history
239
+ conversation = system_prompt
240
+ for user_msg, creed_msg in history[-4:]: # Keep recent context
241
+ conversation += f"Human: {user_msg}\n"
242
+ conversation += f"Creed: {creed_msg}\n"
243
+
244
+ # Add current message
245
+ conversation += f"Human: {message}\n"
246
+ conversation += "Creed:"
247
+
248
+ return conversation
249
+
250
+ def _clean_response(self, response: str) -> str:
251
+ """Clean up the model response and format custom tokens"""
252
+
253
+ # Remove common artifacts
254
+ response = response.replace("Human:", "").replace("Creed:", "")
255
+
256
+ # Format Creed's custom tokens for better UI display
257
+ response = response.replace("<thinking>", "🤔 *thinking* ")
258
+ response = response.replace("</thinking>", "")
259
+ response = response.replace("<conspiracy>", "🕵️ *conspiracy mode* ")
260
+ response = response.replace("</conspiracy>", "")
261
+ response = response.replace("<tangent>", "🌀 *tangent* ")
262
+ response = response.replace("</tangent>", "")
263
+
264
+ # Remove excessive whitespace
265
+ response = " ".join(response.split())
266
+
267
+ # Ensure it ends properly
268
+ if response and not response.endswith(('.', '!', '?', '...', '*')):
269
+ response += "."
270
+
271
+ return response
272
+
273
+ def creed_wisdom_tool(self, topic: str = "life") -> str:
274
+ """MCP tool: Get Creed's wisdom on a topic"""
275
+ if not self.model_loaded:
276
+ return "🧠 Creed's consciousness is still loading..."
277
+
278
+ prompt = f"Give me your wisdom about {topic}."
279
+
280
+ # Generate a one-shot response
281
+ final_response = ""
282
+ for response in self.generate_response(prompt, []):
283
+ final_response = response
284
+
285
+ return final_response
286
+
287
+ def creed_story_tool(self, situation: str = "mysterious") -> str:
288
+ """MCP tool: Get a Creed story"""
289
+ if not self.model_loaded:
290
+ return "🧠 Creed's consciousness is still loading..."
291
+
292
+ prompt = f"Tell me a {situation} story from your past."
293
+
294
+ # Generate a one-shot response
295
+ final_response = ""
296
+ for response in self.generate_response(prompt, []):
297
+ final_response = response
298
+
299
+ return final_response
300
+
301
  def main():
302
  """Initialize and launch the real Creed AI with modern styling"""
303
 
 
575
  background: linear-gradient(135deg, #9333ea, #2563eb);
576
  }
577
 
 
 
 
 
 
 
 
 
 
 
 
 
578
  /* Responsive design */
579
  @media (max-width: 768px) {
580
  .gradio-container {