phxdev commited on
Commit
ab14d1c
·
verified ·
1 Parent(s): 49948e4

Update app.py

Browse files
Files changed (1) hide show
  1. app.py +45 -609
app.py CHANGED
@@ -1,612 +1,48 @@
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
- for response in self.generate_response(prompt, []):
282
- final_response = response
283
-
284
- return final_response
285
-
286
- def creed_story_tool(self, situation: str = "mysterious") -> str:
287
- """MCP tool: Get a Creed story"""
288
- if not self.model_loaded:
289
- return "🧠 Creed's consciousness is still loading..."
290
-
291
- prompt = f"Tell me a {situation} story from your past."
292
-
293
- # Generate a one-shot response
294
- for response in self.generate_response(prompt, []):
295
- final_response = response
296
-
297
- return final_response
298
-
299
- def main():
300
- """Initialize and launch the real Creed AI"""
301
-
302
- print("🎸 Initializing REAL Creed Bratton AI...")
303
- print("📡 Loading Mark's trained model: phxdev/creed-qwen-0.5b-lora")
304
-
305
- # Initialize Creed AI
306
- creed_ai = CreedBrattonAI()
307
-
308
- if SPACES_AVAILABLE:
309
- gpu_placeholder()
310
- print("✅ Spaces compatibility enabled")
311
-
312
- # Simple, readable CSS that actually works
313
- simple_css = """
314
- /* Basic reset */
315
- .gradio-container {
316
- max-width: 1200px !important;
317
- margin: 0 auto !important;
318
- background: #ffffff !important;
319
- font-family: -apple-system, BlinkMacSystemFont, 'Segoe UI', system-ui, sans-serif !important;
320
- color: #1f2937 !important;
321
- padding: 20px !important;
322
- }
323
-
324
- /* Header */
325
- .header {
326
- background: #ffffff !important;
327
- border: 2px solid #e5e7eb !important;
328
- border-radius: 8px !important;
329
- padding: 24px !important;
330
- margin-bottom: 24px !important;
331
- text-align: center !important;
332
- }
333
-
334
- .header h1 {
335
- font-size: 32px !important;
336
- font-weight: 600 !important;
337
- color: #111827 !important;
338
- margin: 0 0 8px 0 !important;
339
- }
340
-
341
- .header p {
342
- font-size: 16px !important;
343
- color: #6b7280 !important;
344
- margin: 0 !important;
345
- }
346
-
347
- /* Info boxes */
348
- .info-box {
349
- background: #f9fafb !important;
350
- border: 1px solid #d1d5db !important;
351
- border-radius: 6px !important;
352
- padding: 16px !important;
353
- margin: 16px 0 !important;
354
- color: #374151 !important;
355
- font-size: 14px !important;
356
- line-height: 1.5 !important;
357
- }
358
-
359
- .status-box {
360
- background: #ecfdf5 !important;
361
- border: 1px solid #a7f3d0 !important;
362
- border-radius: 6px !important;
363
- padding: 12px 16px !important;
364
- margin: 16px 0 !important;
365
- color: #065f46 !important;
366
- font-weight: 500 !important;
367
- }
368
-
369
- /* Chat area */
370
- .chat-area {
371
- background: #ffffff !important;
372
- border: 2px solid #e5e7eb !important;
373
- border-radius: 8px !important;
374
- margin: 16px 0 !important;
375
- overflow: hidden !important;
376
- }
377
-
378
- /* Tools section */
379
- .tools-area {
380
- background: #ffffff !important;
381
- border: 2px solid #e5e7eb !important;
382
- border-radius: 8px !important;
383
- padding: 24px !important;
384
- margin: 24px 0 !important;
385
- }
386
-
387
- .tools-title {
388
- font-size: 20px !important;
389
- font-weight: 600 !important;
390
- color: #111827 !important;
391
- margin: 0 0 16px 0 !important;
392
- padding-bottom: 8px !important;
393
- border-bottom: 1px solid #e5e7eb !important;
394
- }
395
-
396
- /* Form elements - force readable colors */
397
- .gradio-textbox input,
398
- .gradio-textbox textarea {
399
- background: #ffffff !important;
400
- border: 1px solid #d1d5db !important;
401
- color: #111827 !important;
402
- border-radius: 4px !important;
403
- padding: 8px 12px !important;
404
- }
405
-
406
- .gradio-textbox input:focus,
407
- .gradio-textbox textarea:focus {
408
- border-color: #3b82f6 !important;
409
- outline: none !important;
410
- box-shadow: 0 0 0 1px #3b82f6 !important;
411
- }
412
-
413
- /* Labels - force readable */
414
- .gradio-container label {
415
- color: #374151 !important;
416
- font-weight: 500 !important;
417
- font-size: 14px !important;
418
- margin-bottom: 4px !important;
419
- display: block !important;
420
- }
421
-
422
- /* Buttons */
423
- .gradio-container button {
424
- background: #3b82f6 !important;
425
- color: #ffffff !important;
426
- border: none !important;
427
- border-radius: 4px !important;
428
- padding: 8px 16px !important;
429
- font-weight: 500 !important;
430
- cursor: pointer !important;
431
- }
432
-
433
- .gradio-container button:hover {
434
- background: #2563eb !important;
435
- }
436
-
437
- /* Chat messages - ensure readability */
438
- .message {
439
- color: #111827 !important;
440
- background: #ffffff !important;
441
- }
442
-
443
- /* Footer */
444
- .footer {
445
- text-align: center !important;
446
- padding: 24px !important;
447
- color: #6b7280 !important;
448
- border-top: 1px solid #e5e7eb !important;
449
- margin-top: 32px !important;
450
- background: #ffffff !important;
451
- }
452
-
453
- /* Force text to be readable everywhere */
454
- .gradio-container * {
455
- color: inherit !important;
456
- }
457
-
458
- .gradio-container .prose {
459
- color: #111827 !important;
460
- max-width: none !important;
461
- }
462
-
463
- .gradio-container .prose h1,
464
- .gradio-container .prose h2,
465
- .gradio-container .prose h3,
466
- .gradio-container .prose p {
467
- color: #111827 !important;
468
- }
469
-
470
- /* Remove any dark mode for now */
471
- @media (prefers-color-scheme: dark) {
472
- /* Override dark mode completely */
473
- .gradio-container {
474
- background: #ffffff !important;
475
- color: #1f2937 !important;
476
- }
477
- }
478
- """
479
-
480
- # Create the interface
481
- with gr.Blocks(
482
- title="🎸 Creed Bratton AI",
483
- css=simple_css,
484
- theme=gr.themes.Default()
485
- ) as demo:
486
-
487
- # Clean header
488
- gr.HTML("""
489
- <div class="header">
490
- <h1>🎸 Creed Bratton AI</h1>
491
- <p>Powered by phxdev/creed-qwen-0.5b-lora • Enterprise-grade unpredictability</p>
492
- </div>
493
- """)
494
-
495
- # Model info
496
- gr.HTML("""
497
- <div class="info-box">
498
- <strong>Model:</strong> phxdev/creed-qwen-0.5b-lora<br>
499
- <strong>Base:</strong> Qwen 0.5B + LoRA fine-tuning<br>
500
- <strong>Tokens:</strong> &lt;thinking&gt;, &lt;conspiracy&gt;, &lt;tangent&gt;
501
- </div>
502
- """)
503
-
504
- # MCP status
505
- if os.environ.get('GRADIO_MCP_ENABLED'):
506
- gr.HTML("""
507
- <div class="status-box">
508
- ✓ MCP Server Active • Available as tool for Claude Desktop
509
- </div>
510
- """)
511
-
512
- # Main chat interface
513
- with gr.Row(elem_classes="chat-area"):
514
- chatbot = gr.Chatbot(
515
- type='messages',
516
- height=550,
517
- show_copy_button=True,
518
- show_share_button=False,
519
- avatar_images=["👤", "🎸"],
520
- bubble_full_width=False,
521
- show_label=False,
522
- placeholder="🎸 Creed is ready..."
523
- )
524
-
525
- # Input
526
- msg = gr.Textbox(
527
- placeholder="Ask Creed anything...",
528
- container=False,
529
- scale=7,
530
- submit_btn=True,
531
- stop_btn=True
532
- )
533
-
534
- # MCP Tools section
535
- with gr.Row(elem_classes="tools-area"):
536
- gr.HTML('<div class="tools-title">🛠️ MCP Tools</div>')
537
-
538
- with gr.Row():
539
- with gr.Column():
540
- wisdom_topic = gr.Textbox(
541
- label="Wisdom Topic",
542
- placeholder="life, business, relationships..."
543
- )
544
- wisdom_output = gr.Textbox(
545
- label="Creed's Response",
546
- interactive=False
547
- )
548
- wisdom_btn = gr.Button("Ask Creed", variant="primary")
549
-
550
- with gr.Column():
551
- story_situation = gr.Textbox(
552
- label="Story Request",
553
- placeholder="Tell me about..."
554
- )
555
- story_output = gr.Textbox(
556
- label="Creed's Story",
557
- interactive=False
558
- )
559
- story_btn = gr.Button("Get Story", variant="primary")
560
-
561
- # Wire up the tools
562
- wisdom_btn.click(
563
- creed_ai.creed_wisdom_tool,
564
- inputs=[wisdom_topic],
565
- outputs=[wisdom_output]
566
- )
567
-
568
- story_btn.click(
569
- creed_ai.creed_story_tool,
570
- inputs=[story_situation],
571
- outputs=[story_output]
572
- )
573
-
574
- # Chat interface
575
- chat_interface = gr.ChatInterface(
576
- fn=creed_ai.generate_response,
577
- type='messages',
578
- chatbot=chatbot,
579
- textbox=msg,
580
- title=None,
581
- description=None,
582
- examples=[
583
- "What's your philosophy on life?",
584
- "Tell me about your business ventures",
585
- "Share a story from the 60s",
586
- "What's your advice for success?"
587
- ],
588
- cache_examples=False
589
- )
590
-
591
- # Footer
592
- gr.HTML("""
593
- <div class="footer">
594
- <strong>Creed Bratton AI</strong><br>
595
- Model: phxdev/creed-qwen-0.5b-lora • Trained by Mark Scott<br>
596
- <em>"Sometimes a guy's gotta ride the bull, am I right?"</em>
597
- </div>
598
- """)
599
-
600
- # Launch
601
- print("🚀 Launching Real Creed AI...")
602
 
603
- demo.launch(
604
- ssr_mode=False,
605
- server_name="0.0.0.0",
606
- server_port=7860,
607
- share=False,
608
- show_error=True
609
- )
610
-
611
- if __name__ == "__main__":
612
- main()
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ def generate_response(self, message: Dict, history: List[Dict]) -> Iterator[str]:
2
+ """Generate response using the trained Creed model - Fixed for ChatInterface"""
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3
 
4
+ if not self.model_loaded:
5
+ if self.loading:
6
+ yield "🧠 Creed's consciousness is still loading... give me a moment..."
7
+ return
8
+ else:
9
+ yield "❌ Something went wrong loading Creed's mind. Try refreshing the page."
 
 
 
 
 
 
 
 
10
  return
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
11
 
12
+ try:
13
+ # Extract the actual message content from the message dict
14
+ user_message = message.get('text', '') if isinstance(message, dict) else str(message)
15
+
16
+ # Convert message format history to the format your model expects
17
+ formatted_history = []
18
+ if history:
19
+ for msg_pair in history:
20
+ if isinstance(msg_pair, dict):
21
+ # Handle message format
22
+ if msg_pair.get('role') == 'user':
23
+ user_msg = msg_pair.get('content', '')
24
+ # Find corresponding assistant message
25
+ assistant_msg = ""
26
+ formatted_history.append([user_msg, assistant_msg])
27
+ else:
28
+ # Handle legacy format
29
+ formatted_history.append(msg_pair)
30
+
31
+ # Format the conversation
32
+ conversation = self._format_conversation(user_message, formatted_history)
33
+
34
+ # Generate response using GPU function
35
+ response = self.generate_response_gpu(conversation)
36
+
37
+ # Stream the response word by word for effect
38
+ words = response.split()
39
+ current_response = ""
40
+
41
+ for word in words:
42
+ current_response += word + " "
43
+ time.sleep(0.05)
44
+ yield current_response.strip()
45
+
46
+ except Exception as e:
47
+ print(f"❌ Error generating response: {e}")
48
+ yield f"🎸 *Creed scratches his head* Something weird happened in my brain... {str(e)[:100]}"