File size: 12,184 Bytes
e911248
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ae78897
e911248
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
import gradio as gr
import torch
import torch.nn.functional as F
from transformers import AutoModelForCausalLM, AutoTokenizer
import numpy as np
from typing import List, Dict, Tuple
import json
import os
from datetime import datetime

class GRPOTrainer:
    def __init__(self):
        self.model = None
        self.ref_model = None
        self.tokenizer = None
        self.optimizer = None
        self.training_history = []
        
    def load_model(self, model_name: str) -> str:
        """Load the model and tokenizer"""
        try:
            self.tokenizer = AutoTokenizer.from_pretrained(model_name)
            self.model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float16)
            self.ref_model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float16)
            
            # Set padding token
            if self.tokenizer.pad_token is None:
                self.tokenizer.pad_token = self.tokenizer.eos_token
                
            # Freeze reference model
            for param in self.ref_model.parameters():
                param.requires_grad = False
                
            return f"βœ… Successfully loaded model: {model_name}"
        except Exception as e:
            return f"❌ Error loading model: {str(e)}"
    
    def compute_rewards(self, prompts: List[str], responses: List[str]) -> torch.Tensor:
        """Compute rewards for responses (simplified reward function)"""
        rewards = []
        for response in responses:
            # Simple reward based on response length and diversity
            length_reward = min(len(response.split()) / 50, 1.0)
            unique_words = len(set(response.lower().split()))
            diversity_reward = min(unique_words / 20, 1.0)
            reward = (length_reward + diversity_reward) / 2
            rewards.append(reward)
        return torch.tensor(rewards)
    
    def compute_kl_penalty(self, logits: torch.Tensor, ref_logits: torch.Tensor) -> torch.Tensor:
        """Compute KL divergence penalty"""
        probs = F.softmax(logits, dim=-1)
        ref_probs = F.softmax(ref_logits, dim=-1)
        kl = (probs * (probs / ref_probs).log()).sum(-1)
        return kl.mean()
    
    def grpo_step(self, prompts: List[str], beta: float = 0.1) -> Dict:
        """Perform one GRPO training step"""
        if not self.model or not self.tokenizer:
            return {"error": "Model not loaded"}
        
        # Tokenize prompts
        inputs = self.tokenizer(prompts, return_tensors="pt", padding=True, truncation=True)
        
        # Generate responses
        with torch.no_grad():
            outputs = self.model.generate(
                inputs.input_ids,
                max_length=inputs.input_ids.shape[1] + 50,
                do_sample=True,
                temperature=0.8,
                pad_token_id=self.tokenizer.pad_token_id
            )
        
        # Decode responses
        responses = []
        for output in outputs:
            response = self.tokenizer.decode(output[inputs.input_ids.shape[1]:], skip_special_tokens=True)
            responses.append(response)
        
        # Compute rewards
        rewards = self.compute_rewards(prompts, responses)
        
        # Forward pass through both models
        self.model.train()
        model_outputs = self.model(inputs.input_ids)
        ref_outputs = self.ref_model(inputs.input_ids)
        
        # Compute KL penalty
        kl_penalty = self.compute_kl_penalty(model_outputs.logits, ref_outputs.logits)
        
        # Compute loss (simplified GRPO loss)
        loss = -rewards.mean() + beta * kl_penalty
        
        # Backward pass
        if self.optimizer:
            self.optimizer.zero_grad()
            loss.backward()
            self.optimizer.step()
        
        return {
            "loss": loss.item(),
            "reward": rewards.mean().item(),
            "kl_penalty": kl_penalty.item(),
            "responses": responses
        }
    
    def train(self, prompts: List[str], num_steps: int, lr: float, beta: float) -> str:
        """Run GRPO training"""
        if not self.model:
            return "❌ Please load a model first"
        
        # Initialize optimizer
        self.optimizer = torch.optim.AdamW(self.model.parameters(), lr=lr)
        
        results = []
        for step in range(num_steps):
            step_result = self.grpo_step(prompts, beta)
            
            if "error" in step_result:
                return f"❌ Error: {step_result['error']}"
            
            result_str = f"Step {step + 1}/{num_steps} - Loss: {step_result['loss']:.4f}, Reward: {step_result['reward']:.4f}, KL: {step_result['kl_penalty']:.4f}"
            results.append(result_str)
            
            # Store training history
            self.training_history.append({
                "step": step + 1,
                "loss": step_result['loss'],
                "reward": step_result['reward'],
                "kl_penalty": step_result['kl_penalty']
            })
        
        return "\n".join(results)
    
    def generate_response(self, prompt: str, max_length: int = 100, temperature: float = 0.8) -> str:
        """Generate a response using the trained model"""
        if not self.model or not self.tokenizer:
            return "❌ Please load a model first"
        
        inputs = self.tokenizer(prompt, return_tensors="pt")
        
        with torch.no_grad():
            outputs = self.model.generate(
                inputs.input_ids,
                max_length=inputs.input_ids.shape[1] + max_length,
                temperature=temperature,
                do_sample=True,
                pad_token_id=self.tokenizer.pad_token_id
            )
        
        response = self.tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
        return response
    
    def save_model(self, save_path: str) -> str:
        """Save the trained model"""
        if not self.model:
            return "❌ No model to save"
        
        try:
            self.model.save_pretrained(save_path)
            self.tokenizer.save_pretrained(save_path)
            
            # Save training history
            with open(os.path.join(save_path, "training_history.json"), "w") as f:
                json.dump(self.training_history, f)
            
            return f"βœ… Model saved to {save_path}"
        except Exception as e:
            return f"❌ Error saving model: {str(e)}"

# Initialize trainer
trainer = GRPOTrainer()

# Gradio interface
def load_model_interface(model_name):
    return trainer.load_model(model_name)

def train_interface(prompts_text, num_steps, learning_rate, beta):
    prompts = [p.strip() for p in prompts_text.split("\n") if p.strip()]
    if not prompts:
        return "❌ Please provide at least one prompt"
    return trainer.train(prompts, int(num_steps), float(learning_rate), float(beta))

def generate_interface(prompt, max_length, temperature):
    return trainer.generate_response(prompt, int(max_length), float(temperature))

def save_model_interface(save_path):
    return trainer.save_model(save_path)

def get_training_history():
    if not trainer.training_history:
        return "No training history available"
    
    history_str = "Training History:\n"
    history_str += "-" * 50 + "\n"
    for entry in trainer.training_history[-10:]:  # Show last 10 entries
        history_str += f"Step {entry['step']}: Loss={entry['loss']:.4f}, Reward={entry['reward']:.4f}, KL={entry['kl_penalty']:.4f}\n"
    return history_str

# Create Gradio interface
with gr.Blocks(title="GRPO Model Training") as app:
    gr.Markdown("# πŸš€ GRPO (Group Relative Policy Optimization) Training App")
    gr.Markdown("Train language models using GRPO technique with this simple interface")
    
    with gr.Tab("πŸ”§ Model Setup"):
        with gr.Row():
            model_input = gr.Textbox(
                label="Model Name",
                value="Writer/Palmyra-56B-Instruct",
                placeholder="Enter HuggingFace model name (e.g., Palmyra, Qwen, Llama)"
            )
            load_btn = gr.Button("Load Model", variant="primary")
        
        model_status = gr.Textbox(label="Status", lines=2)
        load_btn.click(load_model_interface, inputs=model_input, outputs=model_status)
    
    with gr.Tab("🎯 Training"):
        with gr.Row():
            with gr.Column():
                prompts_input = gr.Textbox(
                    label="Training Prompts (one per line)",
                    lines=5,
                    value="Tell me about artificial intelligence\nExplain quantum computing\nWhat is machine learning?",
                    placeholder="Enter your prompts here..."
                )
            
            with gr.Column():
                num_steps_input = gr.Slider(
                    label="Number of Training Steps",
                    minimum=1,
                    maximum=100,
                    value=10,
                    step=1
                )
                lr_input = gr.Number(
                    label="Learning Rate",
                    value=1e-5,
                    step=1e-6
                )
                beta_input = gr.Number(
                    label="KL Penalty Weight (Ξ²)",
                    value=0.1,
                    step=0.01
                )
        
        train_btn = gr.Button("Start Training", variant="primary")
        training_output = gr.Textbox(label="Training Progress", lines=10)
        
        train_btn.click(
            train_interface,
            inputs=[prompts_input, num_steps_input, lr_input, beta_input],
            outputs=training_output
        )
    
    with gr.Tab("πŸ’¬ Generation"):
        with gr.Row():
            with gr.Column():
                gen_prompt = gr.Textbox(
                    label="Prompt",
                    placeholder="Enter your prompt here...",
                    value="Tell me about"
                )
                max_length = gr.Slider(
                    label="Max Length",
                    minimum=10,
                    maximum=500,
                    value=100,
                    step=10
                )
                temp_slider = gr.Slider(
                    label="Temperature",
                    minimum=0.1,
                    maximum=2.0,
                    value=0.8,
                    step=0.1
                )
            
            with gr.Column():
                gen_btn = gr.Button("Generate", variant="primary")
                gen_output = gr.Textbox(label="Generated Response", lines=10)
        
        gen_btn.click(
            generate_interface,
            inputs=[gen_prompt, max_length, temp_slider],
            outputs=gen_output
        )
    
    with gr.Tab("πŸ’Ύ Save Model"):
        save_path_input = gr.Textbox(
            label="Save Path",
            value="./grpo_trained_model",
            placeholder="Enter path to save the model"
        )
        save_btn = gr.Button("Save Model", variant="primary")
        save_status = gr.Textbox(label="Save Status")
        
        save_btn.click(save_model_interface, inputs=save_path_input, outputs=save_status)
    
    with gr.Tab("πŸ“Š Training History"):
        history_btn = gr.Button("Refresh History", variant="secondary")
        history_output = gr.Textbox(label="Training History", lines=15)
        
        history_btn.click(get_training_history, outputs=history_output)
    
    gr.Markdown("""
    ## πŸ“ Instructions:
    1. **Load Model**: Start by loading a pre-trained model from HuggingFace
    2. **Training**: Add your prompts and configure training parameters
    3. **Generation**: Test your trained model with custom prompts
    4. **Save**: Save your fine-tuned model for later use
    
    ## ⚠️ Note:
    - This is a simplified GRPO implementation for demonstration
    - For production use, consider more sophisticated reward functions
    - GPU recommended for larger models
    """)

# Launch the app
if __name__ == "__main__":
    app.launch(share=True)