File size: 8,280 Bytes
05c5c96
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
LLM Distillation with GGUF Teacher (Correct Tokenizer + Stable)
"""

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.optim import AdamW
from torch.utils.data import DataLoader, Dataset
from transformers import AutoTokenizer, get_cosine_schedule_with_warmup
import logging
from pathlib import Path
from llama_cpp import Llama

logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

# ============================================================================
# GGUF TEACHER
# ============================================================================

class GGUFTeacher:
    def __init__(self, model_path, n_ctx=512, n_gpu_layers=20, n_threads=6):
        self.model = Llama(
            model_path=model_path,
            n_ctx=n_ctx,
            logits_all=True,
            n_gpu_layers=n_gpu_layers,
            n_threads=n_threads,
            verbose=False,
        )
        self.cache = {}

    def get_logits(self, input_ids):
        logits_batch = []

        for seq in input_ids:
            tokens = tuple(seq.tolist())

            if tokens in self.cache:
                logits = self.cache[tokens]
            else:
                try:
                    self.model.reset()
                    self.model.eval(tokens)

                    logits = torch.tensor(self.model._scores, dtype=torch.float32)

                    # Safety: ensure shape matches sequence
                    if logits.shape[0] != len(tokens):
                        logits = logits[:len(tokens)]

                    self.cache[tokens] = logits

                except Exception as e:
                    print("⚠️ GGUF error, skipping sequence:", e)
                    logits = torch.zeros(len(tokens), self.model.n_vocab())

            logits_batch.append(logits)

        return torch.stack(logits_batch)


# ============================================================================
# CONFIG
# ============================================================================

class DistillationConfig:
    def __init__(self):
        self.teacher_gguf_path = "/home/pragadeesh/model/mistral-7b-instruct-v0.2.Q2_K.gguf"

        self.student_hidden_dim = 512
        self.student_num_layers = 8
        self.student_num_heads = 8

        self.batch_size = 2
        self.gradient_accumulation_steps = 4
        self.learning_rate = 5e-4
        self.max_steps = 1000
        self.warmup_steps = 100

        self.temperature = 4.0
        self.max_seq_length = 128

        self.log_interval = 10


# ============================================================================
# DATASET
# ============================================================================

class TextDataset(Dataset):
    def __init__(self, texts, tokenizer, max_length=128):
        self.texts = texts
        self.tokenizer = tokenizer
        self.max_length = max_length

    def __len__(self):
        return len(self.texts)

    def __getitem__(self, idx):
        enc = self.tokenizer(
            self.texts[idx],
            padding="max_length",
            truncation=True,
            max_length=self.max_length,
            return_tensors="pt",
            add_special_tokens=True
        )

        return {
            "input_ids": enc["input_ids"].squeeze()
        }


# ============================================================================
# STUDENT MODEL
# ============================================================================

class StudentModel(nn.Module):
    def __init__(self, config, vocab_size):
        super().__init__()

        self.embedding = nn.Embedding(vocab_size, config.student_hidden_dim)
        self.pos_embedding = nn.Embedding(config.max_seq_length, config.student_hidden_dim)

        self.blocks = nn.ModuleList([
            nn.TransformerEncoderLayer(
                d_model=config.student_hidden_dim,
                nhead=config.student_num_heads,
                dim_feedforward=config.student_hidden_dim * 4,
                batch_first=True
            )
            for _ in range(config.student_num_layers)
        ])

        self.lm_head = nn.Linear(config.student_hidden_dim, vocab_size)

    def forward(self, input_ids):
        x = self.embedding(input_ids)

        pos = torch.arange(input_ids.shape[1], device=input_ids.device).unsqueeze(0)
        x = x + self.pos_embedding(pos)

        for block in self.blocks:
            x = block(x)

        return self.lm_head(x)


# ============================================================================
# LOSS
# ============================================================================

class DistillationLoss(nn.Module):
    def __init__(self, temperature=4.0):
        super().__init__()
        self.temperature = temperature
        self.kl = nn.KLDivLoss(reduction="batchmean")

    def forward(self, student_logits, teacher_logits):
        s = F.log_softmax(student_logits / self.temperature, dim=-1)
        t = F.softmax(teacher_logits / self.temperature, dim=-1)
        return self.kl(s, t)


# ============================================================================
# TRAINER
# ============================================================================

class Trainer:
    def __init__(self, config, device):
        self.config = config
        self.device = device

        logger.info("Loading Mistral tokenizer...")
        self.tokenizer = AutoTokenizer.from_pretrained(
            "mistralai/Mistral-7B-Instruct-v0.2"
        )

        # Fix padding
        self.tokenizer.pad_token = self.tokenizer.eos_token

        logger.info("Loading GGUF teacher...")
        self.teacher = GGUFTeacher(config.teacher_gguf_path)

        logger.info("Creating student...")
        self.student = StudentModel(
            config,
            self.tokenizer.vocab_size
        ).to(device)

        self.optimizer = AdamW(self.student.parameters(), lr=config.learning_rate)

        self.scheduler = get_cosine_schedule_with_warmup(
            self.optimizer,
            config.warmup_steps,
            config.max_steps
        )

        self.criterion = DistillationLoss(config.temperature)

        self.step = 0

    def train_step(self, batch):
        input_ids = batch["input_ids"].to(self.device)

        student_logits = self.student(input_ids)

        with torch.no_grad():
            teacher_logits = self.teacher.get_logits(input_ids).to(self.device)

        # Match sequence length (safety)
        min_len = min(student_logits.shape[1], teacher_logits.shape[1])
        student_logits = student_logits[:, :min_len, :]
        teacher_logits = teacher_logits[:, :min_len, :]

        loss = self.criterion(student_logits, teacher_logits)

        loss.backward()

        if self.step % self.config.gradient_accumulation_steps == 0:
            torch.nn.utils.clip_grad_norm_(self.student.parameters(), 1.0)
            self.optimizer.step()
            self.scheduler.step()
            self.optimizer.zero_grad()

        self.step += 1
        return loss.item()

    def train(self, dataloader):
        self.student.train()

        while self.step < self.config.max_steps:
            for batch in dataloader:
                loss = self.train_step(batch)

                if self.step % self.config.log_interval == 0:
                    logger.info(f"Step {self.step} | Loss: {loss:.4f}")

                if self.step >= self.config.max_steps:
                    break

        Path("checkpoints").mkdir(exist_ok=True)
        torch.save(self.student.state_dict(), "checkpoints/student.pt")

        logger.info("Training complete!")


# ============================================================================
# MAIN
# ============================================================================

def main():
    config = DistillationConfig()
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

    trainer = Trainer(config, device)

    texts = ["AI is transforming the world." * 10 for _ in range(200)]

    dataset = TextDataset(texts, trainer.tokenizer, config.max_seq_length)
    dataloader = DataLoader(dataset, batch_size=config.batch_size, shuffle=True)

    trainer.train(dataloader)


if __name__ == "__main__":
    main()