File size: 7,277 Bytes
db82745
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Knowledge distillation from a teacher LLM into Bee-Nano.

Runs on MacBook MPS / CPU. Downloads a small teacher (SmolLM2-135M),
generates logits on TinyStories, and distills them into Bee using
soft-target cross-entropy (temperature-scaled KL divergence).

This is how Bee learns WITHOUT weeks of pre-training on a GPU cluster.
"""

import argparse
import json
import logging
import os
import sys
import time
from pathlib import Path

import torch
import torch.nn.functional as F
from datasets import load_dataset
from torch.utils.data import DataLoader
from transformers import AutoTokenizer, AutoModelForCausalLM

sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
from bee.register import register
from bee.config import BeeConfig
from bee.modeling_bee import BeeForCausalLM

register()

logging.basicConfig(level=logging.INFO, format="%(asctime)s | %(levelname)s | %(name)s | %(message)s")
logger = logging.getLogger("bee.distill")


def get_args():
    parser = argparse.ArgumentParser(description="Distill teacher into Bee-Nano")
    parser.add_argument("--teacher", type=str, default="HuggingFaceTB/SmolLM2-135M", help="HF teacher model")
    parser.add_argument("--dataset", type=str, default="roneneldan/TinyStories", help="Dataset for distillation")
    parser.add_argument("--output_dir", type=str, required=True)
    parser.add_argument("--max_seq_length", type=int, default=256)
    parser.add_argument("--batch_size", type=int, default=2)
    parser.add_argument("--num_steps", type=int, default=500)
    parser.add_argument("--learning_rate", type=float, default=5e-4)
    parser.add_argument("--temperature", type=float, default=2.0, help="Softmax temperature for distillation")
    parser.add_argument("--alpha", type=float, default=0.7, help="Weight for distillation loss (1-alpha for ground-truth CE)")
    parser.add_argument("--device", type=str, default="mps" if torch.backends.mps.is_available() else "cpu")
    parser.add_argument("--save_every", type=int, default=100)
    return parser.parse_args()


def distill_step(student, teacher, input_ids, attention_mask, temperature, alpha):
    """Single distillation step. Returns loss dict."""
    with torch.no_grad():
        teacher_out = teacher(input_ids=input_ids, attention_mask=attention_mask, use_cache=False)
        teacher_logits = teacher_out.logits / temperature
        teacher_probs = F.softmax(teacher_logits, dim=-1)

    student_out = student(input_ids=input_ids, attention_mask=attention_mask, use_cache=False)
    student_logits = student_out.logits / temperature

    # Distillation loss: KL(student || teacher) on shifted targets
    shift_student = student_logits[:, :-1, :].contiguous().view(-1, student_logits.size(-1))
    shift_teacher = teacher_probs[:, 1:, :].contiguous().view(-1, teacher_probs.size(-1))

    distill_loss = F.kl_div(
        F.log_softmax(shift_student, dim=-1),
        shift_teacher,
        reduction="batchmean",
    ) * (temperature ** 2)

    # Ground-truth CE
    shift_labels = input_ids[:, 1:].contiguous().view(-1)
    ce_loss = F.cross_entropy(shift_student, shift_labels, ignore_index=-100)

    loss = alpha * distill_loss + (1 - alpha) * ce_loss
    return {"loss": loss, "distill": distill_loss.item(), "ce": ce_loss.item()}


def main():
    args = get_args()
    os.makedirs(args.output_dir, exist_ok=True)

    logger.info("Loading teacher: %s", args.teacher)
    teacher = AutoModelForCausalLM.from_pretrained(args.teacher, trust_remote_code=True)
    teacher_tokenizer = AutoTokenizer.from_pretrained(args.teacher, trust_remote_code=True)
    if teacher_tokenizer.pad_token is None:
        teacher_tokenizer.pad_token = teacher_tokenizer.eos_token
    teacher = teacher.to(args.device).eval()

    # Freeze teacher
    for p in teacher.parameters():
        p.requires_grad = False

    logger.info("Initializing Bee-Nano student")
    student_cfg = BeeConfig(
        vocab_size=teacher_tokenizer.vocab_size,
        hidden_size=512,
        num_hidden_layers=8,
        num_attention_heads=8,
        intermediate_size=1024,
        max_position_embeddings=2048,
    )
    student = BeeForCausalLM(student_cfg).to(args.device)
    n_params = sum(p.numel() for p in student.parameters())
    logger.info("Student params: %.2fM", n_params / 1e6)

    optimizer = torch.optim.AdamW(student.parameters(), lr=args.learning_rate)
    scaler = torch.cuda.amp.GradScaler() if args.device == "cuda" else None

    logger.info("Loading dataset: %s", args.dataset)
    ds = load_dataset(args.dataset, split="train", streaming=True)

    def tokenize(ex):
        return teacher_tokenizer(ex["text"], truncation=True, max_length=args.max_seq_length, padding="max_length")

    ds = ds.map(tokenize, remove_columns=["text"])
    def collate_fn(examples):
        input_ids = torch.stack([torch.tensor(ex["input_ids"]) for ex in examples])
        attention_mask = torch.stack([torch.tensor(ex["attention_mask"]) for ex in examples])
        return {"input_ids": input_ids, "attention_mask": attention_mask}
    loader = DataLoader(ds, batch_size=args.batch_size, collate_fn=collate_fn)

    logger.info("Starting distillation: %d steps", args.num_steps)
    step = 0
    losses = []
    start_time = time.perf_counter()

    for batch in loader:
        if step >= args.num_steps:
            break

        input_ids = batch["input_ids"].to(args.device)
        attention_mask = batch["attention_mask"].to(args.device)

        optimizer.zero_grad()

        if scaler:
            with torch.cuda.amp.autocast():
                loss_dict = distill_step(student, teacher, input_ids, attention_mask, args.temperature, args.alpha)
            scaler.scale(loss_dict["loss"]).backward()
            scaler.step(optimizer)
            scaler.update()
        else:
            loss_dict = distill_step(student, teacher, input_ids, attention_mask, args.temperature, args.alpha)
            loss_dict["loss"].backward()
            optimizer.step()

        losses.append(loss_dict["loss"].item())
        step += 1

        if step % 10 == 0:
            recent = losses[-10:]
            logger.info("Step %d | loss=%.4f | distill=%.4f | ce=%.4f | tok/s=%.1f",
                        step,
                        sum(recent) / len(recent),
                        loss_dict["distill"],
                        loss_dict["ce"],
                        (step * args.batch_size * args.max_seq_length) / (time.perf_counter() - start_time),
            )

        if step % args.save_every == 0:
            ckpt_dir = os.path.join(args.output_dir, f"checkpoint-{step}")
            os.makedirs(ckpt_dir, exist_ok=True)
            student.save_pretrained(ckpt_dir)
            teacher_tokenizer.save_pretrained(ckpt_dir)
            logger.info("Saved checkpoint to %s", ckpt_dir)

    # Final save
    student.save_pretrained(args.output_dir)
    teacher_tokenizer.save_pretrained(args.output_dir)

    # Save loss curve
    with open(os.path.join(args.output_dir, "loss_curve.json"), "w") as f:
        json.dump({"steps": list(range(1, len(losses) + 1)), "losses": losses}, f)

    logger.info("Distillation complete. Final avg loss (last 50): %.4f", sum(losses[-50:]) / min(len(losses), 50))


if __name__ == "__main__":
    main()