File size: 9,393 Bytes
d2693e0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import argparse
import math
import os
from collections import defaultdict

import torch
import torch.nn as nn
from tqdm import tqdm
from datasets import Dataset, DatasetDict

# --- Model Architecture (Must match the trained model) ---
def modulate(x, shift, scale):
    return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)

class TimestepEmbedder(nn.Module):
    def __init__(self, hidden_size):
        super().__init__()
        self.mlp = nn.Sequential(
            nn.Linear(1, hidden_size, bias=True), nn.SiLU(),
            nn.Linear(hidden_size, hidden_size, bias=True),
        )
    def forward(self, t):
        return self.mlp(t.unsqueeze(-1))

class DiTBlock(nn.Module):
    def __init__(self, hidden_size, n_heads):
        super().__init__()
        self.norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
        self.attn = nn.MultiheadAttention(hidden_size, n_heads, batch_first=True)
        self.norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
        self.mlp = nn.Sequential(
            nn.Linear(hidden_size, 4 * hidden_size), nn.GELU(),
            nn.Linear(4 * hidden_size, hidden_size)
        )
        self.adaLN_modulation = nn.Sequential(
            nn.SiLU(), nn.Linear(hidden_size, 6 * hidden_size, bias=True)
        )
    def forward(self, x, c):
        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(c).chunk(6, dim=1)
        x_norm1 = modulate(self.norm1(x), shift_msa, scale_msa)
        attn_output, _ = self.attn(x_norm1, x_norm1, x_norm1)
        x = x + gate_msa.unsqueeze(1) * attn_output
        x_norm2 = modulate(self.norm2(x), shift_mlp, scale_mlp)
        mlp_output = self.mlp(x_norm2)
        x = x + gate_mlp.unsqueeze(1) * mlp_output
        return x

class MDLM(nn.Module):
    def __init__(self, vocab_size, seq_len, model_dim, n_heads, n_layers):
        super().__init__()
        self.vocab_size = vocab_size
        self.seq_len = seq_len
        self.model_dim = model_dim
        self.mask_token_id = vocab_size
        self.token_embedder = nn.Embedding(vocab_size + 1, model_dim)
        self.pos_embedder = nn.Parameter(torch.randn(1, seq_len, model_dim))
        self.time_embedder = TimestepEmbedder(model_dim)
        self.transformer_blocks = nn.ModuleList([DiTBlock(model_dim, n_heads) for _ in range(n_layers)])
        self.final_norm = nn.LayerNorm(model_dim)
        self.lm_head = nn.Linear(model_dim, vocab_size)
    def forward(self, x, t):
        seq_len = x.shape[1]
        x_embed = self.token_embedder(x) + self.pos_embedder[:, :seq_len, :]
        t_embed = self.time_embedder(t)
        for block in self.transformer_blocks:
            x_embed = block(x_embed, t_embed)
        x_embed = self.final_norm(x_embed)
        logits = self.lm_head(x_embed)
        return logits

# --- Generation & Utility Functions ---

def generate_x1_from_x0(model, device, x0_batch, steps, temperature):
    model.eval()
    x = x0_batch.clone()
    num_samples, seq_len = x.shape
    keep_schedule = torch.cos(torch.linspace(math.pi / 2, 0, steps, device=device)) * seq_len
    keep_schedule = torch.round(keep_schedule).long()
    with torch.no_grad():
        for i in range(steps):
            t_continuous = torch.full((num_samples,), 1.0 - (i / steps), device=device)
            logits = model(x, t_continuous)
            scaled_logits = logits / temperature
            probs = torch.nn.functional.softmax(scaled_logits, dim=-1)
            sampled_tokens = torch.multinomial(probs.view(-1, model.vocab_size), 1).view(x.shape)
            if i == steps - 1:
                x = sampled_tokens
                break
            confidence = torch.gather(probs, 2, sampled_tokens.unsqueeze(-1)).squeeze(-1)
            num_to_keep = keep_schedule[i]
            _, indices_to_keep = torch.topk(confidence, num_to_keep, largest=True, dim=-1)
            keep_mask = torch.zeros_like(x, dtype=torch.bool).scatter_(1, indices_to_keep, True)
            x = torch.where(keep_mask, sampled_tokens, x)
    return x

def is_sample_valid(sample_x1):
    """
    Checks if special tokens [0, 1, 2, 3] appear in the middle of the sequence.
    """
    middle_sequence = sample_x1[1:-1]
    invalid_tokens = {0, 1, 2, 3}
    for token in middle_sequence:
        if token in invalid_tokens:
            return False
    return True

def create_prebatched_dataset(dataset, max_tokens_per_batch=500):
    """
    Groups samples into batches and restructures the dataset.
    Each row in the new dataset is a complete batch.
    """
    # Group samples by their length
    data_by_length = defaultdict(list)
    for sample in dataset:
        length = len(sample['input_ids_x1'])
        data_by_length[length].append(sample)

    # Create the actual batches
    batched_data = {'input_ids_x0': [], 'input_ids_x1': []}
    for length, samples in data_by_length.items():
        samples_per_batch = max(1, max_tokens_per_batch // length)
        for i in range(0, len(samples), samples_per_batch):
            batch_samples = samples[i:i + samples_per_batch]
            
            batch_x0 = [s['input_ids_x0'] for s in batch_samples]
            batch_x1 = [s['input_ids_x1'] for s in batch_samples]
            
            batched_data['input_ids_x0'].append(batch_x0)
            batched_data['input_ids_x1'].append(batch_x1)
            
    return Dataset.from_dict(batched_data)

# --- Main Execution ---

def main(args):
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    print(f"Using device: {device}")

    print(f"Loading checkpoint from {args.checkpoint}...")
    try:
        checkpoint = torch.load(args.checkpoint, map_location=device, weights_only=False)
        model_args = checkpoint['args']
    except Exception as e:
        print(f"Error loading checkpoint: {e}")
        return

    print("Initializing model...")
    model = MDLM(
        vocab_size=model_args.vocab_size,
        seq_len=model_args.seq_len,
        model_dim=model_args.model_dim,
        n_heads=model_args.n_heads,
        n_layers=model_args.n_layers
    ).to(device)
    model.load_state_dict(checkpoint['model_state_dict'])
    print("Model loaded successfully.")

    all_x0 = []
    all_x1 = []

    # 1. Generate samples for each length
    for length in range(args.min_len, args.max_len + 1):
        print(f"Generating {args.samples_per_len} valid samples for length {length}...")
        valid_samples_count = 0
        pbar = tqdm(total=args.samples_per_len)
        while valid_samples_count < args.samples_per_len:
            remaining = args.samples_per_len - valid_samples_count
            batch_size = min(args.batch_size, remaining)
            
            shape = (batch_size, length)
            x0_batch = torch.randint(0, model.vocab_size, shape, dtype=torch.long, device=device)
            x1_batch = generate_x1_from_x0(model, device, x0_batch, args.gen_steps, args.temperature)

            # 2. Perform sanity check on each sample
            for x0, x1 in zip(x0_batch, x1_batch):
                if is_sample_valid(x1.tolist()):
                    all_x0.append(x0.cpu().tolist())
                    all_x1.append(x1.cpu().tolist())
                    valid_samples_count += 1
                    pbar.update(1)
                    if valid_samples_count >= args.samples_per_len:
                        break
        pbar.close()

    # 3. Create dataset and split
    print("Splitting dataset...")
    rectified_data = {'input_ids_x0': all_x0, 'input_ids_x1': all_x1}
    dataset = Dataset.from_dict(rectified_data)
    train_test_split = dataset.train_test_split(test_size=0.2, seed=42)
    valid_test_split = train_test_split['test'].train_test_split(test_size=0.5, seed=42)
    final_dataset_dict = DatasetDict({
        'train': train_test_split['train'],
        'validation': valid_test_split['train'],
        'test': valid_test_split['test']
    })

    # 4. Pre-batch each split
    print("Pre-batching splits...")
    batched_dataset_dict = DatasetDict()
    for split_name, split_dataset in final_dataset_dict.items():
        print(f"Processing {split_name} split...")
        batched_dataset_dict[split_name] = create_prebatched_dataset(split_dataset)

    # 5. Save the final dataset
    output_path = f"{args.output_path}/v{args.version}"
    print(f"Saving new batched dataset to {output_path}...")
    batched_dataset_dict.save_to_disk(output_path)
    
    print("Rectification complete.")
    print(f"Train on this by updating your training script's dataset path to '{output_path}'.")

if __name__ == "__main__":
    parser = argparse.ArgumentParser(description="Generate a rectified dataset with variable lengths and pre-batching.")

    parser.add_argument("--checkpoint", type=str, required=True)
    parser.add_argument("--output_path", type=str, default="./rectified_datasets")
    parser.add_argument("--version", type=str, default='1')
    parser.add_argument("--samples_per_len", type=int, default=10000)
    parser.add_argument("--min_len", type=int, default=6)
    parser.add_argument("--max_len", type=int, default=49)
    parser.add_argument("--gen_steps", type=int, default=16)
    parser.add_argument("--temperature", type=float, default=1.0)
    parser.add_argument("--batch_size", type=int, default=128)

    args = parser.parse_args()
    main(args)