File size: 9,035 Bytes
ffbd655
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""

Parallel (Non-autoregressive) Inference for B2NL-IntelligentTokenizer

Faster inference by generating all tokens at once

"""

import torch
import torch.nn.functional as F
import sys
import time
import io
from pathlib import Path

# Fix Windows Unicode
if sys.platform == 'win32':
    sys.stdout = io.TextIOWrapper(sys.stdout.buffer, encoding='utf-8')
    sys.stderr = io.TextIOWrapper(sys.stderr.buffer, encoding='utf-8')

# Add paths
sys.path.insert(0, 'core')

from unified_model import IntelligentTokenizerV62
from tokenizer import ByteTokenizerV62


class ParallelTokenizer:
    """Fast parallel generation (non-autoregressive)"""

    def __init__(self, checkpoint_path: str = None):
        self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

        if checkpoint_path is None:
            checkpoint_path = "D:/intelligent-tokenizer/intelligent-tokenizer_v6.2.1/checkpoints/v62/16.0/epoch_100.pt"

        # Load model
        self.model = IntelligentTokenizerV62()
        checkpoint = torch.load(checkpoint_path, map_location=self.device, weights_only=False)
        self.model.load_state_dict(checkpoint['model_state_dict'])
        self.model = self.model.to(self.device)
        self.model.eval()

        print(f"Model loaded on {self.device}")

    def parallel_generate(self, text: str) -> str:
        """

        Parallel generation - generate all 48 tokens at once

        This uses teacher forcing with dummy inputs

        """
        tokenizer = self.model.tokenizer

        # Encode input
        encoded = tokenizer.encode(text)
        if isinstance(encoded, dict):
            input_ids = encoded['input_ids'].unsqueeze(0) if encoded['input_ids'].dim() == 1 else encoded['input_ids']
            attention_mask = encoded['attention_mask'].unsqueeze(0) if encoded['attention_mask'].dim() == 1 else encoded['attention_mask']
        else:
            input_ids = encoded.unsqueeze(0) if encoded.dim() == 1 else encoded
            attention_mask = torch.ones_like(input_ids)

        input_ids = input_ids.to(self.device)
        attention_mask = attention_mask.to(self.device)

        # Encode
        with torch.no_grad():
            encoder_outputs = self.model.encoder(
                input_ids=input_ids,
                attention_mask=attention_mask
            )

            # Prepare all hidden states
            if 'all_hidden_states' in encoder_outputs:
                encoder_all_hidden = encoder_outputs['all_hidden_states']
            else:
                compressed = encoder_outputs.get('compressed', encoder_outputs.get('hidden_states'))
                encoder_all_hidden = [compressed] * 4

            # Create dummy decoder input (all MASK tokens)
            batch_size = input_ids.size(0)
            dummy_input = torch.full((batch_size, 48), tokenizer.MASK, device=self.device)
            dummy_input[:, 0] = tokenizer.BOS  # Start with BOS

            # Single forward pass - generate all tokens at once
            decoder_outputs = self.model.decoder(
                encoder_all_hidden=encoder_all_hidden,
                decoder_input_ids=dummy_input,
                attention_mask=torch.ones_like(dummy_input),
                use_cache=False
            )

            # Get predictions for all positions
            logits = decoder_outputs['logits']  # [batch, 48, vocab]

            # Take argmax to get predicted tokens
            predicted = torch.argmax(logits, dim=-1)  # [batch, 48]

            # Decode to text
            if predicted.dim() > 1:
                text = tokenizer.decode(predicted[0])
            else:
                text = tokenizer.decode(predicted)

        return text

    def iterative_refinement(self, text: str, iterations: int = 2) -> str:
        """

        Iterative refinement - generate multiple times and refine

        Similar to BERT-style masked prediction

        """
        tokenizer = self.model.tokenizer

        # Encode input
        encoded = tokenizer.encode(text)
        if isinstance(encoded, dict):
            input_ids = encoded['input_ids'].unsqueeze(0) if encoded['input_ids'].dim() == 1 else encoded['input_ids']
            attention_mask = encoded['attention_mask'].unsqueeze(0) if encoded['attention_mask'].dim() == 1 else encoded['attention_mask']
        else:
            input_ids = encoded.unsqueeze(0) if encoded.dim() == 1 else encoded
            attention_mask = torch.ones_like(input_ids)

        input_ids = input_ids.to(self.device)
        attention_mask = attention_mask.to(self.device)

        # Encode once
        with torch.no_grad():
            encoder_outputs = self.model.encoder(
                input_ids=input_ids,
                attention_mask=attention_mask
            )

            if 'all_hidden_states' in encoder_outputs:
                encoder_all_hidden = encoder_outputs['all_hidden_states']
            else:
                compressed = encoder_outputs.get('compressed', encoder_outputs.get('hidden_states'))
                encoder_all_hidden = [compressed] * 4

            batch_size = input_ids.size(0)

            # Start with all MASK tokens
            current = torch.full((batch_size, 48), tokenizer.MASK, device=self.device)
            current[:, 0] = tokenizer.BOS

            # Iteratively refine
            for iteration in range(iterations):
                # Forward pass with current tokens
                decoder_outputs = self.model.decoder(
                    encoder_all_hidden=encoder_all_hidden,
                    decoder_input_ids=current,
                    attention_mask=torch.ones_like(current),
                    use_cache=False
                )

                logits = decoder_outputs['logits']

                # Gradually unmask tokens (confidence-based)
                probs = F.softmax(logits, dim=-1)
                confidence = torch.max(probs, dim=-1)[0]  # [batch, 48]

                # Update tokens with high confidence
                threshold = 0.7 - (iteration * 0.1)  # Lower threshold over iterations
                high_conf_mask = confidence > threshold

                new_tokens = torch.argmax(logits, dim=-1)
                current = torch.where(high_conf_mask, new_tokens, current)

                # Add some randomness to break loops
                if iteration < iterations - 1:
                    # Randomly mask 10% of tokens for next iteration
                    rand_mask = torch.rand_like(confidence) < 0.1
                    current = torch.where(rand_mask, tokenizer.MASK, current)

        # Final decode
        if current.dim() > 1:
            text = tokenizer.decode(current[0])
        else:
            text = tokenizer.decode(current)

        return text


def test_parallel_vs_autoregressive():
    """Compare parallel vs autoregressive generation"""
    print("="*60)
    print("Parallel vs Autoregressive Comparison")
    print("="*60)

    # Load both models
    parallel = ParallelTokenizer()

    # Also test with the original autoregressive
    from inference import B2NLTokenizer
    autoregressive = B2NLTokenizer()

    test_texts = [
        "Hello, world!",
        "The quick brown fox",
        "์•ˆ๋…•ํ•˜์„ธ์š”, ๋ฐ˜๊ฐ‘์Šต๋‹ˆ๋‹ค.",
        "Testing 123",
    ]

    print("\n1. PARALLEL GENERATION (Single Pass)")
    print("-"*40)
    for text in test_texts:
        start = time.time()
        result = parallel.parallel_generate(text)
        elapsed = (time.time() - start) * 1000

        accuracy = sum(1 for i in range(min(len(text), len(result))) if text[i] == result[i]) / len(text) * 100
        print(f"Input: {text}")
        print(f"Output: {result}")
        print(f"Accuracy: {accuracy:.1f}%, Time: {elapsed:.1f}ms\n")

    print("\n2. ITERATIVE REFINEMENT (2 iterations)")
    print("-"*40)
    for text in test_texts:
        start = time.time()
        result = parallel.iterative_refinement(text, iterations=2)
        elapsed = (time.time() - start) * 1000

        accuracy = sum(1 for i in range(min(len(text), len(result))) if text[i] == result[i]) / len(text) * 100
        print(f"Input: {text}")
        print(f"Output: {result}")
        print(f"Accuracy: {accuracy:.1f}%, Time: {elapsed:.1f}ms\n")

    print("\n3. AUTOREGRESSIVE (Original - 48 steps)")
    print("-"*40)
    for text in test_texts:
        start = time.time()
        result = autoregressive.reconstruct(text, temperature=0.1)
        elapsed = (time.time() - start) * 1000

        accuracy = sum(1 for i in range(min(len(text), len(result))) if text[i] == result[i]) / len(text) * 100
        print(f"Input: {text}")
        print(f"Output: {result}")
        print(f"Accuracy: {accuracy:.1f}%, Time: {elapsed:.1f}ms\n")


if __name__ == "__main__":
    test_parallel_vs_autoregressive()