File size: 10,814 Bytes
ba1b797
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d5537e3
ba1b797
 
 
 
 
d5537e3
 
 
 
ba1b797
 
d5537e3
ba1b797
d5537e3
ba1b797
d5537e3
 
 
 
ba1b797
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
Mixed-Precision Quantization Script for Small Language Models
Supports selective quantization of different model components with configurable bitwidths.
"""

import torch
import torch.nn as nn
from transformers import AutoModelForCausalLM, AutoTokenizer, AutoConfig
import argparse
import os
import json
from pathlib import Path
from typing import Dict, Optional, Tuple
import time

class MixedPrecisionQuantizer:
    """
    Quantizes model components with different precision levels.
    Supports more aggressive quantization for attention layers while
    preserving higher precision for FFN layers.
    """
    
    def __init__(
        self,
        model_name: str,
        attention_bits: int = 4,
        ffn_bits: int = 8,
        embedding_bits: int = 8,
        output_dir: str = "./quantized_models",
        device: str = "cuda" if torch.cuda.is_available() else "cpu"
    ):
        self.model_name = model_name
        self.attention_bits = attention_bits
        self.ffn_bits = ffn_bits
        self.embedding_bits = embedding_bits
        self.output_dir = Path(output_dir)
        self.device = device
        
        # Create output directory
        self.output_dir.mkdir(parents=True, exist_ok=True)
        
        print(f"Initializing quantizer for {model_name}")
        print(f"Attention layers: {attention_bits}-bit")
        print(f"FFN layers: {ffn_bits}-bit")
        print(f"Embeddings: {embedding_bits}-bit")
        print(f"Device: {device}")
    
    def load_model(self) -> Tuple[nn.Module, AutoTokenizer]:
        """Load the pretrained model and tokenizer."""
        print(f"\nLoading model: {self.model_name}")
        start_time = time.time()
        
        # Load with low_cpu_mem_usage for large models
        model = AutoModelForCausalLM.from_pretrained(
            self.model_name,
            torch_dtype=torch.float32,
            low_cpu_mem_usage=True,
            trust_remote_code=True
        )
        
        tokenizer = AutoTokenizer.from_pretrained(
            self.model_name,
            trust_remote_code=True
        )
        
        load_time = time.time() - start_time
        print(f"Model loaded in {load_time:.2f} seconds")
        
        # Calculate original model size
        param_count = sum(p.numel() for p in model.parameters())
        param_size_mb = sum(p.numel() * p.element_size() for p in model.parameters()) / (1024 ** 2)
        print(f"Parameters: {param_count:,} ({param_size_mb:.2f} MB)")
        
        return model, tokenizer
    
    def quantize_linear_layer(self, layer: nn.Linear, bits: int) -> nn.Linear:
        """
        Quantize a linear layer to specified bit width using symmetric quantization.
        """
        if bits == 32:
            return layer
        
        weight = layer.weight.data.clone()
        
        # Symmetric quantization
        qmin = -(2 ** (bits - 1))
        qmax = 2 ** (bits - 1) - 1
        
        # Calculate scale per-channel (per output channel)
        # This provides better accuracy than per-tensor quantization
        max_val = torch.max(torch.abs(weight), dim=1, keepdim=True)[0]
        max_val = torch.clamp(max_val, min=1e-5)  # Avoid division by zero
        scale = max_val / qmax
        
        # Quantize and dequantize (fake quantization)
        weight_q = torch.clamp(torch.round(weight / scale), qmin, qmax)
        weight_dq = weight_q * scale
        
        # Store dequantized weights as float (required for autograd)
        layer.weight.data = weight_dq.contiguous()
        
        # Store quantization metadata as layer attributes
        layer.weight_scale = scale
        layer.quantized = True
        layer.bits = bits
        
        return layer
    
    def identify_layer_type(self, name: str, module: nn.Module) -> str:
        """
        Identify if a layer is part of attention, FFN, embedding, or other components.
        """
        name_lower = name.lower()
        
        # Attention-related patterns
        attention_patterns = [
            'attn', 'attention', 'q_proj', 'k_proj', 'v_proj', 
            'qkv', 'query', 'key', 'value', 'o_proj', 'out_proj',
            'c_attn', 'c_proj'
        ]
        
        # FFN-related patterns
        ffn_patterns = [
            'mlp', 'ffn', 'fc', 'dense', 'intermediate',
            'gate_proj', 'up_proj', 'down_proj', 'w1', 'w2', 'w3'
        ]
        
        # Embedding patterns
        embedding_patterns = ['embed', 'wte', 'wpe', 'lm_head']
        
        if any(pattern in name_lower for pattern in attention_patterns):
            return 'attention'
        elif any(pattern in name_lower for pattern in ffn_patterns):
            return 'ffn'
        elif any(pattern in name_lower for pattern in embedding_patterns):
            return 'embedding'
        else:
            return 'other'
    
    def quantize_model(self, model: nn.Module) -> Tuple[nn.Module, Dict]:
        """
        Apply mixed-precision quantization to the model.
        """
        print("\nApplying mixed-precision quantization...")
        start_time = time.time()
        
        stats = {
            'attention_layers': 0,
            'ffn_layers': 0,
            'embedding_layers': 0,
            'other_layers': 0,
            'total_quantized': 0
        }
        
        # Iterate through all modules
        for name, module in model.named_modules():
            if isinstance(module, nn.Linear):
                layer_type = self.identify_layer_type(name, module)
                
                # Select quantization bitwidth based on layer type
                if layer_type == 'attention':
                    bits = self.attention_bits
                    stats['attention_layers'] += 1
                elif layer_type == 'ffn':
                    bits = self.ffn_bits
                    stats['ffn_layers'] += 1
                elif layer_type == 'embedding':
                    bits = self.embedding_bits
                    stats['embedding_layers'] += 1
                else:
                    bits = self.ffn_bits  # Default to FFN bitwidth
                    stats['other_layers'] += 1
                
                # Quantize the layer
                self.quantize_linear_layer(module, bits)
                stats['total_quantized'] += 1
        
        quant_time = time.time() - start_time
        print(f"\nQuantization completed in {quant_time:.2f} seconds")
        print(f"Quantized layers breakdown:")
        print(f"  - Attention: {stats['attention_layers']} layers ({self.attention_bits}-bit)")
        print(f"  - FFN: {stats['ffn_layers']} layers ({self.ffn_bits}-bit)")
        print(f"  - Embedding: {stats['embedding_layers']} layers ({self.embedding_bits}-bit)")
        print(f"  - Other: {stats['other_layers']} layers ({self.ffn_bits}-bit)")
        print(f"  - Total quantized: {stats['total_quantized']} layers")
        
        return model, stats
    
    def save_quantized_model(
        self,
        model: nn.Module,
        tokenizer: AutoTokenizer,
        stats: Dict
    ) -> str:
        """Save the quantized model, tokenizer, and metadata."""
        # Create model-specific output directory
        model_short_name = self.model_name.split('/')[-1]
        quant_config = f"attn{self.attention_bits}_ffn{self.ffn_bits}_emb{self.embedding_bits}"
        save_dir = self.output_dir / f"{model_short_name}_{quant_config}"
        save_dir.mkdir(parents=True, exist_ok=True)
        
        print(f"\nSaving quantized model to: {save_dir}")
        
        # Save model
        model.save_pretrained(save_dir)
        
        # Save tokenizer
        tokenizer.save_pretrained(save_dir)
        
        # Calculate quantized model size
        quantized_size_mb = sum(
            p.numel() * p.element_size() for p in model.parameters()
        ) / (1024 ** 2)
        
        # Save metadata
        metadata = {
            'original_model': self.model_name,
            'quantization_config': {
                'attention_bits': self.attention_bits,
                'ffn_bits': self.ffn_bits,
                'embedding_bits': self.embedding_bits
            },
            'layer_stats': stats,
            'model_size_mb': quantized_size_mb,
            'quantization_timestamp': time.strftime('%Y-%m-%d %H:%M:%S')
        }
        
        with open(save_dir / 'quantization_metadata.json', 'w') as f:
            json.dump(metadata, f, indent=2)
        
        print(f"Quantized model size: {quantized_size_mb:.2f} MB")
        print(f"Metadata saved to: {save_dir / 'quantization_metadata.json'}")
        
        return str(save_dir)
    
    def run(self) -> str:
        """Execute the full quantization pipeline."""
        print("=" * 80)
        print("MIXED-PRECISION QUANTIZATION PIPELINE")
        print("=" * 80)
        
        # Load model
        model, tokenizer = self.load_model()
        
        # Quantize model
        quantized_model, stats = self.quantize_model(model)
        
        # Save quantized model
        save_path = self.save_quantized_model(quantized_model, tokenizer, stats)
        
        print("\n" + "=" * 80)
        print("QUANTIZATION COMPLETE")
        print("=" * 80)
        print(f"Saved to: {save_path}")
        
        return save_path


def main():
    parser = argparse.ArgumentParser(
        description="Mixed-Precision Quantization for Small Language Models"
    )
    parser.add_argument(
        '--model_name',
        type=str,
        required=True,
        help='HuggingFace model name or path'
    )
    parser.add_argument(
        '--attention_bits',
        type=int,
        default=4,
        help='Bit width for attention layers (default: 4)'
    )
    parser.add_argument(
        '--ffn_bits',
        type=int,
        default=8,
        help='Bit width for FFN layers (default: 8)'
    )
    parser.add_argument(
        '--embedding_bits',
        type=int,
        default=8,
        help='Bit width for embedding layers (default: 8)'
    )
    parser.add_argument(
        '--output_dir',
        type=str,
        default='./quantized_models',
        help='Output directory for quantized models'
    )
    parser.add_argument(
        '--device',
        type=str,
        default='cuda' if torch.cuda.is_available() else 'cpu',
        help='Device to use (cuda/cpu)'
    )
    
    args = parser.parse_args()
    
    # Initialize quantizer
    quantizer = MixedPrecisionQuantizer(
        model_name=args.model_name,
        attention_bits=args.attention_bits,
        ffn_bits=args.ffn_bits,
        embedding_bits=args.embedding_bits,
        output_dir=args.output_dir,
        device=args.device
    )
    
    # Run quantization
    quantizer.run()


if __name__ == "__main__":
    main()