File size: 8,990 Bytes
3d8856d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
Model evaluation and testing utilities for TTV-1B
"""

import torch
import torch.nn as nn
from video_ttv_1b import VideoTTV1B, create_model
import time
from typing import Dict, Tuple
import numpy as np


def count_parameters(model: nn.Module) -> Dict[str, int]:
    """Count parameters by component"""
    total = 0
    breakdown = {}
    
    # Text encoder
    text_params = sum(p.numel() for p in model.text_encoder.parameters())
    breakdown['text_encoder'] = text_params
    total += text_params
    
    # Patch embedding
    patch_params = sum(p.numel() for p in model.patch_embed.parameters())
    breakdown['patch_embed'] = patch_params
    total += patch_params
    
    # DiT blocks
    dit_params = sum(p.numel() for p in model.blocks.parameters())
    breakdown['dit_blocks'] = dit_params
    total += dit_params
    
    # Other
    other_params = sum(p.numel() for p in model.parameters()) - total
    breakdown['other'] = other_params
    total += other_params
    
    breakdown['total'] = total
    
    return breakdown


def measure_inference_speed(
    model: nn.Module,
    batch_size: int = 1,
    num_iterations: int = 10,
    device: str = 'cuda',
) -> Dict[str, float]:
    """Measure inference speed"""
    model.eval()
    
    # Prepare dummy inputs
    videos = torch.randn(batch_size, 3, 16, 256, 256).to(device)
    timesteps = torch.randint(0, 1000, (batch_size,)).to(device)
    text_tokens = torch.randint(0, 50257, (batch_size, 256)).to(device)
    
    # Warmup
    with torch.no_grad():
        for _ in range(3):
            _ = model(videos, timesteps, text_tokens)
    
    # Measure
    if device == 'cuda':
        torch.cuda.synchronize()
    
    start_time = time.time()
    
    with torch.no_grad():
        for _ in range(num_iterations):
            _ = model(videos, timesteps, text_tokens)
            if device == 'cuda':
                torch.cuda.synchronize()
    
    end_time = time.time()
    
    total_time = end_time - start_time
    avg_time = total_time / num_iterations
    throughput = batch_size / avg_time
    
    return {
        'total_time': total_time,
        'avg_time_per_batch': avg_time,
        'throughput': throughput,
        'time_per_sample': avg_time / batch_size,
    }


def measure_memory_usage(
    model: nn.Module,
    batch_size: int = 1,
    device: str = 'cuda',
) -> Dict[str, float]:
    """Measure memory usage"""
    if device != 'cuda':
        return {'error': 'Memory measurement only available on CUDA'}
    
    torch.cuda.reset_peak_memory_stats()
    torch.cuda.empty_cache()
    
    # Model memory
    model_memory = sum(p.numel() * p.element_size() for p in model.parameters())
    model_memory_mb = model_memory / (1024 ** 2)
    
    # Forward pass memory
    videos = torch.randn(batch_size, 3, 16, 256, 256).to(device)
    timesteps = torch.randint(0, 1000, (batch_size,)).to(device)
    text_tokens = torch.randint(0, 50257, (batch_size, 256)).to(device)
    
    torch.cuda.reset_peak_memory_stats()
    
    with torch.no_grad():
        _ = model(videos, timesteps, text_tokens)
    
    peak_memory = torch.cuda.max_memory_allocated()
    peak_memory_mb = peak_memory / (1024 ** 2)
    
    return {
        'model_memory_mb': model_memory_mb,
        'peak_memory_mb': peak_memory_mb,
        'activation_memory_mb': peak_memory_mb - model_memory_mb,
    }


def test_model_correctness(model: nn.Module, device: str = 'cuda') -> bool:
    """Test model correctness with various inputs"""
    model.eval()
    
    tests_passed = 0
    total_tests = 0
    
    # Test 1: Output shape
    total_tests += 1
    x = torch.randn(2, 3, 16, 256, 256).to(device)
    t = torch.randint(0, 1000, (2,)).to(device)
    tokens = torch.randint(0, 50257, (2, 256)).to(device)
    
    with torch.no_grad():
        output = model(x, t, tokens)
    
    if output.shape == x.shape:
        tests_passed += 1
        print("βœ“ Test 1 passed: Output shape matches input")
    else:
        print(f"βœ— Test 1 failed: Expected {x.shape}, got {output.shape}")
    
    # Test 2: No NaN values
    total_tests += 1
    if not torch.isnan(output).any():
        tests_passed += 1
        print("βœ“ Test 2 passed: No NaN values in output")
    else:
        print("βœ— Test 2 failed: NaN values detected in output")
    
    # Test 3: Different timesteps produce different outputs
    total_tests += 1
    t1 = torch.full((2,), 0).to(device)
    t2 = torch.full((2,), 999).to(device)
    
    with torch.no_grad():
        out1 = model(x, t1, tokens)
        out2 = model(x, t2, tokens)
    
    if not torch.allclose(out1, out2, rtol=1e-3):
        tests_passed += 1
        print("βœ“ Test 3 passed: Different timesteps produce different outputs")
    else:
        print("βœ— Test 3 failed: Outputs identical for different timesteps")
    
    # Test 4: Different text produces different outputs
    total_tests += 1
    tokens1 = torch.randint(0, 50257, (2, 256)).to(device)
    tokens2 = torch.randint(0, 50257, (2, 256)).to(device)
    
    with torch.no_grad():
        out1 = model(x, t, tokens1)
        out2 = model(x, t, tokens2)
    
    if not torch.allclose(out1, out2, rtol=1e-3):
        tests_passed += 1
        print("βœ“ Test 4 passed: Different text produces different outputs")
    else:
        print("βœ— Test 4 failed: Outputs identical for different text")
    
    # Test 5: Gradient flow (training mode)
    total_tests += 1
    model.train()
    x.requires_grad = True
    output = model(x, t, tokens)
    loss = output.mean()
    loss.backward()
    
    if x.grad is not None and not torch.isnan(x.grad).any():
        tests_passed += 1
        print("βœ“ Test 5 passed: Gradients computed correctly")
    else:
        print("βœ— Test 5 failed: Gradient computation error")
    
    model.eval()
    
    print(f"\nTests passed: {tests_passed}/{total_tests}")
    return tests_passed == total_tests


def benchmark_full_pipeline(device: str = 'cuda'):
    """Comprehensive benchmark of the model"""
    print("="*60)
    print("TTV-1B Model Benchmark")
    print("="*60)
    
    # Create model
    print("\n1. Creating model...")
    model = create_model(device)
    print(f"   Device: {device}")
    
    # Count parameters
    print("\n2. Parameter count:")
    param_counts = count_parameters(model)
    for name, count in param_counts.items():
        print(f"   {name:20s}: {count:>12,} ({count/1e6:>6.1f}M)")
    
    # Memory usage
    if device == 'cuda':
        print("\n3. Memory usage:")
        mem_stats = measure_memory_usage(model, batch_size=1, device=device)
        for name, value in mem_stats.items():
            print(f"   {name:25s}: {value:>8.1f} MB")
    
    # Inference speed
    print("\n4. Inference speed:")
    speed_stats = measure_inference_speed(model, batch_size=1, num_iterations=10, device=device)
    print(f"   Average time per batch:  {speed_stats['avg_time_per_batch']:.3f} seconds")
    print(f"   Time per sample:         {speed_stats['time_per_sample']:.3f} seconds")
    print(f"   Throughput:              {speed_stats['throughput']:.2f} samples/sec")
    
    # Correctness tests
    print("\n5. Correctness tests:")
    all_passed = test_model_correctness(model, device)
    
    print("\n" + "="*60)
    if all_passed:
        print("βœ“ All tests passed!")
    else:
        print("βœ— Some tests failed")
    print("="*60)


def estimate_training_time(
    num_samples: int = 1_000_000,
    batch_size: int = 16,
    num_epochs: int = 100,
    seconds_per_batch: float = 2.0,
) -> Dict[str, float]:
    """Estimate training time"""
    steps_per_epoch = num_samples // batch_size
    total_steps = steps_per_epoch * num_epochs
    total_seconds = total_steps * seconds_per_batch
    
    return {
        'steps_per_epoch': steps_per_epoch,
        'total_steps': total_steps,
        'total_hours': total_seconds / 3600,
        'total_days': total_seconds / (3600 * 24),
    }


if __name__ == "__main__":
    # Run full benchmark
    device = 'cuda' if torch.cuda.is_available() else 'cpu'
    benchmark_full_pipeline(device)
    
    # Training time estimates
    print("\n" + "="*60)
    print("Training Time Estimates")
    print("="*60)
    
    configs = [
        {'name': 'Single A100 (bs=2, grad_accum=8)', 'batch_size': 16, 'seconds_per_batch': 3.0},
        {'name': '8x A100 (bs=16, grad_accum=8)', 'batch_size': 128, 'seconds_per_batch': 3.0},
    ]
    
    for config in configs:
        print(f"\n{config['name']}:")
        estimates = estimate_training_time(
            num_samples=10_000_000,
            batch_size=config['batch_size'],
            num_epochs=10,
            seconds_per_batch=config['seconds_per_batch'],
        )
        print(f"  Steps per epoch: {estimates['steps_per_epoch']:,}")
        print(f"  Total steps:     {estimates['total_steps']:,}")
        print(f"  Estimated time:  {estimates['total_days']:.1f} days ({estimates['total_hours']:.1f} hours)")