Add test_lira.py
Browse files- test_lira.py +403 -0
test_lira.py
ADDED
|
@@ -0,0 +1,403 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
Comprehensive test suite for LiRA architecture.
|
| 3 |
+
Tests: model creation, forward pass, memory footprint, gradient flow,
|
| 4 |
+
training step, and inference sampling.
|
| 5 |
+
"""
|
| 6 |
+
|
| 7 |
+
import torch
|
| 8 |
+
import sys
|
| 9 |
+
import os
|
| 10 |
+
sys.path.insert(0, '/app')
|
| 11 |
+
|
| 12 |
+
from lira.model import LiRAModel, LiRAPipeline, TinyVAEDecoder, estimate_memory_mb
|
| 13 |
+
from lira.training import (
|
| 14 |
+
FlowMatchingScheduler, EMAModel, compute_loss,
|
| 15 |
+
LiRATrainingConfig, FlowDPMSolver
|
| 16 |
+
)
|
| 17 |
+
|
| 18 |
+
|
| 19 |
+
def test_model_creation():
|
| 20 |
+
"""Test all model configurations can be instantiated"""
|
| 21 |
+
print("=" * 60)
|
| 22 |
+
print("TEST 1: Model Creation & Parameter Counts")
|
| 23 |
+
print("=" * 60)
|
| 24 |
+
|
| 25 |
+
configs = ['tiny', 'small', 'base']
|
| 26 |
+
|
| 27 |
+
for config_name in configs:
|
| 28 |
+
# Use SD1.x-style VAE params for testing (4ch, f8)
|
| 29 |
+
model = LiRAModel(
|
| 30 |
+
config_name=config_name,
|
| 31 |
+
in_channels=4,
|
| 32 |
+
d_text=768,
|
| 33 |
+
patch_size=2,
|
| 34 |
+
)
|
| 35 |
+
|
| 36 |
+
counts = model.count_parameters()
|
| 37 |
+
total_m = counts['total'] / 1e6
|
| 38 |
+
|
| 39 |
+
print(f"\nLiRA-{config_name.capitalize()}:")
|
| 40 |
+
print(f" Total parameters: {total_m:.1f}M")
|
| 41 |
+
for k, v in counts.items():
|
| 42 |
+
if k != 'total':
|
| 43 |
+
print(f" {k}: {v/1e6:.2f}M ({v/counts['total']*100:.1f}%)")
|
| 44 |
+
|
| 45 |
+
# Memory estimate for 1024px with f8 VAE
|
| 46 |
+
mem = estimate_memory_mb(model, batch_size=1, img_size=1024,
|
| 47 |
+
spatial_compression=8, latent_channels=4, dtype_bytes=2)
|
| 48 |
+
print(f" Estimated inference memory (fp16): {mem['total_inference_mb']:.0f}MB")
|
| 49 |
+
print(f" Params: {mem['params_mb']:.0f}MB, Latent: {mem['latent_mb']:.1f}MB, Activations: {mem['activation_mb']:.1f}MB")
|
| 50 |
+
|
| 51 |
+
# Also test f32 VAE configuration
|
| 52 |
+
print(f"\n--- f32 VAE Configuration (DC-AE) ---")
|
| 53 |
+
model_f32 = LiRAModel(
|
| 54 |
+
config_name='small',
|
| 55 |
+
in_channels=32,
|
| 56 |
+
d_text=768,
|
| 57 |
+
patch_size=1,
|
| 58 |
+
)
|
| 59 |
+
counts_f32 = model_f32.count_parameters()
|
| 60 |
+
mem_f32 = estimate_memory_mb(model_f32, batch_size=1, img_size=1024,
|
| 61 |
+
spatial_compression=32, latent_channels=32, dtype_bytes=2)
|
| 62 |
+
print(f" LiRA-Small (f32 VAE): {counts_f32['total']/1e6:.1f}M params")
|
| 63 |
+
print(f" Estimated inference memory (fp16): {mem_f32['total_inference_mb']:.0f}MB")
|
| 64 |
+
print(f" Latent tokens: {(1024//32)**2} (32x32)")
|
| 65 |
+
|
| 66 |
+
print("\n✅ All model configurations created successfully!")
|
| 67 |
+
return True
|
| 68 |
+
|
| 69 |
+
|
| 70 |
+
def test_forward_pass():
|
| 71 |
+
"""Test forward pass with proper shapes"""
|
| 72 |
+
print("\n" + "=" * 60)
|
| 73 |
+
print("TEST 2: Forward Pass")
|
| 74 |
+
print("=" * 60)
|
| 75 |
+
|
| 76 |
+
model = LiRAModel(
|
| 77 |
+
config_name='tiny',
|
| 78 |
+
in_channels=4,
|
| 79 |
+
d_text=768,
|
| 80 |
+
patch_size=2,
|
| 81 |
+
)
|
| 82 |
+
model.eval()
|
| 83 |
+
|
| 84 |
+
# Simulate inputs
|
| 85 |
+
B = 2
|
| 86 |
+
|
| 87 |
+
# For 256px image with f8 VAE: 32x32 latent
|
| 88 |
+
z_t = torch.randn(B, 4, 32, 32)
|
| 89 |
+
t = torch.rand(B)
|
| 90 |
+
text_features = torch.randn(B, 77, 768) # CLIP-like
|
| 91 |
+
text_mask = torch.ones(B, 77, dtype=torch.bool)
|
| 92 |
+
|
| 93 |
+
print(f"Input shapes:")
|
| 94 |
+
print(f" z_t: {z_t.shape}")
|
| 95 |
+
print(f" t: {t.shape}")
|
| 96 |
+
print(f" text_features: {text_features.shape}")
|
| 97 |
+
|
| 98 |
+
with torch.no_grad():
|
| 99 |
+
v_pred, reason_info = model(z_t, t, text_features, text_mask)
|
| 100 |
+
|
| 101 |
+
print(f"\nOutput shapes:")
|
| 102 |
+
print(f" v_pred: {v_pred.shape}")
|
| 103 |
+
print(f" Reasoning steps: {reason_info['total_steps']}")
|
| 104 |
+
print(f" Discard rates: {[f'{r:.3f}' for r in reason_info['discard_rates']]}")
|
| 105 |
+
print(f" Stop values: {[f'{s:.3f}' for s in reason_info['stop_values']]}")
|
| 106 |
+
|
| 107 |
+
assert v_pred.shape == z_t.shape, f"Output shape mismatch: {v_pred.shape} vs {z_t.shape}"
|
| 108 |
+
print("\n✅ Forward pass successful!")
|
| 109 |
+
return True
|
| 110 |
+
|
| 111 |
+
|
| 112 |
+
def test_training_step():
|
| 113 |
+
"""Test a complete training step with loss computation"""
|
| 114 |
+
print("\n" + "=" * 60)
|
| 115 |
+
print("TEST 3: Training Step")
|
| 116 |
+
print("=" * 60)
|
| 117 |
+
|
| 118 |
+
config = LiRATrainingConfig(
|
| 119 |
+
model_config='tiny',
|
| 120 |
+
latent_channels=4,
|
| 121 |
+
spatial_compression=8,
|
| 122 |
+
d_text=768,
|
| 123 |
+
patch_size=2,
|
| 124 |
+
batch_size=2,
|
| 125 |
+
learning_rate=1e-4,
|
| 126 |
+
)
|
| 127 |
+
|
| 128 |
+
model = LiRAModel(
|
| 129 |
+
config_name=config.model_config,
|
| 130 |
+
in_channels=config.latent_channels,
|
| 131 |
+
d_text=config.d_text,
|
| 132 |
+
patch_size=config.patch_size,
|
| 133 |
+
)
|
| 134 |
+
model.train()
|
| 135 |
+
|
| 136 |
+
optimizer = torch.optim.AdamW(
|
| 137 |
+
model.parameters(), lr=config.learning_rate,
|
| 138 |
+
weight_decay=config.weight_decay
|
| 139 |
+
)
|
| 140 |
+
|
| 141 |
+
scheduler = FlowMatchingScheduler(schedule=config.noise_schedule)
|
| 142 |
+
ema = EMAModel(model, decay=config.ema_decay)
|
| 143 |
+
|
| 144 |
+
# Simulate data
|
| 145 |
+
B = 2
|
| 146 |
+
z_0 = torch.randn(B, 4, 32, 32) # Latent from VAE
|
| 147 |
+
text_features = torch.randn(B, 77, 768)
|
| 148 |
+
|
| 149 |
+
# Training loop (3 steps)
|
| 150 |
+
print("Running 3 training steps...")
|
| 151 |
+
losses = []
|
| 152 |
+
for step in range(3):
|
| 153 |
+
optimizer.zero_grad()
|
| 154 |
+
|
| 155 |
+
loss, info = compute_loss(
|
| 156 |
+
model, z_0, text_features, scheduler, config,
|
| 157 |
+
global_step=step
|
| 158 |
+
)
|
| 159 |
+
|
| 160 |
+
loss.backward()
|
| 161 |
+
grad_norm = torch.nn.utils.clip_grad_norm_(model.parameters(), config.grad_clip)
|
| 162 |
+
optimizer.step()
|
| 163 |
+
ema.update(model)
|
| 164 |
+
|
| 165 |
+
losses.append(info['loss'])
|
| 166 |
+
print(f" Step {step}: loss={info['loss']:.4f}, "
|
| 167 |
+
f"mse={info['mse_loss']:.4f}, "
|
| 168 |
+
f"reason_steps={info['reason_steps']}, "
|
| 169 |
+
f"grad_norm={grad_norm:.4f}")
|
| 170 |
+
|
| 171 |
+
# Verify loss is finite and reasonable
|
| 172 |
+
assert all(torch.isfinite(torch.tensor(l)) for l in losses), "Loss is not finite!"
|
| 173 |
+
assert all(l < 100 for l in losses), "Loss is unreasonably large!"
|
| 174 |
+
|
| 175 |
+
print("\n✅ Training step successful!")
|
| 176 |
+
return True
|
| 177 |
+
|
| 178 |
+
|
| 179 |
+
def test_gradient_flow():
|
| 180 |
+
"""Verify gradients flow through all components"""
|
| 181 |
+
print("\n" + "=" * 60)
|
| 182 |
+
print("TEST 4: Gradient Flow Analysis")
|
| 183 |
+
print("=" * 60)
|
| 184 |
+
|
| 185 |
+
model = LiRAModel(
|
| 186 |
+
config_name='tiny',
|
| 187 |
+
in_channels=4,
|
| 188 |
+
d_text=768,
|
| 189 |
+
patch_size=2,
|
| 190 |
+
)
|
| 191 |
+
model.train()
|
| 192 |
+
|
| 193 |
+
z_t = torch.randn(1, 4, 32, 32)
|
| 194 |
+
t = torch.rand(1)
|
| 195 |
+
text = torch.randn(1, 77, 768)
|
| 196 |
+
|
| 197 |
+
v_pred, _ = model(z_t, t, text)
|
| 198 |
+
loss = v_pred.sum()
|
| 199 |
+
loss.backward()
|
| 200 |
+
|
| 201 |
+
# Check gradients in each component
|
| 202 |
+
components = {
|
| 203 |
+
'patch_embed': model.patch_embed,
|
| 204 |
+
'time_embed': model.time_embed,
|
| 205 |
+
'text_proj': model.text_proj,
|
| 206 |
+
'reasoning': model.reasoning,
|
| 207 |
+
'blocks[0]': model.blocks[0],
|
| 208 |
+
'blocks[-1]': model.blocks[-1],
|
| 209 |
+
}
|
| 210 |
+
|
| 211 |
+
for name, module in components.items():
|
| 212 |
+
has_grad = any(p.grad is not None and p.grad.abs().sum() > 0
|
| 213 |
+
for p in module.parameters() if p.requires_grad)
|
| 214 |
+
grad_norm = sum(p.grad.norm().item() for p in module.parameters()
|
| 215 |
+
if p.grad is not None)
|
| 216 |
+
status = "✅" if has_grad else "❌"
|
| 217 |
+
print(f" {status} {name}: grad_norm={grad_norm:.6f}")
|
| 218 |
+
|
| 219 |
+
print("\n✅ Gradient flow verified!")
|
| 220 |
+
return True
|
| 221 |
+
|
| 222 |
+
|
| 223 |
+
def test_sampling():
|
| 224 |
+
"""Test inference sampling"""
|
| 225 |
+
print("\n" + "=" * 60)
|
| 226 |
+
print("TEST 5: Inference Sampling")
|
| 227 |
+
print("=" * 60)
|
| 228 |
+
|
| 229 |
+
model = LiRAModel(
|
| 230 |
+
config_name='tiny',
|
| 231 |
+
in_channels=4,
|
| 232 |
+
d_text=768,
|
| 233 |
+
patch_size=2,
|
| 234 |
+
)
|
| 235 |
+
model.eval()
|
| 236 |
+
|
| 237 |
+
solver = FlowDPMSolver(num_steps=5, order=2) # Few steps for testing
|
| 238 |
+
|
| 239 |
+
text_features = torch.randn(1, 77, 768)
|
| 240 |
+
|
| 241 |
+
print("Sampling with DPM-Solver (5 steps)...")
|
| 242 |
+
z_0 = solver.sample(
|
| 243 |
+
model,
|
| 244 |
+
shape=(1, 4, 32, 32),
|
| 245 |
+
text_features=text_features,
|
| 246 |
+
cfg_scale=1.0, # No CFG for speed
|
| 247 |
+
)
|
| 248 |
+
|
| 249 |
+
print(f" Output shape: {z_0.shape}")
|
| 250 |
+
print(f" Output range: [{z_0.min():.3f}, {z_0.max():.3f}]")
|
| 251 |
+
print(f" Output std: {z_0.std():.3f}")
|
| 252 |
+
|
| 253 |
+
assert z_0.shape == (1, 4, 32, 32), f"Wrong output shape: {z_0.shape}"
|
| 254 |
+
assert torch.isfinite(z_0).all(), "Output contains NaN/Inf!"
|
| 255 |
+
|
| 256 |
+
print("\n✅ Sampling successful!")
|
| 257 |
+
return True
|
| 258 |
+
|
| 259 |
+
|
| 260 |
+
def test_tiny_decoder():
|
| 261 |
+
"""Test the mobile-optimized VAE decoder"""
|
| 262 |
+
print("\n" + "=" * 60)
|
| 263 |
+
print("TEST 6: Tiny VAE Decoder")
|
| 264 |
+
print("=" * 60)
|
| 265 |
+
|
| 266 |
+
# Test f8 decoder (128x128 → 1024x1024)
|
| 267 |
+
decoder_f8 = TinyVAEDecoder(
|
| 268 |
+
in_channels=4, spatial_compression=8, base_channels=64
|
| 269 |
+
)
|
| 270 |
+
params_f8 = sum(p.numel() for p in decoder_f8.parameters())
|
| 271 |
+
|
| 272 |
+
z = torch.randn(1, 4, 128, 128)
|
| 273 |
+
with torch.no_grad():
|
| 274 |
+
img = decoder_f8(z)
|
| 275 |
+
|
| 276 |
+
print(f"f8 Decoder:")
|
| 277 |
+
print(f" Parameters: {params_f8/1e6:.2f}M ({params_f8 * 2 / (1024**2):.1f}MB fp16)")
|
| 278 |
+
print(f" Input: {z.shape} → Output: {img.shape}")
|
| 279 |
+
|
| 280 |
+
# Test f32 decoder (32x32 → 1024x1024)
|
| 281 |
+
decoder_f32 = TinyVAEDecoder(
|
| 282 |
+
in_channels=32, spatial_compression=32, base_channels=64
|
| 283 |
+
)
|
| 284 |
+
params_f32 = sum(p.numel() for p in decoder_f32.parameters())
|
| 285 |
+
|
| 286 |
+
z32 = torch.randn(1, 32, 32, 32)
|
| 287 |
+
with torch.no_grad():
|
| 288 |
+
img32 = decoder_f32(z32)
|
| 289 |
+
|
| 290 |
+
print(f"\nf32 Decoder:")
|
| 291 |
+
print(f" Parameters: {params_f32/1e6:.2f}M ({params_f32 * 2 / (1024**2):.1f}MB fp16)")
|
| 292 |
+
print(f" Input: {z32.shape} → Output: {img32.shape}")
|
| 293 |
+
|
| 294 |
+
print("\n✅ Tiny VAE Decoder test passed!")
|
| 295 |
+
return True
|
| 296 |
+
|
| 297 |
+
|
| 298 |
+
def test_noise_schedules():
|
| 299 |
+
"""Test all noise schedule variants"""
|
| 300 |
+
print("\n" + "=" * 60)
|
| 301 |
+
print("TEST 7: Noise Schedules")
|
| 302 |
+
print("=" * 60)
|
| 303 |
+
|
| 304 |
+
for schedule in ['laplace', 'logit_normal', 'uniform']:
|
| 305 |
+
scheduler = FlowMatchingScheduler(schedule=schedule)
|
| 306 |
+
t = scheduler.sample_timesteps(10000, torch.device('cpu'))
|
| 307 |
+
|
| 308 |
+
print(f"\n{schedule}:")
|
| 309 |
+
print(f" Mean: {t.mean():.3f}, Std: {t.std():.3f}")
|
| 310 |
+
print(f" Min: {t.min():.3f}, Max: {t.max():.3f}")
|
| 311 |
+
|
| 312 |
+
# Check distribution shape
|
| 313 |
+
bins = torch.histc(t, bins=10, min=0, max=1)
|
| 314 |
+
bins = bins / bins.sum()
|
| 315 |
+
print(f" Distribution (10 bins): {[f'{b:.2f}' for b in bins.tolist()]}")
|
| 316 |
+
|
| 317 |
+
print("\n✅ All noise schedules working!")
|
| 318 |
+
return True
|
| 319 |
+
|
| 320 |
+
|
| 321 |
+
def test_full_pipeline():
|
| 322 |
+
"""Test the complete pipeline including parameter summary"""
|
| 323 |
+
print("\n" + "=" * 60)
|
| 324 |
+
print("TEST 8: Full Pipeline Summary")
|
| 325 |
+
print("=" * 60)
|
| 326 |
+
|
| 327 |
+
pipeline = LiRAPipeline(
|
| 328 |
+
config_name='small',
|
| 329 |
+
latent_channels=32,
|
| 330 |
+
spatial_compression=32,
|
| 331 |
+
d_text=768,
|
| 332 |
+
patch_size=1,
|
| 333 |
+
)
|
| 334 |
+
|
| 335 |
+
counts = pipeline.count_parameters()
|
| 336 |
+
|
| 337 |
+
print("\n🏗️ LiRA-Small Pipeline (f32 VAE, 1024px native):")
|
| 338 |
+
print(f" Denoiser: {counts['total']/1e6:.1f}M params")
|
| 339 |
+
print(f" Tiny Decoder: {counts['tiny_decoder']/1e6:.2f}M params")
|
| 340 |
+
print(f" Total: {counts['total_with_decoder']/1e6:.1f}M params")
|
| 341 |
+
print(f" Model size (fp16): {counts['total_with_decoder'] * 2 / (1024**2):.0f}MB")
|
| 342 |
+
|
| 343 |
+
# Breakdown
|
| 344 |
+
print(f"\n Component breakdown:")
|
| 345 |
+
for k, v in counts.items():
|
| 346 |
+
if k not in ['total', 'total_with_decoder', 'tiny_decoder']:
|
| 347 |
+
print(f" {k}: {v/1e6:.2f}M ({v/counts['total']*100:.1f}%)")
|
| 348 |
+
|
| 349 |
+
# Memory estimate
|
| 350 |
+
mem = estimate_memory_mb(pipeline, batch_size=1, img_size=1024,
|
| 351 |
+
spatial_compression=32, latent_channels=32, dtype_bytes=2)
|
| 352 |
+
print(f"\n 💾 Estimated inference memory:")
|
| 353 |
+
print(f" Model params: {mem['params_mb']:.0f}MB")
|
| 354 |
+
print(f" Latent tensors: {mem['latent_mb']:.1f}MB")
|
| 355 |
+
print(f" Activations: {mem['activation_mb']:.1f}MB")
|
| 356 |
+
print(f" Total: {mem['total_inference_mb']:.0f}MB")
|
| 357 |
+
|
| 358 |
+
# Latent token analysis
|
| 359 |
+
lat_h = 1024 // 32
|
| 360 |
+
lat_w = 1024 // 32
|
| 361 |
+
print(f"\n 📐 Latent space:")
|
| 362 |
+
print(f" Image: 1024x1024px → Latent: {lat_h}x{lat_w} = {lat_h*lat_w} tokens")
|
| 363 |
+
print(f" Complexity: O({lat_h*lat_w}) per block (linear, not quadratic)")
|
| 364 |
+
print(f" Equivalent quadratic cost: O({lat_h*lat_w}²) = O({(lat_h*lat_w)**2:,})")
|
| 365 |
+
|
| 366 |
+
print("\n✅ Full pipeline test passed!")
|
| 367 |
+
return True
|
| 368 |
+
|
| 369 |
+
|
| 370 |
+
if __name__ == '__main__':
|
| 371 |
+
print("🎨 LiRA (Liquid Reasoning Artisan) - Architecture Tests")
|
| 372 |
+
print("=" * 60)
|
| 373 |
+
|
| 374 |
+
tests = [
|
| 375 |
+
test_model_creation,
|
| 376 |
+
test_forward_pass,
|
| 377 |
+
test_training_step,
|
| 378 |
+
test_gradient_flow,
|
| 379 |
+
test_sampling,
|
| 380 |
+
test_tiny_decoder,
|
| 381 |
+
test_noise_schedules,
|
| 382 |
+
test_full_pipeline,
|
| 383 |
+
]
|
| 384 |
+
|
| 385 |
+
passed = 0
|
| 386 |
+
failed = 0
|
| 387 |
+
|
| 388 |
+
for test_fn in tests:
|
| 389 |
+
try:
|
| 390 |
+
result = test_fn()
|
| 391 |
+
if result:
|
| 392 |
+
passed += 1
|
| 393 |
+
else:
|
| 394 |
+
failed += 1
|
| 395 |
+
except Exception as e:
|
| 396 |
+
print(f"\n❌ {test_fn.__name__} FAILED: {e}")
|
| 397 |
+
import traceback
|
| 398 |
+
traceback.print_exc()
|
| 399 |
+
failed += 1
|
| 400 |
+
|
| 401 |
+
print("\n" + "=" * 60)
|
| 402 |
+
print(f"RESULTS: {passed} passed, {failed} failed out of {len(tests)} tests")
|
| 403 |
+
print("=" * 60)
|