File size: 12,158 Bytes
fee179c | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 | #!/usr/bin/env python3
"""
============================================================================
LatentRecurrentFlow (LRF) β End-to-End Test Script
============================================================================
Tests the full pipeline on CPU:
1. Model creation and parameter counting
2. VAE forward pass
3. Flow matching forward pass
4. Recursive latent core forward pass
5. Full training loop (few steps)
6. Sample generation
7. Checkpoint save/load
Run: python test_lrf.py
"""
import sys
import os
import time
import torch
import traceback
# Add project root
sys.path.insert(0, '/app')
def test_model_creation():
"""Test model creation with different configs."""
print("\n[TEST 1] Model Creation")
print("-" * 40)
from lrf.model import LatentRecurrentFlow
# Test tiny config
model = LatentRecurrentFlow(LatentRecurrentFlow.tiny_config())
counts = model.count_parameters()
print("Tiny config parameters:")
for name, count in counts.items():
print(f" {name}: {count:,}")
assert counts['total'] > 0, "Model has no parameters!"
# Test default config
model_default = LatentRecurrentFlow(LatentRecurrentFlow.default_config())
counts_default = model_default.count_parameters()
print("\nDefault config parameters:")
for name, count in counts_default.items():
print(f" {name}: {count:,}")
assert counts_default['total'] > counts['total'], "Default should be larger than tiny"
print("β Model creation passed")
return True
def test_vae():
"""Test VAE forward and backward."""
print("\n[TEST 2] VAE Forward/Backward")
print("-" * 40)
from lrf.model import CompactVAE
vae = CompactVAE(in_channels=3, latent_channels=16, encoder_base_ch=32, decoder_base_ch=64)
# Count params
enc_params = sum(p.numel() for p in vae.encoder.parameters())
dec_params = sum(p.numel() for p in vae.decoder.parameters())
print(f"Encoder params: {enc_params:,}")
print(f"Decoder params: {dec_params:,}")
# Forward
x = torch.randn(2, 3, 64, 64)
recon, mean, logvar = vae(x)
print(f"Input shape: {x.shape}")
print(f"Latent shape: {mean.shape}")
print(f"Recon shape: {recon.shape}")
assert recon.shape == x.shape, f"Reconstruction shape mismatch: {recon.shape} != {x.shape}"
assert mean.shape[1] == 16, f"Latent channels mismatch: {mean.shape[1]}"
# Backward
loss = F.l1_loss(recon, x) - 0.5 * torch.mean(1 + logvar - mean.pow(2) - logvar.exp()) * 1e-6
loss.backward()
# Check gradients
grad_ok = all(p.grad is not None for p in vae.parameters() if p.requires_grad)
print(f"Gradients computed: {grad_ok}")
print("β VAE test passed")
return True
def test_gla():
"""Test Gated Linear Attention."""
print("\n[TEST 3] Gated Linear Attention")
print("-" * 40)
from lrf.model import GatedLinearAttention
gla = GatedLinearAttention(dim=64, num_heads=4, head_dim=16)
B, H, W, D = 2, 8, 8, 64
x = torch.randn(B, H * W, D)
t0 = time.time()
out = gla(x, h=H, w=W)
dt = time.time() - t0
print(f"Input: {x.shape}")
print(f"Output: {out.shape}")
print(f"Time: {dt*1000:.1f}ms")
assert out.shape == x.shape, f"Shape mismatch: {out.shape}"
# Test with larger sequence
B, H, W, D = 1, 32, 32, 64
x_large = torch.randn(B, H * W, D)
t0 = time.time()
out_large = gla(x_large, h=H, w=W)
dt_large = time.time() - t0
print(f"\nLarger input (32x32={H*W} tokens):")
print(f" Time: {dt_large*1000:.1f}ms")
print("β GLA test passed")
return True
def test_recursive_core():
"""Test the Recursive Latent Core."""
print("\n[TEST 4] Recursive Latent Core")
print("-" * 40)
from lrf.model import RecursiveLatentCore
core = RecursiveLatentCore(
dim=32,
cond_dim=64,
num_blocks=2,
num_heads=2,
head_dim=16,
T_inner=2,
T_outer=1,
use_ift_training=False,
)
params = sum(p.numel() for p in core.parameters())
print(f"Core params: {params:,}")
B, C, H, W = 2, 32, 4, 4
z_t = torch.randn(B, C, H, W)
t = torch.rand(B)
text_emb = torch.randn(B, 10, 64)
text_global = torch.randn(B, 64)
# Forward
t0 = time.time()
v = core(z_t, t, text_emb, text_global)
dt = time.time() - t0
print(f"Input shape: {z_t.shape}")
print(f"Output shape: {v.shape}")
print(f"Time: {dt*1000:.1f}ms")
assert v.shape == z_t.shape, f"Shape mismatch: {v.shape}"
# Backward
loss = v.pow(2).mean()
loss.backward()
grad_ok = sum(1 for p in core.parameters() if p.grad is not None and p.requires_grad)
total_params = sum(1 for p in core.parameters() if p.requires_grad)
print(f"Params with grad: {grad_ok}/{total_params}")
print("β Recursive core test passed")
return True
def test_ift_training():
"""Test IFT (Implicit Function Theorem) training mode."""
print("\n[TEST 5] IFT Training Mode")
print("-" * 40)
from lrf.model import RecursiveLatentCore
# Test with IFT enabled
core_ift = RecursiveLatentCore(
dim=32, cond_dim=64, num_blocks=2, num_heads=2, head_dim=16,
T_inner=3, T_outer=2, use_ift_training=True,
)
core_ift.train()
z_t = torch.randn(2, 32, 4, 4, requires_grad=True)
t = torch.rand(2)
v = core_ift(z_t, t)
loss = v.pow(2).mean()
loss.backward()
print(f"IFT mode: loss={loss.item():.4f}")
print(f" T_outer={core_ift.T_outer}, T_inner={core_ift.T_inner}")
print(f" Effective depth: {core_ift.T_outer * core_ift.T_inner * core_ift.num_blocks} layers")
print(f" Actual blocks: {core_ift.num_blocks}")
print("β IFT training test passed")
return True
def test_flow_matching():
"""Test flow matching scheduler."""
print("\n[TEST 6] Flow Matching Scheduler")
print("-" * 40)
from lrf.training import RectifiedFlowScheduler
scheduler = RectifiedFlowScheduler(shift=1.0)
z_0 = torch.randn(2, 16, 4, 4)
noise = torch.randn_like(z_0)
t = torch.tensor([0.0, 0.5])
z_t = scheduler.add_noise(z_0, noise, t)
v_target = scheduler.get_velocity_target(z_0, noise)
print(f"z_0 shape: {z_0.shape}")
print(f"z_t shape: {z_t.shape}")
print(f"v_target shape: {v_target.shape}")
# At t=0, z_t should equal z_0
t_zero = torch.tensor([0.0, 0.0])
z_t_zero = scheduler.add_noise(z_0, noise, t_zero)
diff = (z_t_zero - z_0).abs().max().item()
print(f"At t=0, |z_t - z_0| max = {diff:.6f}")
assert diff < 1e-5, f"At t=0, z_t should equal z_0, got diff={diff}"
# At t=1, z_t should equal noise
t_one = torch.tensor([1.0, 1.0])
z_t_one = scheduler.add_noise(z_0, noise, t_one)
diff_one = (z_t_one - noise).abs().max().item()
print(f"At t=1, |z_t - noise| max = {diff_one:.6f}")
assert diff_one < 1e-5, f"At t=1, z_t should equal noise, got diff={diff_one}"
print("β Flow matching test passed")
return True
def test_full_training():
"""Test full training pipeline."""
print("\n[TEST 7] Full Training Pipeline")
print("-" * 40)
from lrf.model import LatentRecurrentFlow
from lrf.training import LRFTrainer, SyntheticImageTextDataset
from torch.utils.data import DataLoader
config = LatentRecurrentFlow.tiny_config()
model = LatentRecurrentFlow(config)
trainer = LRFTrainer(model, torch.device('cpu'), '/app/test_checkpoints')
dataset = SyntheticImageTextDataset(num_samples=16, image_size=64, max_text_length=32)
dataloader = DataLoader(dataset, batch_size=4, shuffle=True)
# Stage 1: VAE
print(" Training VAE...")
vae_opt = torch.optim.AdamW(model.vae.parameters(), lr=1e-3)
for i, batch in enumerate(dataloader):
if i >= 3:
break
losses = trainer.train_vae_step(batch['image'], vae_opt)
print(f" VAE step {i}: loss={losses['total']:.4f}")
# Stage 2: Flow matching
print(" Training flow matching...")
for p in model.vae.parameters():
p.requires_grad = False
flow_params = list(model.core.parameters()) + list(model.text_encoder.parameters())
flow_opt = torch.optim.AdamW(flow_params, lr=1e-3)
for i, batch in enumerate(dataloader):
if i >= 3:
break
losses = trainer.train_flow_step(
batch['image'], batch['token_ids'], batch['attention_mask'],
flow_opt,
)
print(f" Flow step {i}: loss={losses['flow_loss']:.4f}")
# Generate
print(" Generating samples...")
sample_tokens = torch.randint(1, 31999, (2, 32))
sample_mask = torch.ones(2, 32)
images = trainer.generate(
sample_tokens, sample_mask,
num_steps=5, cfg_scale=1.0,
latent_h=4, latent_w=4,
)
print(f" Generated: {images.shape}, range=[{images.min():.3f}, {images.max():.3f}]")
# Save/load checkpoint
print(" Saving checkpoint...")
trainer.save_checkpoint('/app/test_checkpoints/test.pt', 'test', 0)
trainer.load_checkpoint('/app/test_checkpoints/test.pt')
print("β Full training pipeline test passed")
return True
def test_memory_estimate():
"""Estimate memory usage for different configs."""
print("\n[TEST 8] Memory Estimation")
print("-" * 40)
from lrf.model import LatentRecurrentFlow
configs = {
'tiny': LatentRecurrentFlow.tiny_config(),
'default': LatentRecurrentFlow.default_config(),
}
for name, config in configs.items():
model = LatentRecurrentFlow(config)
counts = model.count_parameters()
# Estimate memory
param_bytes = counts['total'] * 4 # float32
param_mb = param_bytes / (1024 * 1024)
# INT8 deployment
param_int8_mb = counts['total'] * 1 / (1024 * 1024)
print(f"\n{name} config:")
print(f" Total params: {counts['total']:,}")
print(f" FP32 size: {param_mb:.1f} MB")
print(f" INT8 size: {param_int8_mb:.1f} MB")
# Estimate activation memory for 256x256 generation
latent_h = 256 // 16
latent_w = 256 // 16
latent_tokens = latent_h * latent_w
act_bytes = 2 * latent_tokens * config['latent_channels'] * 4 # Conservative
act_mb = act_bytes / (1024 * 1024)
print(f" Est. activation memory (256x256): {act_mb:.1f} MB")
del model
print("\nβ Memory estimation passed")
return True
# Import F for backward test
import torch.nn.functional as F
def main():
"""Run all tests."""
print("=" * 60)
print("LatentRecurrentFlow (LRF) - End-to-End Tests")
print("=" * 60)
tests = [
("Model Creation", test_model_creation),
("VAE", test_vae),
("GLA", test_gla),
("Recursive Core", test_recursive_core),
("IFT Training", test_ift_training),
("Flow Matching", test_flow_matching),
("Full Training", test_full_training),
("Memory Estimate", test_memory_estimate),
]
results = []
for name, test_fn in tests:
try:
passed = test_fn()
results.append((name, passed))
except Exception as e:
print(f"\nβ {name} FAILED: {e}")
traceback.print_exc()
results.append((name, False))
print("\n" + "=" * 60)
print("Test Summary")
print("=" * 60)
all_passed = True
for name, passed in results:
status = "β PASS" if passed else "β FAIL"
print(f" {status}: {name}")
if not passed:
all_passed = False
if all_passed:
print("\nβ ALL TESTS PASSED!")
else:
print("\nβ SOME TESTS FAILED!")
sys.exit(1)
return all_passed
if __name__ == '__main__':
main()
|