File size: 12,837 Bytes
388fd6e | 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 | """Tests for Gamma SSM Block."""
import torch
import pytest
from gamma_space_model import GammaSingleBlock
class TestGammaSingleBlockInitialization:
"""Test GammaSingleBlock initialization."""
def test_direct_parameter_init(self):
"""Test GammaSingleBlock with direct parameters (no config)."""
block = GammaSingleBlock(
d_model=16,
hidden_dim=32,
delta_t=0.1,
kernel_length=4,
A_type="tridiagonal",
prenorm=True,
residual_scale=1.0,
dropout=0.0,
)
assert block.d_model == 16
assert block.prenorm is True
assert block.residual_scale == 1.0
assert block.dropout_p == 0.0
def test_default_parameters(self):
"""Test that default parameters are set correctly."""
block = GammaSingleBlock(d_model=16, hidden_dim=32)
assert block.d_model == 16
assert block.prenorm is True
assert block.residual_scale == 1.0
assert block.dropout_p == 0.0
assert block.ssm.delta_t == 0.1
def test_ssm_instantiation(self):
"""Test that SSM block is correctly instantiated."""
block = GammaSingleBlock(
d_model=16,
hidden_dim=32,
delta_t=0.2,
A_type="tridiagonal",
)
assert block.ssm.state_dim == 16
assert block.ssm.hidden_dim == 32
assert block.ssm.delta_t == 0.2
assert block.ssm.A_type == "tridiagonal"
class TestGammaSingleBlockForwardPass:
"""Test GammaSingleBlock forward pass."""
def test_forward_output_shape(self):
"""Test that forward pass produces correct output shape."""
batch_size, seq_len, d_model = 4, 32, 16
hidden_dim = 32
block = GammaSingleBlock(d_model=d_model, hidden_dim=hidden_dim)
x = torch.randn(batch_size, seq_len, d_model)
output, final_state = block(x)
assert output.shape == (batch_size, seq_len, d_model)
assert final_state.shape == (batch_size, hidden_dim)
def test_forward_with_initial_state(self):
"""Test forward pass with provided initial state."""
batch_size, seq_len, d_model, hidden_dim = 2, 16, 8, 16
block = GammaSingleBlock(d_model=d_model, hidden_dim=hidden_dim)
x = torch.randn(batch_size, seq_len, d_model)
initial_state = torch.zeros(batch_size, hidden_dim)
output1, final_state1 = block(x, state=initial_state)
output2, final_state2 = block(x, state=None)
# Should produce the same results since initial_state defaults to zeros
assert torch.allclose(output1, output2, atol=1e-5)
class TestGammaSingleBlockNormalization:
"""Test GammaSingleBlock normalization (prenorm vs postnorm)."""
def test_prenorm_configuration(self):
"""Test prenorm configuration."""
block = GammaSingleBlock(d_model=16, hidden_dim=32, prenorm=True)
x = torch.randn(2, 10, 16)
output, _ = block(x)
assert output.shape == (2, 10, 16)
def test_postnorm_configuration(self):
"""Test postnorm configuration."""
block = GammaSingleBlock(d_model=16, hidden_dim=32, prenorm=False)
x = torch.randn(2, 10, 16)
output, _ = block(x)
assert output.shape == (2, 10, 16)
def test_prenorm_vs_postnorm_outputs_differ(self):
"""Test that prenorm and postnorm produce different outputs."""
x = torch.randn(2, 10, 16)
prenorm_block = GammaSingleBlock(d_model=16, hidden_dim=32, prenorm=True)
postnorm_block = GammaSingleBlock(d_model=16, hidden_dim=32, prenorm=False)
output_pre, _ = prenorm_block(x)
output_post, _ = postnorm_block(x)
# Outputs should differ but have same shape
assert output_pre.shape == output_post.shape
assert not torch.allclose(output_pre, output_post)
class TestGammaSingleBlockResidualConnection:
"""Test GammaSingleBlock residual connection."""
def test_residual_with_scale_1(self):
"""Test residual connection with scale=1.0."""
block = GammaSingleBlock(
d_model=16,
hidden_dim=32,
residual_scale=1.0,
prenorm=True,
)
x = torch.randn(2, 10, 16)
output, _ = block(x)
# Output should be x + SSM_output (approximately close to x)
assert torch.allclose(output, x, atol=2.0)
def test_residual_with_scale_0(self):
"""Test residual connection with scale=0.0 (no residual)."""
block = GammaSingleBlock(
d_model=16,
hidden_dim=32,
residual_scale=0.0,
prenorm=True,
)
x = torch.randn(2, 10, 16)
output, _ = block(x)
# Output should be purely from norm + SSM (not affected by input)
# It will still be different from x
assert not torch.allclose(output, x)
def test_residual_scale_effect(self):
"""Test that residual_scale parameter affects output."""
x = torch.randn(2, 10, 16)
block1 = GammaSingleBlock(
d_model=16,
hidden_dim=32,
residual_scale=0.5,
prenorm=True,
)
block2 = GammaSingleBlock(
d_model=16,
hidden_dim=32,
residual_scale=2.0,
prenorm=True,
)
output1, _ = block1(x)
output2, _ = block2(x)
# Different scales should produce different outputs
assert not torch.allclose(output1, output2)
class TestGammaSingleBlockDropout:
"""Test GammaSingleBlock dropout."""
def test_dropout_train_mode(self):
"""Test that dropout is applied during training."""
block = GammaSingleBlock(
d_model=16,
hidden_dim=32,
dropout=0.5,
)
block.train()
x = torch.randn(2, 10, 16)
# Multiple forward passes should give different results due to dropout
output1, _ = block(x)
output2, _ = block(x)
assert not torch.allclose(output1, output2)
def test_dropout_eval_mode(self):
"""Test that dropout is not applied during evaluation."""
block = GammaSingleBlock(
d_model=16,
hidden_dim=32,
dropout=0.5,
)
block.eval()
x = torch.randn(2, 10, 16)
# Multiple forward passes should give same results in eval mode
output1, _ = block(x)
output2, _ = block(x)
assert torch.allclose(output1, output2)
def test_no_dropout_with_zero_dropout_rate(self):
"""Test that no dropout is applied when dropout=0."""
block = GammaSingleBlock(
d_model=16,
hidden_dim=32,
dropout=0.0,
)
# Should not have dropout layer
assert block.dropout is None
class TestGammaSingleBlockMasking:
"""Test GammaSingleBlock masking functionality."""
def test_forward_with_mask(self):
"""Test forward pass with masking."""
batch_size, seq_len, d_model = 2, 10, 16
hidden_dim = 32
block = GammaSingleBlock(d_model=d_model, hidden_dim=hidden_dim)
x = torch.randn(batch_size, seq_len, d_model)
mask = torch.ones(batch_size, seq_len, dtype=torch.bool)
mask[1, 5:] = False
output, _ = block(x, mask=mask)
assert output.shape == (batch_size, seq_len, d_model)
class TestGammaSingleBlockGradients:
"""Test gradient flow through GammaSingleBlock."""
def test_backward_pass(self):
"""Test that gradients flow correctly."""
batch_size, seq_len, d_model = 2, 10, 16
hidden_dim = 32
block = GammaSingleBlock(d_model=d_model, hidden_dim=hidden_dim)
x = torch.randn(batch_size, seq_len, d_model, requires_grad=True)
output, _ = block(x)
loss = output.sum()
loss.backward()
assert x.grad is not None
assert x.grad.shape == x.shape
# Check that block parameters have gradients
for param in block.parameters():
if param.requires_grad:
assert param.grad is not None
def test_gradient_flow_prenorm(self):
"""Test gradient flow with prenorm."""
block = GammaSingleBlock(d_model=16, hidden_dim=32, prenorm=True)
x = torch.randn(2, 10, 16, requires_grad=True)
output, _ = block(x)
loss = output.sum()
loss.backward()
assert x.grad is not None
def test_gradient_flow_postnorm(self):
"""Test gradient flow with postnorm."""
block = GammaSingleBlock(d_model=16, hidden_dim=32, prenorm=False)
x = torch.randn(2, 10, 16, requires_grad=True)
output, _ = block(x)
loss = output.sum()
loss.backward()
assert x.grad is not None
class TestGammaSingleBlockIntegration:
"""Integration tests for GammaSingleBlock."""
def test_stacked_blocks(self):
"""Test stacking multiple blocks together."""
d_model, hidden_dim = 16, 32
num_blocks = 3
blocks = [
GammaSingleBlock(d_model=d_model, hidden_dim=hidden_dim)
for _ in range(num_blocks)
]
x = torch.randn(2, 10, d_model)
states = []
# Forward through blocks
for block in blocks:
x, state = block(x)
states.append(state)
assert x.shape == (2, 10, d_model)
assert len(states) == num_blocks
def test_device_transfer(self):
"""Test that block can be transferred between devices."""
block = GammaSingleBlock(d_model=16, hidden_dim=32)
# Test on CPU
x_cpu = torch.randn(2, 10, 16)
output_cpu, _ = block(x_cpu)
assert output_cpu.device.type == "cpu"
if torch.cuda.is_available():
# Transfer to GPU
block = block.cuda()
x_gpu = torch.randn(2, 10, 16).cuda()
output_gpu, _ = block(x_gpu)
assert output_gpu.device.type == "cuda"
# Transfer back to CPU
block = block.cpu()
output_cpu2, _ = block(x_cpu)
assert output_cpu2.device.type == "cpu"
def test_state_dict_save_load(self):
"""Test saving and loading state dict."""
block1 = GammaSingleBlock(d_model=16, hidden_dim=32)
block2 = GammaSingleBlock(d_model=16, hidden_dim=32)
# Save state dict from block1
state_dict = block1.state_dict()
# Load into block2
block2.load_state_dict(state_dict)
# They should produce same output
x = torch.randn(2, 10, 16)
with torch.no_grad():
out1, _ = block1(x)
out2, _ = block2(x)
assert torch.allclose(out1, out2, atol=1e-6)
def test_train_eval_mode_switching(self):
"""Test switching between train and eval modes."""
block = GammaSingleBlock(d_model=16, hidden_dim=32, dropout=0.5)
# Train mode
block.train()
assert block.training
# Eval mode
block.eval()
assert not block.training
# Train mode again
block.train()
assert block.training
def test_different_d_models_and_hidden_dims(self):
"""Test blocks with various dimensions."""
configs = [
(8, 16),
(16, 32),
(64, 128),
(256, 512),
]
for d_model, hidden_dim in configs:
block = GammaSingleBlock(d_model=d_model, hidden_dim=hidden_dim)
x = torch.randn(2, 10, d_model)
output, state = block(x)
assert output.shape == (2, 10, d_model)
assert state.shape == (2, hidden_dim)
if __name__ == "__main__":
pytest.main([__file__, "-v"])
|