File size: 9,346 Bytes
3f42614 | 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 | """
GLADIUS Plug — Test Suite
Verifies membrane projection, kernel freeze, gradient flow, and save/load.
Runs WITHOUT a real GLADIUS checkpoint by mocking the kernel.
"""
import torch
import torch.nn as nn
import tempfile
import os
import sys
from pathlib import Path
# Import directly from the module file to avoid package __init__.py conflicts
plug_dir = str(Path(__file__).parent)
sys.path.insert(0, str(Path(__file__).parent.parent))
from plug.plug import Membrane
def test_membrane_shape():
"""Membrane projects external_dim → gladius_dim with correct shapes."""
membrane = Membrane(external_dim=768, gladius_dim=640)
x = torch.randn(2, 128, 768)
out = membrane(x)
assert out.shape == (2, 128, 640), f"Expected (2, 128, 640), got {out.shape}"
print("[PASS] test_membrane_shape: (2, 128, 768) -> (2, 128, 640)")
def test_membrane_different_dims():
"""Membrane works with various external dimensions."""
for ext_dim, gladius_dim in [(768, 640), (2048, 640), (256, 640), (4096, 256)]:
membrane = Membrane(external_dim=ext_dim, gladius_dim=gladius_dim)
x = torch.randn(1, 32, ext_dim)
out = membrane(x)
assert out.shape == (1, 32, gladius_dim), \
f"ext={ext_dim}->gladius={gladius_dim}: expected (1,32,{gladius_dim}), got {out.shape}"
print("[PASS] test_membrane_different_dims: all dimension pairs work")
def test_membrane_gradients():
"""Membrane parameters have gradients after backward pass."""
membrane = Membrane(external_dim=768, gladius_dim=640)
x = torch.randn(2, 128, 768)
out = membrane(x)
loss = out.sum()
loss.backward()
for name, param in membrane.named_parameters():
assert param.grad is not None, f"Membrane param '{name}' has no gradient"
assert param.grad.abs().sum() > 0, f"Membrane param '{name}' has zero gradient"
print("[PASS] test_membrane_gradients: all membrane params receive gradients")
def test_freeze_kernel_simulation():
"""
Simulate the freeze behavior: kernel params frozen, membrane params trainable.
Uses a simple nn.Module as kernel stand-in.
"""
# Mock kernel: 3-layer transformer
class MockKernel(nn.Module):
def __init__(self):
super().__init__()
self.layers = nn.ModuleList([nn.Linear(640, 640) for _ in range(3)])
self.final_norm = nn.LayerNorm(640)
def forward(self, x):
for layer in self.layers:
x = layer(x)
return self.final_norm(x)
kernel = MockKernel()
membrane = Membrane(external_dim=768, gladius_dim=640)
# Freeze kernel
for p in kernel.parameters():
p.requires_grad = False
# Forward through membrane + kernel
x = torch.randn(2, 128, 768)
projected = membrane(x)
enriched = kernel(projected)
loss = enriched.sum()
loss.backward()
# Kernel params: frozen (no grad)
for name, param in kernel.named_parameters():
assert param.grad is None, f"Kernel param '{name}' should have no gradient (frozen)"
# Membrane params: trainable (has grad)
for name, param in membrane.named_parameters():
assert param.grad is not None, f"Membrane param '{name}' should have gradient"
print("[PASS] test_freeze_kernel_simulation: kernel frozen, membrane learns")
def test_membrane_save_load():
"""Membrane state roundtrips through save/load correctly."""
membrane = Membrane(external_dim=768, gladius_dim=640)
# Run a forward + backward to give params non-trivial values
x = torch.randn(2, 64, 768)
out = membrane(x)
loss = out.sum()
loss.backward()
# Manually update weights to make them non-default
with torch.no_grad():
for p in membrane.parameters():
p.add_(torch.randn_like(p) * 0.1)
# Save
with tempfile.NamedTemporaryFile(suffix='.pt', delete=False) as f:
save_path = f.name
try:
torch.save({
'membrane_state_dict': membrane.state_dict(),
'external_dim': membrane.external_dim,
'gladius_dim': membrane.gladius_dim,
}, save_path)
# Load into new membrane
membrane2 = Membrane(external_dim=768, gladius_dim=640)
data = torch.load(save_path, map_location='cpu')
membrane2.load_state_dict(data['membrane_state_dict'])
# Verify weights match
for (n1, p1), (n2, p2) in zip(membrane.named_parameters(), membrane2.named_parameters()):
assert torch.equal(p1, p2), f"Mismatch on '{n1}' after load"
# Verify outputs match
x_test = torch.randn(1, 32, 768)
membrane.eval()
membrane2.eval()
with torch.no_grad():
out1 = membrane(x_test)
out2 = membrane2(x_test)
assert torch.allclose(out1, out2, atol=1e-6), "Output mismatch after load"
print("[PASS] test_membrane_save_load: roundtrip preserves weights and outputs")
finally:
os.unlink(save_path)
def test_membrane_param_count():
"""Verify parameter count math: Linear(ext, gladius) + LayerNorm(gladius)."""
membrane = Membrane(external_dim=768, gladius_dim=640)
expected_linear = 768 * 640 + 640 # weight + bias
expected_ln = 640 + 640 # weight + bias
expected_total = expected_linear + expected_ln
actual = sum(p.numel() for p in membrane.parameters())
assert actual == expected_total, \
f"Expected {expected_total:,} params, got {actual:,}"
print(f"[PASS] test_membrane_param_count: {actual:,} params (Linear: {expected_linear:,} + LN: {expected_ln:,})")
def test_membrane_layernorm_output():
"""LayerNorm in membrane normalizes the projected output."""
membrane = Membrane(external_dim=768, gladius_dim=640)
x = torch.randn(2, 128, 768) * 100 # Large scale input
out = membrane(x)
# After LayerNorm, last dim should be approximately zero-mean, unit-var
mean = out.mean(dim=-1)
var = out.var(dim=-1, unbiased=False)
assert mean.abs().max() < 0.5, f"Post-LN mean too high: {mean.abs().max():.4f}"
assert (var - 1.0).abs().max() < 0.5, f"Post-LN var far from 1: {var.mean():.4f}"
print("[PASS] test_membrane_layernorm_output: output is approximately normalized")
def test_plug_forward_simulation():
"""
Full Plug-style forward: membrane → frozen layers → output.
Simulates what GladiusPlug.forward() does without needing a checkpoint.
"""
class MockGladiusStack(nn.Module):
"""Simulates GLADIUS layer stack + final norm."""
def __init__(self, dim=640, num_layers=14):
super().__init__()
self.layers = nn.ModuleList([
nn.Sequential(nn.LayerNorm(dim), nn.Linear(dim, dim))
for _ in range(num_layers)
])
self.final_norm = nn.LayerNorm(dim)
def forward(self, x):
for layer in self.layers:
x = x + layer(x) # residual
return self.final_norm(x)
# Build Plug-like setup
membrane = Membrane(external_dim=768, gladius_dim=640)
kernel = MockGladiusStack(dim=640, num_layers=14)
# Freeze kernel
for p in kernel.parameters():
p.requires_grad = False
# Forward
external_hidden = torch.randn(2, 128, 768)
projected = membrane(external_hidden)
enriched = kernel(projected)
assert enriched.shape == (2, 128, 640)
# Backward — only membrane should get gradients
loss = enriched.sum()
loss.backward()
membrane_grads = sum(1 for p in membrane.parameters() if p.grad is not None)
kernel_grads = sum(1 for p in kernel.parameters() if p.grad is not None)
assert membrane_grads > 0, "Membrane should have gradients"
assert kernel_grads == 0, "Kernel should have zero gradients (frozen)"
print("[PASS] test_plug_forward_simulation: full pipeline, correct gradient flow")
def test_membrane_batch_sizes():
"""Membrane handles various batch sizes including batch=1."""
membrane = Membrane(external_dim=768, gladius_dim=640)
for batch in [1, 2, 4, 16]:
x = torch.randn(batch, 64, 768)
out = membrane(x)
assert out.shape == (batch, 64, 640), f"Batch {batch}: wrong shape {out.shape}"
print("[PASS] test_membrane_batch_sizes: handles batch 1, 2, 4, 16")
# ─── Run all tests ───
if __name__ == '__main__':
tests = [
test_membrane_shape,
test_membrane_different_dims,
test_membrane_gradients,
test_freeze_kernel_simulation,
test_membrane_save_load,
test_membrane_param_count,
test_membrane_layernorm_output,
test_plug_forward_simulation,
test_membrane_batch_sizes,
]
passed = 0
failed = 0
for test in tests:
try:
test()
passed += 1
except Exception as e:
print(f"[FAIL] {test.__name__}: {e}")
failed += 1
print(f"\n{'='*50}")
print(f" {passed}/{passed+failed} tests passed")
if failed > 0:
print(f" {failed} FAILED")
else:
print(f" ALL PASS ✓")
print(f"{'='*50}")
|