File size: 6,222 Bytes
ebad435 | 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 | import unittest
from unittest.mock import patch
import torch
import torch.nn.functional as F
from torch import nn
from predictor_training.atc_fusion import ATCFusion, MultiInputMLP, StateTransport
from predictor_training.single_block import SingleBlockPredictor
from wan.modules.model import rope_params
def small_freqs(head_dim: int = 8) -> torch.Tensor:
return torch.cat(
[
rope_params(32, head_dim - 4 * (head_dim // 6)),
rope_params(32, 2 * (head_dim // 6)),
rope_params(32, 2 * (head_dim // 6)),
],
dim=1,
)
def reference_attention(
q: torch.Tensor, k: torch.Tensor, v: torch.Tensor
) -> torch.Tensor:
scores = torch.einsum("blhd,bshd->bhls", q, k) / q.shape[-1] ** 0.5
probability = scores.float().softmax(dim=-1).to(v.dtype)
return torch.einsum("bhls,bshd->blhd", probability, v)
class DummyBlock(nn.Module):
num_heads = 1
def forward(self, value: torch.Tensor, **_kwargs) -> torch.Tensor:
return value
class ATCFusionTest(unittest.TestCase):
def setUp(self) -> None:
torch.manual_seed(0)
self.grid = torch.tensor([[2, 2, 2]], dtype=torch.long)
self.freqs = small_freqs()
self.current = torch.randn(1, 8, 8)
self.previous = torch.randn(1, 8, 8)
def test_multi_input_mlp_matches_explicit_concat(self) -> None:
module = MultiInputMLP(3, dim=8, hidden_dim=12)
values = [torch.randn(2, 5, 8) for _ in range(3)]
combined_weight = torch.cat(
[layer.weight for layer in module.inputs], dim=1
)
combined_bias = sum(layer.bias for layer in module.inputs)
reference_hidden = F.linear(
torch.cat(values, dim=-1), combined_weight, combined_bias
)
reference = module.out(module.activation(reference_hidden))
torch.testing.assert_close(module(*values), reference)
@patch(
"predictor_training.atc_fusion.attention",
side_effect=reference_attention,
)
def test_last_frame_scope_cannot_read_earlier_frames(self, _mock) -> None:
transport = StateTransport(
dim=8,
num_heads=1,
previous_scope="last_frame",
gradient_checkpointing=False,
diagnostic_queries=4,
).eval()
first, _ = transport(
self.current, self.previous, self.grid, self.freqs, current_start=8
)
perturbed_early = self.previous.clone()
perturbed_early[:, :4] += torch.randn_like(perturbed_early[:, :4])
second, _ = transport(
self.current,
perturbed_early,
self.grid,
self.freqs,
current_start=8,
)
torch.testing.assert_close(first, second, rtol=0, atol=0)
perturbed_last = self.previous.clone()
perturbed_last[:, 4:] += torch.randn_like(perturbed_last[:, 4:])
third, _ = transport(
self.current,
perturbed_last,
self.grid,
self.freqs,
current_start=8,
)
self.assertFalse(torch.equal(first, third))
@patch(
"predictor_training.atc_fusion.attention",
side_effect=reference_attention,
)
def test_zero_initialized_correction_bypass(self, _mock) -> None:
fusion = ATCFusion(
dim=8,
num_heads=1,
freq_dim=8,
mlp_hidden_dim=16,
gate_hidden_dim=4,
previous_scope="chunk",
gradient_checkpointing=False,
).eval()
condition = torch.randn_like(self.current)
backbone, gated_correction = fusion(
self.current,
torch.randn_like(self.current),
self.previous,
condition,
torch.tensor([62.5]),
self.grid,
self.freqs,
current_start=8,
)
self.assertEqual(backbone.shape, self.current.shape)
torch.testing.assert_close(
gated_correction,
torch.zeros_like(gated_correction),
rtol=0,
atol=0,
)
self.assertAlmostEqual(float(fusion.last_gate.mean()), 0.3, delta=0.01)
def test_atc_token_gate_is_a_stage1_parameter(self) -> None:
predictor = SingleBlockPredictor(
DummyBlock(),
dim=8,
gradient_checkpointing=False,
input_variant="atc",
atc_freq_dim=8,
atc_mlp_hidden_dim=16,
atc_gate_hidden_dim=4,
)
token_gate_ids = {
id(parameter)
for parameter in predictor.fusion.token_gate.parameters()
}
stage1_ids = {
id(parameter) for parameter in predictor.stage1_input_parameters()
}
self.assertTrue(token_gate_ids)
self.assertTrue(token_gate_ids.issubset(stage1_ids))
self.assertEqual(predictor.gate_parameters(), [])
@patch(
"predictor_training.atc_fusion.attention",
side_effect=reference_attention,
)
def test_zero_initialized_predictor_returns_anchor(self, _mock) -> None:
predictor = SingleBlockPredictor(
DummyBlock(),
dim=8,
gradient_checkpointing=False,
input_variant="atc",
atc_freq_dim=8,
atc_mlp_hidden_dim=16,
atc_gate_hidden_dim=4,
).eval()
anchor = torch.randn_like(self.current)
empty_history = torch.empty(1, 0, 1, 8)
prediction = predictor(
current_tokens=self.current,
anchor_hidden=anchor,
previous_hidden=self.previous,
timestep_modulation=torch.randn(1, 2, 6, 8),
grid_sizes=self.grid,
freqs=self.freqs,
history_k=empty_history,
history_v=empty_history,
cross_k=torch.empty(1, 0, 1, 8),
cross_v=torch.empty(1, 0, 1, 8),
current_start=8,
condition_tokens=torch.randn_like(self.current),
anchor_distance=torch.tensor([62.5]),
)
torch.testing.assert_close(prediction, anchor, rtol=0, atol=0)
if __name__ == "__main__":
unittest.main()
|