"""RFM contracts: direction, loss, ODE/CFG, source balance and resume cursor.""" import unittest import torch from diffusion.rectified_flow import flow_matching_loss, sample_rfm from datasets.rfm_data import BalancedFlowSampler class RectifiedFlowTests(unittest.TestCase): def test_path_target_time_and_gradient(self): clean = torch.randn(3, 4, 2, 2) noise = torch.randn_like(clean) time = torch.tensor([0., .37, 1.]) offset = torch.tensor(.25, requires_grad=True) def oracle(x, embedding_time, condition, mask): self.assertTrue(torch.equal(x[0], noise[0])) self.assertTrue(torch.equal(x[-1], clean[-1])) torch.testing.assert_close(embedding_time, 999 * (1-time)) torch.testing.assert_close(x[1], .63*noise[1]+.37*clean[1]) return clean - noise + offset loss = flow_matching_loss(oracle, clean, noise, time, None, None).mean() torch.testing.assert_close(loss, torch.tensor(.25**2)) loss.backward() torch.testing.assert_close(offset.grad, torch.tensor(.5)) def test_cfg_and_solver_direction_all_four_channels(self): noise = torch.randn(2, 4, 2, 2) condition = torch.ones(2, 1, 1, 1) empty = torch.zeros(1, 1, 1, 1) mask = torch.ones(2, 1, dtype=torch.bool) def model(x, time, y, mask): return y.expand_as(x) for method in ('heun', 'dopri5'): result, report = sample_rfm(model, noise, condition, mask, empty, mask[:1], cfg_scale=2., method=method, steps=4) torch.testing.assert_close(result, noise + 2, atol=2e-5, rtol=2e-5) self.assertGreater(report['nfe'], 0) if method == 'heun': self.assertEqual(report['nfe'], 8) def test_unequal_sources_balance_coverage_and_resume(self): lengths = [50, 77] rank_batches = [list(BalancedFlowSampler(lengths, 8, 4, rank, 43, 0, 20)) for rank in range(4)] seen = [set(), set()] for step in range(20): for rank in range(4): batch = rank_batches[rank][step] self.assertEqual(sum(index < 50 for index in batch), 4) self.assertEqual(sum(index >= 50 for index in batch), 4) self.assertEqual(len(set(batch)), 8) seen[0].update(index for index in batch if index < 50) seen[1].update(index for index in batch if index >= 50) if step < 3: # Full global half-batches before either source's padded tail. combined = sum([rank_batches[rank][step] for rank in range(4)], []) self.assertEqual(len(set(combined)), 32) self.assertEqual(seen, [set(range(50)), set(range(50, 127))]) for rank in range(4): resumed = list(BalancedFlowSampler(lengths, 8, 4, rank, 43, 7, 20)) self.assertEqual(resumed, rank_batches[rank][7:]) if __name__ == '__main__': unittest.main()