(self)
| 621 | self.assertLess((y_q - y_hat).abs().max(), 1e-3) |
| 622 | |
| 623 | def test_non_multiples(self): |
| 624 | w = mx.random.normal(shape=(33, 256)) |
| 625 | w_q, scales, biases = mx.quantize(w) |
| 626 | w_hat = mx.dequantize(w_q, scales, biases) |
| 627 | |
| 628 | # Test qmv |
| 629 | x = mx.random.normal(shape=(1, 256)) |
| 630 | y_q = mx.quantized_matmul(x, w_q, scales, biases, transpose=True) |
| 631 | y_hat = x @ w_hat.T |
| 632 | self.assertLess((y_q - y_hat).abs().max(), 1e-3) |
| 633 | |
| 634 | # Test qmm_t |
| 635 | x = mx.random.normal(shape=(10, 256)) |
| 636 | y_q = mx.quantized_matmul(x, w_q, scales, biases, transpose=True) |
| 637 | y_hat = x @ w_hat.T |
| 638 | self.assertEqual(y_q.shape, y_hat.shape) |
| 639 | self.assertLess((y_q - y_hat).abs().max(), 1e-3) |
| 640 | |
| 641 | # Test qvm |
| 642 | x = mx.random.normal(shape=(1, 33)) |
| 643 | y_q = mx.quantized_matmul(x, w_q, scales, biases, transpose=False) |
| 644 | y_hat = x @ w_hat |
| 645 | self.assertEqual(y_q.shape, y_hat.shape) |
| 646 | self.assertLess((y_q - y_hat).abs().max(), 1e-3) |
| 647 | |
| 648 | # Test qmm |
| 649 | x = mx.random.normal(shape=(10, 33)) |
| 650 | y_q = mx.quantized_matmul(x, w_q, scales, biases, transpose=False) |
| 651 | y_hat = x @ w_hat |
| 652 | self.assertEqual(y_q.shape, y_hat.shape) |
| 653 | self.assertLess((y_q - y_hat).abs().max(), 1e-3) |
| 654 | |
| 655 | # Smaller than 8 |
| 656 | w = mx.random.normal(shape=(3, 256)) |
| 657 | w_q, scales, biases = mx.quantize(w) |
| 658 | w_hat = mx.dequantize(w_q, scales, biases) |
| 659 | |
| 660 | # Test qmv |
| 661 | x = mx.random.normal(shape=(1, 256)) |
| 662 | y_q = mx.quantized_matmul(x, w_q, scales, biases, transpose=True) |
| 663 | y_hat = x @ w_hat.T |
| 664 | self.assertLess((y_q - y_hat).abs().max(), 1e-3) |
| 665 | |
| 666 | # Test qmm_t |
| 667 | x = mx.random.normal(shape=(10, 256)) |
| 668 | y_q = mx.quantized_matmul(x, w_q, scales, biases, transpose=True) |
| 669 | y_hat = x @ w_hat.T |
| 670 | self.assertEqual(y_q.shape, y_hat.shape) |
| 671 | self.assertLess((y_q - y_hat).abs().max(), 1e-3) |
| 672 | |
| 673 | # Test qvm |
| 674 | x = mx.random.normal(shape=(1, 3)) |
| 675 | y_q = mx.quantized_matmul(x, w_q, scales, biases, transpose=False) |
| 676 | y_hat = x @ w_hat |
| 677 | self.assertEqual(y_q.shape, y_hat.shape) |
| 678 | self.assertLess((y_q - y_hat).abs().max(), 1e-3) |
| 679 | |
| 680 | # Test qmm |
nothing calls this directly
no test coverage detected