已合并
test:This PR adds missing test cases for torch._utils._unflatten_dense_tensors, as there are currently no community-provided tests for this function. #42048
创建于 7月18日
test:This PR adds missing test cases for torch._utils._unflatten_dense_tensors, as there are currently no community-provided tests for this function. #42048
已合并
从已删除 :test_unflatten_dense_tensors_v2.7.1合入到Ascend/pytorchv2.7.1
共 1 个文件变更+132-0
| @@ -0,0 +1,132 @@ | |||
| 1 | +""" | ||
| 2 | +Add validation cases for torch._utils APIs on NPU: | ||
| 3 | +1. PyTorch community lacks direct Python unit tests for torch._utils._unflatten_dense_tensors. | ||
| 4 | +2. This file validates torch._utils._unflatten_dense_tensors (extendable). | ||
| 5 | +""" | ||
| 6 | + | ||
| 7 | +import torch | ||
| 8 | +from torch.testing._internal.common_utils import TestCase, run_tests | ||
| 9 | + | ||
| 10 | + | ||
| 11 | +class TestUnflattenDenseTensors(TestCase): | ||
| 12 | + """Test cases for torch._utils._unflatten_dense_tensors.""" | ||
| 13 | + | ||
| 14 | + def setUp(self): | ||
| 15 | + super().setUp() | ||
| 16 | + acc = torch.accelerator.current_accelerator() | ||
| 17 | + self.device = acc.type if acc else "cpu" | ||
| 18 | + | ||
| 19 | + def _to_device(self, t): | ||
| 20 | + return t.to(self.device) | ||
| 21 | + | ||
| 22 | + def test_round_trip_basic(self): | ||
| 23 | + # Round-trip: flatten then unflatten should recover original tensors | ||
| 24 | + t1 = self._to_device(torch.ones(4, 4)) | ||
| 25 | + t2 = self._to_device(torch.zeros(2, 3)) | ||
| 26 | + t3 = self._to_device(torch.randn(5)) | ||
| 27 | + tensors = [t1, t2, t3] | ||
| 28 | + | ||
| 29 | + flat = torch._utils._flatten_dense_tensors(tensors) | ||
| 30 | + result = torch._utils._unflatten_dense_tensors(flat, tensors) | ||
| 31 | + | ||
| 32 | + self.assertEqual(len(result), len(tensors)) | ||
| 33 | + for r, t in zip(result, tensors): | ||
| 34 | + self.assertEqual(r.shape, t.shape) | ||
| 35 | + self.assertEqual(r, t) | ||
| 36 | + | ||
| 37 | + def test_single_tensor(self): | ||
| 38 | + t = self._to_device(torch.randn(3, 5, 2)) | ||
| 39 | + flat = torch._utils._flatten_dense_tensors([t]) | ||
| 40 | + result = torch._utils._unflatten_dense_tensors(flat, [t]) | ||
| 41 | + | ||
| 42 | + self.assertEqual(len(result), 1) | ||
| 43 | + self.assertEqual(result[0].shape, t.shape) | ||
| 44 | + self.assertEqual(result[0], t) | ||
| 45 | + | ||
| 46 | + def test_multiple_tensors_different_sizes(self): | ||
| 47 | + sizes = [(1,), (2, 3), (4, 5, 6), (7, 8)] | ||
| 48 | + tensors = [self._to_device(torch.randn(*s)) for s in sizes] | ||
| 49 | + | ||
| 50 | + flat = torch._utils._flatten_dense_tensors(tensors) | ||
| 51 | + result = torch._utils._unflatten_dense_tensors(flat, tensors) | ||
| 52 | + | ||
| 53 | + self.assertEqual(len(result), len(tensors)) | ||
| 54 | + for i, (r, t) in enumerate(zip(result, tensors)): | ||
| 55 | + self.assertEqual(r.shape, t.shape, f"tensor {i} shape mismatch") | ||
| 56 | + self.assertEqual(r, t, f"tensor {i} value mismatch") | ||
| 57 | + | ||
| 58 | + def test_empty_tensor_in_list(self): | ||
| 59 | + t1 = self._to_device(torch.ones(3, 2)) | ||
| 60 | + t2 = self._to_device(torch.tensor([])) | ||
| 61 | + t3 = self._to_device(torch.randn(4)) | ||
| 62 | + tensors = [t1, t2, t3] | ||
| 63 | + | ||
| 64 | + flat = torch._utils._flatten_dense_tensors(tensors) | ||
| 65 | + result = torch._utils._unflatten_dense_tensors(flat, tensors) | ||
| 66 | + | ||
| 67 | + self.assertEqual(len(result), 3) | ||
| 68 | + self.assertEqual(result[0].numel(), 6) | ||
| 69 | + self.assertEqual(result[1].numel(), 0) | ||
| 70 | + self.assertEqual(result[2].numel(), 4) | ||
| 71 | + self.assertEqual(result[0], t1) | ||
| 72 | + # Verify empty tensor preserves shape, dtype and device | ||
| 73 | + self.assertEqual(result[1].shape, t2.shape) | ||
| 74 | + self.assertEqual(result[1].dtype, t2.dtype) | ||
| 75 | + self.assertEqual(result[1], t2) | ||
| 76 | + self.assertEqual(result[2], t3) | ||
| 77 | + | ||
| 78 | + def test_all_empty_tensors(self): | ||
| 79 | + tensors = [ | ||
| 80 | + self._to_device(torch.tensor([])), | ||
| 81 | + self._to_device(torch.tensor([])), | ||
| 82 | + ] | ||
| 83 | + | ||
| 84 | + flat = torch._utils._flatten_dense_tensors(tensors) | ||
| 85 | + result = torch._utils._unflatten_dense_tensors(flat, tensors) | ||
| 86 | + | ||
| 87 | + self.assertEqual(len(result), 2) | ||
| 88 | + for r, t in zip(result, tensors): | ||
| 89 | + self.assertEqual(r.shape, t.shape) | ||
| 90 | + self.assertEqual(r, t) | ||
| 91 | + | ||
| 92 | + def test_different_dtypes(self): | ||
| 93 | + for dtype in [torch.float32, torch.float16, torch.int32]: | ||
| 94 | + with self.subTest(dtype=dtype): | ||
| 95 | + t1 = self._to_device(torch.ones(2, 3, dtype=dtype)) | ||
| 96 | + t2 = self._to_device(torch.zeros(4, dtype=dtype)) | ||
| 97 | + | ||
| 98 | + flat = torch._utils._flatten_dense_tensors([t1, t2]) | ||
| 99 | + result = torch._utils._unflatten_dense_tensors(flat, [t1, t2]) | ||
| 100 | + | ||
| 101 | + self.assertEqual(result[0].dtype, dtype) | ||
| 102 | + self.assertEqual(result[1].dtype, dtype) | ||
| 103 | + self.assertEqual(result[0], t1) | ||
| 104 | + self.assertEqual(result[1], t2) | ||
| 105 | + | ||
| 106 | + def test_large_num_tensors(self): | ||
| 107 | + n = 50 | ||
| 108 | + tensors = [self._to_device(torch.randn(i + 1)) for i in range(n)] | ||
| 109 | + | ||
| 110 | + flat = torch._utils._flatten_dense_tensors(tensors) | ||
| 111 | + result = torch._utils._unflatten_dense_tensors(flat, tensors) | ||
| 112 | + | ||
| 113 | + self.assertEqual(len(result), n) | ||
| 114 | + for r, t in zip(result, tensors): | ||
| 115 | + self.assertEqual(r.shape, t.shape) | ||
| 116 | + self.assertEqual(r, t) | ||
| 117 | + | ||
| 118 | + def test_tuple_input(self): | ||
| 119 | + t1 = self._to_device(torch.ones(3, 3)) | ||
| 120 | + t2 = self._to_device(torch.zeros(2)) | ||
| 121 | + tensors = (t1, t2) # tuple, not list | ||
| 122 | + | ||
| 123 | + flat = torch._utils._flatten_dense_tensors(tensors) | ||
| 124 | + result = torch._utils._unflatten_dense_tensors(flat, tensors) | ||
| 125 | + | ||
| 126 | + self.assertEqual(len(result), 2) | ||
| 127 | + for r, t in zip(result, tensors): | ||
| 128 | + self.assertEqual(r, t) | ||
| 129 | + | ||
| 130 | + | ||
| 131 | +if __name__ == "__main__": | ||
| 132 | + run_tests() | ||