Testing a normal tensor
(self)
| 48 | self.assertEqual(flatbuffer_tensor.dim_order, dim_order) |
| 49 | |
| 50 | def test_normal_tensor_conversion(self) -> None: |
| 51 | """Testing a normal tensor""" |
| 52 | |
| 53 | normal_tensor = torch.randn(2, 2, 3) |
| 54 | flatbuffer_tensor = make_tensor_value( |
| 55 | 1, 0, TensorSpec.from_tensor(normal_tensor) |
| 56 | ) |
| 57 | self.compare_tensors(normal_tensor, flatbuffer_tensor) |
| 58 | |
| 59 | # Test zero size tensor |
| 60 | normal_tensor = torch.randn(2, 2, 0) |
| 61 | flatbuffer_tensor = make_tensor_value( |
| 62 | 1, 0, TensorSpec.from_tensor(normal_tensor) |
| 63 | ) |
| 64 | self.compare_tensors(normal_tensor, flatbuffer_tensor) |
| 65 | |
| 66 | # Test zero size tensor |
| 67 | normal_tensor = torch.randn(2, 0, 3) |
| 68 | flatbuffer_tensor = make_tensor_value( |
| 69 | 1, 0, TensorSpec.from_tensor(normal_tensor) |
| 70 | ) |
| 71 | self.compare_tensors(normal_tensor, flatbuffer_tensor) |
| 72 | |
| 73 | # Test zero size tensor |
| 74 | normal_tensor = torch.randn(0, 2, 3) |
| 75 | flatbuffer_tensor = make_tensor_value( |
| 76 | 1, 0, TensorSpec.from_tensor(normal_tensor) |
| 77 | ) |
| 78 | self.compare_tensors(normal_tensor, flatbuffer_tensor) |
| 79 | |
| 80 | # Compare dim order |
| 81 | normal_tensor = torch.rand((2, 2, 3, 4)) |
| 82 | flatbuffer_tensor = make_tensor_value( |
| 83 | 1, 0, TensorSpec.from_tensor(normal_tensor) |
| 84 | ) |
| 85 | self.compare_tensors(normal_tensor, flatbuffer_tensor, dim_order=[0, 1, 2, 3]) |
| 86 | # cannot compare torch.memory_format = torch.channels_last because make_tensor_value |
| 87 | # infers strides from sizes assuming tensor dimensions are laid out in memory |
| 88 | # in the same order as indicated by dimension order of sizes array. |
| 89 | # e.g. for sizes = (2, 3, 4, 5), it assumes dimension order is (0, 1, 2, 3) and |
| 90 | # thus strides = (3*4*5, 4*5, 5, 1) |
| 91 | # whereas strides for torch.memory_format = torch.channels_last is |
| 92 | # (3*4*5, 1, 5*3, 3)) |
| 93 | |
| 94 | def test_fp8_tensor_conversion(self) -> None: |
| 95 | for dtype in (torch.float8_e5m2, torch.float8_e4m3fn): |
nothing calls this directly
no test coverage detected