已合并
【bugfix】transfer_to_npu add has_triton patch for compile_fx.py #37371
kkjocker创建于 6月1日
【bugfix】transfer_to_npu add has_triton patch for compile_fx.py #37371
已合并
共 2 个文件变更+3-0
| @@ -31,6 +31,7 @@ class TestActivations(TestCase): | |||
| 31 | res.backward(torch.ones_like(res)) | 31 | res.backward(torch.ones_like(res)) |
| 32 | return res.detach().cpu(), input1.grad.cpu() | 32 | return res.detach().cpu(), input1.grad.cpu() |
| 33 | 33 | ||
| 34 | + | ||
| 34 | def test_mish(self): | 35 | def test_mish(self): |
| 35 | dtype_list = [np.float16, np.float32] | 36 | dtype_list = [np.float16, np.float32] |
| 36 | format_list = [-1, 0, 2] | 37 | format_list = [-1, 0, 2] |
| @@ -81,6 +82,7 @@ class TestActivations(TestCase): | |||
| 81 | output.backward() | 82 | output.backward() |
| 82 | return res.detach().cpu(), input1.grad.cpu() | 83 | return res.detach().cpu(), input1.grad.cpu() |
| 83 | 84 | ||
| 85 | + | ||
| 84 | def test_silu(self): | 86 | def test_silu(self): |
| 85 | dtype_list = [np.float32, np.float16] | 87 | dtype_list = [np.float32, np.float16] |
| 86 | format_list = [-1, 0, 2] | 88 | format_list = [-1, 0, 2] |
| @@ -513,6 +513,7 @@ def _init(): | |||
| 513 | setattr(torch.utils._triton, 'has_triton', _patch_has_triton) | 513 | setattr(torch.utils._triton, 'has_triton', _patch_has_triton) |
| 514 | setattr(torch._dynamo.utils, 'has_triton', _patch_has_triton) | 514 | setattr(torch._dynamo.utils, 'has_triton', _patch_has_triton) |
| 515 | setattr(torch._inductor.runtime.autotune_cache, 'has_triton', _patch_has_triton) | 515 | setattr(torch._inductor.runtime.autotune_cache, 'has_triton', _patch_has_triton) |
| 516 | + setattr(torch._inductor.compile_fx, 'has_triton', _patch_has_triton) | ||
| 516 | 517 | ||
| 517 | setattr(torch._utils, '_get_available_device_type', _patch_get_available_device_type) | 518 | setattr(torch._utils, '_get_available_device_type', _patch_get_available_device_type) |
| 518 | setattr(torch.distributed.checkpoint.filesystem._OverlappingCpuLoader, '__init__', | 519 | setattr(torch.distributed.checkpoint.filesystem._OverlappingCpuLoader, '__init__', |