[MPS] Tracking issue for ModuleInfo failures when enabling testing for torch.float16 #119108
Description
🐛 Describe the bug
Issue accompanying #119039. The following tests fail when I enable torch.float16 ModuleInfo testing for MPS
Not specific to MPS (also fail on CPU and CUDA)
-
test_check_inplace_nn_CELU_mps_float16 -
test_check_inplace_nn_ELU_mps_float16
Errors with "error: input types 'tensor' and 'tensor<...f16>' are not broadcast compatible"
-
test_forward_nn_AvgPool2d_mps_float16 -
test_forward_nn_BCELoss_mps_float16 -
test_if_train_and_eval_modes_differ_nn_AvgPool2d_mps_float16 -
test_if_train_and_eval_modes_differ_nn_BCELoss_mps_float16 -
test_memory_format_nn_AvgPool2d_mps_float16 -
test_non_contiguous_tensors_nn_AvgPool2d_mps_float16 -
test_non_contiguous_tensors_nn_BCELoss_mps_float16 -
test_pickle_nn_AvgPool2d_mps_float16 -
test_pickle_nn_BCELoss_mps_float16 -
test_non_contiguous_tensors_nn_MSELoss_mps_float16 -
test_non_contiguous_tensors_nn_SmoothL1Loss_mps_float16
Errors with "MPSNDArrayConvolutionA14.mm:3976: failed assertion `destination datatype must be fp32'"
-
test_memory_format_nn_Conv1d_mps_float16 -
test_memory_format_nn_Conv2d_mps_float16 -
test_memory_format_nn_ConvTranspose1d_mps_float16 -
test_memory_format_nn_ConvTranspose2d_mps_float16 -
test_memory_format_nn_LazyConv1d_mps_float16 -
test_memory_format_nn_LazyConv2d_mps_float16 -
test_memory_format_nn_LazyConvTranspose1d_mps_float16 -
test_memory_format_nn_LazyConvTranspose2d_mps_float16 -
test_non_contiguous_tensors_nn_Conv1d_mps_float16 -
test_non_contiguous_tensors_nn_Conv2d_mps_float16 -
test_non_contiguous_tensors_nn_ConvTranspose1d_mps_float16 -
test_non_contiguous_tensors_nn_ConvTranspose2d_mps_float16 -
test_non_contiguous_tensors_nn_LazyConv1d_mps_float16 -
test_non_contiguous_tensors_nn_LazyConv2d_mps_float16 -
test_non_contiguous_tensors_nn_LazyConvTranspose1d_mps_float16 -
test_non_contiguous_tensors_nn_LazyConvTranspose2d_mps_float16
Tolerance errors
-
test_forward_nn_Bilinear_mps_float16 -
test_forward_nn_BCEWithLogitsLoss_mps_float16 -
test_forward_nn_GELU_mps_float16 -
test_forward_nn_LogSigmoid_mps_float16 -
test_forward_nn_MSELoss_mps_float16 -
test_forward_nn_NLLLoss_mps_float16 -
test_forward_nn_SoftMarginLoss_mps_float16 -
test_forward_nn_Softmax2d_mps_float16
inf nan error
-
test_forward_nn_LogSoftmax_mps_float16
Incorrect output dtype? values for attribute 'dtype' do not match: torch.float16 != torch.float32.
-
test_forward_nn_HuberLoss_mps_float16
Versions
main