pytorch - 💡(How to fix) Fix Combining `bfloat16` with complex scalars raises `NotImplementedError` [1 participants]

<python-input-3>:1: UserWarning: BComplex32 support is experimental and many operators don't support it yet. (Triggered internally at /Users/habbasi/Quansight/torch-build/git/pytorch/aten/src/ATen/EmptyTensor.cpp:58.) Traceback (most recent call last): File "<python-input-3>", line 1, in <module> a + b ^ NotImplementedError: "add_stub" not implemented for 'BComplex32'

Fix Plan

The fix involves avoiding the use of experimental BComplex32 data type and instead using a supported data type for complex numbers.

• Step 1: Convert the torch.bfloat16 tensor to a supported data type, such as torch.float32 or torch.complex64.
• Step 2: Perform the desired operation using the converted tensor.

Example Code

import torch

# Create a tensor with a supported data type
a = torch.zeros((1,), dtype=torch.float32)
b = complex(0, 0)  # Any complex scalar or array with ndim=0

# Convert the complex number to a torch tensor
b_tensor = torch.tensor([b.real, b.imag], dtype=torch.float32)

# Perform the desired operation
result = a + b_tensor[0]  # Add the real part of b to a

print(result)

Alternatively, you can use torch.complex64 data type to represent complex numbers:

import torch

# Create a tensor with complex data type
a = torch.zeros((1,), dtype=torch.complex64)
b = complex(0, 0)  # Any complex scalar or array with ndim=0

# Convert the complex number to a torch tensor
b_tensor = torch.tensor(b, dtype=torch.complex64)

# Perform the desired operation
result = a + b_tensor

print(result)

Verification

To verify that the fix worked, you can check the output of the example code and ensure that it produces the expected result without raising any errors.

Extra Tips

• Be cautious when using experimental data types, as they may not be fully supported and can lead to unexpected behavior.
• Always check the PyTorch documentation for the latest information on supported data types and operations.