Tensor_totype issues with device doesn't support double

1

When the device doesn’t support double (eg. intel iGPUs with torch-directml), this function will throw RuntimError:

def tensor_totype(t):
    dtype = torch.float if t.is_mps else torch.double
    return t.to(dtype=dtype)

Example:

>>> import torch
>>> import torch_directml
>>> device = torch_directml.device(1)  # UHD Graphics 770
>>> x = torch.ones(5, device=device)
>>> x
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "C:\Users\u7755376\AppData\Local\Programs\Python\Python312\Lib\site-packages\torch\_tensor.py", line 463, in __repr__
    return torch._tensor_str._str(self, tensor_contents=tensor_contents)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\u7755376\AppData\Local\Programs\Python\Python312\Lib\site-packages\torch\_tensor_str.py", line 698, in _str
    return _str_intern(self, tensor_contents=tensor_contents)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\u7755376\AppData\Local\Programs\Python\Python312\Lib\site-packages\torch\_tensor_str.py", line 618, in _str_intern
    tensor_str = _tensor_str(self, indent)
                 ^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\u7755376\AppData\Local\Programs\Python\Python312\Lib\site-packages\torch\_tensor_str.py", line 350, in _tensor_str
    formatter = _Formatter(get_summarized_data(self) if summarize else self)
                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\u7755376\AppData\Local\Programs\Python\Python312\Lib\site-packages\torch\_tensor_str.py", line 147, in __init__
    nonzero_finite_abs = tensor_totype(nonzero_finite_vals.abs())
                         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\u7755376\AppData\Local\Programs\Python\Python312\Lib\site-packages\torch\_tensor_str.py", line 119, in tensor_totype
    return t.to(dtype=dtype)
           ^^^^^^^^^^^^^^^^^
RuntimeError: The GPU device does not support Double (Float64) operations!

Would there be any way to fix this?

2

@chn-lee-yumi , It is due to that the iGPU does not support double. Does the fix here mean XPU should not raise RuntimeError?

3

By the way, you can check if the device supports FP64 through get_device_capability.

import torch
xpu_cap = torch.xpu.get_device_capability()

# Take Intel(R) Data Center GPU Max 1550 as an example, its device capability is as follows:
# {
#    'architecture': 13136561920,
#    'gpu_eu_count': 512, 
#    'gpu_subslice_count': 64,
#    'has_atomic64': True,
#    'has_bfloat16_conversions': True,
#    'has_fp16': True, 
#    'has_fp64': True, 
#    'has_subgroup_2d_block_io': True, 
#    'has_subgroup_matrix_multiply_accumulate': True, 
#    'has_subgroup_matrix_multiply_accumulate_tensor_float32': False,
#    'max_compute_units': 512,
#    'max_num_sub_groups': 64,
#    'max_work_group_size': 1024, 
#    'name': 'Intel(R) Data Center GPU Max 1550',
#    'platform_name': 'Intel(R) oneAPI Unified Runtime over Level-Zero',
#    'sub_group_sizes': [16, 32], 
#    'total_memory': 68719476736, 
#    'type': 'gpu',
#    'vendor': 'Intel(R) Corporation',
#    'version': '12.60.7'
# }

if xpu_cap["has_fp64"]:
  # Do something