What's the expected output from Device.allDevices ?

[xihui-wu]

I ran this on colab with CPU, GPT and TPU respectively. Here are their outputs:

CPU:
1 element
▿ 0 : Device(kind: .CPU, ordinal: 0, backend: .XLA)
- kind : TensorFlow.Device.Kind.CPU
- ordinal : 0
- backend : TensorFlow.Device.Backend.XLA

GPU:
1 element
▿ 0 : Device(kind: .GPU, ordinal: 0, backend: .XLA)
- kind : TensorFlow.Device.Kind.GPU
- ordinal : 0
- backend : TensorFlow.Device.Backend.XLA

TPU:
9 elements
▿ 0 : Device(kind: .CPU, ordinal: 0, backend: .XLA)
- kind : TensorFlow.Device.Kind.CPU
- ordinal : 0
- backend : TensorFlow.Device.Backend.XLA
▿ 1 : Device(kind: .TPU, ordinal: 0, backend: .XLA)
- kind : TensorFlow.Device.Kind.TPU
- ordinal : 0
- backend : TensorFlow.Device.Backend.XLA
▿ 2 : Device(kind: .TPU, ordinal: 1, backend: .XLA)
- kind : TensorFlow.Device.Kind.TPU
- ordinal : 1
- backend : TensorFlow.Device.Backend.XLA
...
▿ 8 : Device(kind: .TPU, ordinal: 7, backend: .XLA)
- kind : TensorFlow.Device.Kind.TPU
- ordinal : 7
- backend : TensorFlow.Device.Backend.XLA

My confusion is why GPU output doesn't contain CPU device but that of TPU does ? What are expected ?

[brettkoonce]

see also tensorflow/swift#524 #1077

[BradLarson]

This is due to the cause behind issue #1059, in that the X10 backend searches out and automatically exposes the best accelerator it finds on CPU / GPU systems. In doing so, it doesn't expose the other devices on the system, including the CPU on a CPU + GPU system, or additional GPUs or CPUs beyond the first.

This can be done manually at the command line by constructing the kind of environment variable string I list in #1059, but the proper mappings should be detected and taken care of for you internally. I believe that's the focus of that other issue, and it's being worked on in parallel to the visibility issues with the eager-mode devices.

[marcrasi]

@xihui-wu will verify and close after #1059 is fixed.

[xihui-wu]

Verified this issue is gone.