How to use IPU accelerator in paperspace

1

Hi, Hope you are doing great.

I recently discovered that Paperspace offers some IPU accelerators for free. I decided to use these feature, but It seems that pytorch lightning cannot detect IPUs. Because whenever I run the code with accelerator='auto' it print the following log:

GPU available: False, used: False
TPU available: False, using: 0 TPU cores
IPU available: False, using: 0 IPUs
HPU available: False, using: 0 HPUs

If I set the device with accelerator=ipu. it produces the following error:

Traceback (most recent call last):
  File "train.py", line 154, in <module>
    main()
  File "train.py", line 140, in main
    trainer = Trainer(
  File "/opt/pytorch/lib/python3.8/site-packages/lightning/pytorch/utilities/argparse.py", line 70, in insert_env_defaults
    return fn(self, **kwargs)
  File "/opt/pytorch/lib/python3.8/site-packages/lightning/pytorch/trainer/trainer.py", line 402, in __init__
    self._accelerator_connector = _AcceleratorConnector(
  File "/opt/pytorch/lib/python3.8/site-packages/lightning/pytorch/trainer/connectors/accelerator_connector.py", line 139, in __init__
    self._check_config_and_set_final_flags(
  File "/opt/pytorch/lib/python3.8/site-packages/lightning/pytorch/trainer/connectors/accelerator_connector.py", line 220, in _check_config_and_set_final_flags
    raise ValueError(
ValueError: You selected an invalid accelerator name: `accelerator='ipu'`. Available names are: auto, cpu, cuda, mps, tpu.

Any ideas about how to resolve this issue?

2

The graphcore implementation in Lightning is no longer maintained. It moved to a separate package:

But it existed in Lightning until version 1.9.5:
https://lightning.ai/docs/pytorch/1.9.5/accelerators/ipu.html?highlight=ipu

So with a bit of luck, you will be able to run it with pytorch-lightning==1.9.5. Good luck :))