Support Intel Arc

[johnnynunez]

Hi,
Is it possible to support Intel Arc as a device, as it is done with apple ARM(MPS), nvidia GPU(GPU) or tpus, so I don't have to change the code using the intel extension?

import torch
import torch.nn as nn
# Import intel_extension_for_pytorch
import intel_extension_for_pytorch as ipex

# oneDNN graph fusion is enabled by default, uncomment the line below to disable it explicitly
# ipex.enable_onednn_fusion(False)

class Model(nn.Module):
    def __init__(self):
        super(Model, self).__init__()
        self.linear = nn.Linear(4, 5)

    def forward(self, input):
        return self.linear(input)

input = torch.randn(2, 4)
model = Model()
model.eval()
# Invoke optimize function against the model object
model = ipex.optimize(model, dtype=torch.float32)
model = torch.jit.trace(model, torch.randn(2, 4))
model = torch.jit.freeze(model)
res = model(input)

Intel Extension

cc @Borda @akihironitta @justusschock

[awaelchli]

This looks great. @johnnynunez do you know if calling ipex.optimize jointly on the optimizer and model is required for training? Or would it be ok to call them separately?

model = ipex.optimize(model, dtype=torch.float32)
optimizer = ipex.optimize(optimizer, dtype=torch.float32)

instead of

model, optimizer = ipex.optimize(model, optimizer, dtype=torch.float32)

?

One could investigate whether creating a IntelCPUAccelerator is enough to support this in Lightning Trainer.

[justusschock]

@awaelchli since the optimize function does not have a default argument for model, I'd assume that they need to be called jointly. What is yet unclear to me is: if

• are multiple optimizers supported?
• how common arc devices are in general?

[johnnynunez]

In the CES, Intell will show second generation: Intel Battlemage GPU, so it would great if calling gpu or ipex as device, it would be more generic than call ipx

[Borda]

cc: @lantiga

[johnnynunez]

https://www.phoronix.com/news/Intel-oneDNN-3.0

[HadiSDev]

Any progress on this? :D

[tabahi]

I have been waiting for an update on this. PyTorch is full-on supporting XPUs now:
https://pytorch.org/blog/intel-gpu-support-pytorch-2-5/

Now just waiting for lightning to catch up.

Has anyone tried this experimental and "hypothetical" XPU integration from

https://lightning.ai/docs/pytorch/stable/extensions/accelerator.html