How to use this with SX-Aurora TSUBASA

[yunkai1841]

Is there any documentation or sample code?

[mergian]

Hey @yunkai1841,

this project is a helper library for integration of the SX-Aurora into PyTorch for the NEC SOL compiler. Without SOL it's not really useful.

If you still want to try, here an example:

import torch
import veda.pytorch # loads the extension

A = torch.rand(1, 2, 3).ve()
B = torch.rand(1, 2, 3).ve()
C = A + B
print(C)

[owainkenwayucl]

As someone with one of these machines this seems... quite awkward. Is the only way to get Pytorch running on this device to get our sales rep to sign us up for the SOL closed beta?

[mergian]

Dear @owainkenwayucl

short answer:
Yes, unfortunately it's the only way. But you can request access to the SOL4VE closed beta yourself. Follow this link and fill out the closed beta form or send me an email.

long answer:
The SX-Aurora is not build as AI accelerator and has non of the AI relevant technologies implemented, e.g., int8, fp4, fp8, fp16, bfloat16, tf32, tensor cores, etc. So the expected performance of SX-Aurora running AI is rather low compared to other AI specialized hardware. Therefore an investment into maintaining a 100% functioning PyTorch, TensorFlow, JAX, ... implementation for a hardware that is not specifically designed to run AI workloads, is very expensive and a huge waste of resources. However, SOL is an NEC internally used AI compiler that allows to easily add and maintain different hardware architectures with little effort. So we added a SX-Aurora plugin and created this minimalistic library for basic functionality within PyTorch (memalloc, memfree, copy, +, -, ...) as demonstrator, on minimal effort basis.

As said, if you are really interested, apply to the SOL4VE closed beta. There will also be a new release of SOL end of this month. But don't expect quantum leaps on hardware that is not specialized for AI.

Best