Supported Ops
The list of individual operations that have been verified for functional accuracy in the eager mode execution environment is as follows.
The below operations are mostly accelerated on the Rebellions NPU, while the remaining PyTorch operations are executed on the CPU. The number of operations accelerated on the Rebellions NPU is expected to increase gradually.
| PyTorch Ops | Note |
|---|---|
| torch.abs | - |
| torch.add | - |
| torch.addmm | - |
| torch.bmm | - |
| torch.ceil | - |
| torch.clamp | - |
| torch.div | - |
| torch.floor | - |
| torch.log | - |
| torch.masked_fill | - |
| torch.max | - |
| torch.mean | - |
| torch.min | - |
| torch.mm | - |
| torch.mul | - |
| torch.neg | - |
| torch.pow | - |
| torch.rsqrt | - |
| torch.silu | - |
| torch.softmax | - |
| torch.sub | - |
| torch.trunc | - |
| torch.where | - |