5.1. PopTorch

3.4.0

None

Fix corner-case where model parameters with requires_grad=False were not handled correctly during training.
Propagate the requires_grad setting for parameters that are set per-replica.

None

PopTorch always uses the tanh approximation for GELU, even if approximate is set to ‘none’. (See the PyTorch documentation for torch.nn.functional.gelu and GELU).

A workaround is to use a custom GELU implementation that uses the error function. See the DINO example for one approach (ERF GELU implementation and its use in the model).

None

Added support for the following ops:
- torch.nn.Mish
- torch.bucketize
- torch.cdist
- torch.sort
- torch.take_along_dim
- torch.bincount
Added support for the following torch_cluster ops:
- radius
- grid
- knn
Added drop-in replacements for the following torch_cluster operations (replace the torch_cluster ops with poptorch ops ones): * fps * nearest
Added a cond op for inference models only. This op conditionally executes one of two branches depending on the input condition.
Extended the API for copying data structures to the IPU. copyNamedBuffersToDevice allows for a named buffer to be copied.
FixedSizeCollator can now pad to different node and edge types.
The PopTorch wheel now has dependencies on torchvision and torchaudio. This will prevent upgrade to an unsupported PyTorch version when installing other third-party packages that depend on torchvision or torchaudio.
Added support for the largest=False` option in the torch.topk op.
Added the compilationTime function to extract the total compilation time from the compiled PopTorch model.

Fixed compilation of models with torch.norm inside the for_loop op.
Fixed the clamp dtype mismatch error. torch.clamp used to raise an incompatible type error.
torch.var used to raise an error when an input dimension was negative. The fix converts the input negative integer to a positive integer so it can be used as an index of the input shape vector.
Fixed the torch.round behaviour in PopTorch to use a “round half down” method to match the behaviour in PyTorch. PopTorch previously used a “round half up” method.
Fixed the Int32-Float32 op not being processed.

torch.compile is not supported in PopTorch. This has been documented in PyTorch for the IPU: User Guide.

None

None