5.7. PopXL

3.1.0

New features

• Added support for float8 inference (host-based conversion, variables, constants, view changers, cast, matmul and conv).

• Added support for replicated tensor sharding (RTS) and replica-grouped initialisation with multiple instances for remote variables.

• Added the following ops:

  • scatter_reduce

  • pow

  • subsample

  • batch_norm_inference

• Added the diag method to the tensor class.

Bug Fixes

• None

Other improvements

• Disabled storing variable data with executable caching.

• Used subsample op to handle slicing with step > 1.

Known issues

None

Compatibility changes

None

3.0.0

New features

• Add abs, cos and sin operations providing additional operation coverage.

• Add ability to compile without acquiring IPUs providing improved IPU resource utilisation.

• Add DecomposeSum transform.

• Add support for if operation allowing for more expressive control flow.

• Add ability to broadcast binary operations.

• Enable support for all stride configurations collectives and replicated variables.

  The ability to stride replicas in collectives is useful for advanced use cases like Tensor Model Parallel (TMP).

  For more details refer to Available operations in popxl.ops.collectives in the section on Supported operations in the PopXL User Guide and API (experimental).

Bug Fixes

None

Other improvements

• Improvements to PopXL User Guide and API (experimental):

  • add description of environment variables

  • improved sections for custom operations

Known issues

None

Compatibility changes

• Change replicated_all_gather op so that the output tensor has shape [size, *shape].