veda-pytorch 14.0.1

VEDA PyTorch

VEDA PyTorch

VEDA PyTorch is a library to add device support for the NEC SX-Aurora TSUBASA into PyTorch.

Release Notes

Version	Comment
v14.0.1	Increased c10d compatibility (tested on v2.5-2.9) Added better error handling when running on systems with missing VEOS installation.
v14	Starting with v14, VEDA PyTorch is no longer distributed as precompiled binary but gets compiled as PyTorch C++ extension on the target machine. So you don't need to install a matching binary package anymore! We further added a experimental implementation for using NEC MPI. You can create the process group as follows: torch.distributed.init_process_group( backend = 'veda', world_size = os.environ['MPISIZE'], rank = os.environ['MPIRANK'], store = torch.distributed.Store() ) Further changes: Added support for PyTorch v2.9.0 Added arange.start_out Added function tracing. Activate using TUNGL_LOG=TRACE Bugfix for aten::cat.out Bugfix for copy_ Bugfix for torch.load(location='ve') Removed unnecessary context sync
v13	Fixed torch.ve.set_device Fixed allocation on wrong VE in multi-process execution Improved error messages Upgraded build script for PyTorch >=2.7!
v12	Added auto plugin loading for Pytorch. import veda.pytorch is no longer required with PyTorch >=2.5!
v11	Fixed shutdown problem in mixed GPU/VE use cases.
v10	Support for PyTorch v2.3.1 Support for SX-Aurora VE3
v9	Support for PyTorch v2.3.0
v8	Added torch.logical_not
v7	Support for PyTorch v2.0.0 Support for PyTorch v1.13.0 Added torch.log1p
v6	Support for PyTorch v1.12.0 and v1.12.1
v5	Added torch.clamp torch.clamp_max torch.clamp_min torch.exp torch.log torch.norm torch.pow torch.where Fixed conversion from numeric value to bool Fixed calling torch.ve.memory_allocated() without device id Preventing 0-byte allocations from PyTorch to be passed on to VEDA
v4	fixed possible segfault in Tensor resize if no storage is initialized fixed dtype handling in Scalar to Tensor operations
v3	added squeeze and unsqueeze handlers
v2	Minor changes to enable PyTorch v1.11.0 Fixed vedaInit error checking to ignore if already initialized
v1	Initial Release