Skip to content

v0.12.0

Latest
Latest
Compare
Choose a tag to compare
@dfalbel dfalbel released this 24 Jan 11:22
v0.12.0
cf9819d
This commit was signed with the committer’s verified signature.
dfalbel Daniel Falbel
SSH Key Fingerprint: rbExZVSaXO4uvLL2/Yr03fwRVBu7vtDz6+l5PUTCY8E Learn about vigilant mode.

Breaking changes

  • New torch_save serialization format. It's ~10x faster and since it's based on safetensors, files can be read with any safetensors implementation. (#1071)
  • Updated to LibTorch 2.0.1. (#1085)
  • torch_load no longer supports device=NULL to load weights in the same device they were saved. (#1085)
  • Lantern binaries and torch pre-built binaries are now built on Ubuntu 20.04. (#1124)

New features

  • Added support for CUDA 11.8. (#1089)
  • Added support for iterable datasets. (#1095)

Bug fixes

  • fix printer of torch device (add new line at the end)
  • as.array now moves tensors to the cpu before copying data into R. (#1080)
  • Fixed segfault caused by comparing a dtype with a NULL. (#1090)
  • Fixed incorrect naming of complex data type names, such as torch_cfloat64. (#1091)
  • Fixed name of the out_features attribute in the nn_linear module. (#1097)
  • Fixed issues when loading the state dict of optimizers and learning rate schedulers. (#1100)
  • Fixed bug when cloning nn_modules with empty state dicts. (#1108)
  • distr_multivariate_normal now correctly handles precision matrix's. (#1110)
  • Moved length.torch_tensor implementation to R7 to avoid problems when a torch dataset has the torch_tensor class. (#1111)
  • Fixed problem when deep cloning a nn_module. (#1123)
Assets 2