Prerelease: v1.1.1
Pre-release
Pre-release
·
35 commits
to develop
since this release
Overview
Fixed broken CLI API and improved JIT function @torch.compile(...) decorator usage. Also greatly improved README installation guide for CUDA compatible installlation.
Changes
- Explictly run GitHub Action for unit tests with CPU.
- Moved CLI scripts from
/src/bin/**<SCRIPT>**to/src/flat_bug/cli/<SCRIPT>submodule. (and changepyproject.tomlto use the new path) - Improved README installation guide, particularly with steps related to PyTorch integration with
uv/pip. (https://docs.astral.sh/uv/guides/integration/pytorch/#configuring-accelerators-with-optional-dependencies) - Change all use of
@torch.compileto@torch.compile(fullgraph=True, mode="max-autotune-no-cudagraphs"). - Add
torch.set_float32_matmul_precision("medium")to the start offlat_bug.cli.fb_predict.py:maini.e. the main CLI API for this package.
Full Changelog: v1.1-beta.0...v1.1-beta.1