v0.1.5

What's New

• FIX: pytorch 1.11 adam maximize=False #23
• FIX: buffer size not aligned
• FIX: as_inf_or_nan cuda invalid configuration argument
• FIX: eval not working in CheckpointingBlock
• FIX: KeyError when getting inspected tensor by name #22

New Contributors

• @xcjthu made their first contribution in #22
• @yushengsu-thu made their first contribution in #23

Full Changelog: 0.1.4...0.1.5

v0.1.4

• Fixed RuntimeError when parameter size is smaller than world size
• Set default loss scale to 1
• Fixed distributed.AllGather in pytorch 1.11
• Fixed address is already in use error

v 0.1.3

• Fixed expanded size mismatch.

v0.1.2

• Add bmtrain model wrapper #2
• Add docs #3 #13 #15 #14
• Support PyTorch 1.11 (fixed slices are only supported in _UntypedStorage.__getitem__) #16
• Add more lr schedulers #12
• Adapt to OpenDelta #7
• Fixed bugs #1 #8
• Speed up BMTrain's optimizers

v0.1.1

• Fixed a lot of bugs ...

0.0.15

remove run.sh