This update utilizes new PyTorch 1.8 features, most importantly native complex operations for the index_add_ function in the adjoint on the GPU. It also removes some thread management code that led to a performance regression on PyTorch 1.8 (Issue #25). A further list of changes is below:
Update requirements.txt and dev-requirements.txt to latest packages.
Remove calc_split_sizes - we can now use tensor_split.
Removed some calls to tensor attributes - these can be expensive.
Removal of kwarg usage for some torch.jit.script functions - these can behave strangely with scripted functions.
Removal of index_put for accumulation. We now only use index_add.
This update utilizes new PyTorch 1.8 features, most importantly native complex operations for the
index_add_function in the adjoint on the GPU. It also removes some thread management code that led to a performance regression on PyTorch 1.8 (Issue #25). A further list of changes is below:requirements.txtanddev-requirements.txtto latest packages.calc_split_sizes- we can now usetensor_split.kwargusage for sometorch.jit.scriptfunctions - these can behave strangely with scripted functions.index_putfor accumulation. We now only useindex_add.