pytorch/xla

Pull Requests Commits

Allow the PyTorch include folder to be overridden via environment.

dlibenzi committed 6 years ago

a77dd070

Add gradient clipping patch

asuhan committed 6 years ago

8fcdcb39

Track post order size in op-by-op mode.

dlibenzi committed 6 years ago

d91a8acc

Remove more JIT leftovers

asuhan committed 6 years ago

28dcfbef

Remove autodiff patch as well

asuhan committed 6 years ago

3ce4e992

Remove the JIT code and all its dependencies.

dlibenzi committed 6 years ago

b7c0ab3f

Added split executor capability, which runs a given IR graph one IR node at a time.

dlibenzi committed 6 years ago

c96789ef

Don't use torchscript version for aten::max_pool2d autodiff

asuhan committed 6 years ago

4f99844f

Remove patch which landed

asuhan committed 6 years ago

262c360b

Use two-pass lowerings for argmin and argmax

asuhan committed 6 years ago

c9d654ad

Fix pending PR extension

asuhan committed 6 years ago

a5d996d0

Fix XlaModule constructor after PyTorch changes

asuhan committed 6 years ago

470b15e1

Local fix for max_pool2d differentiation

asuhan committed 6 years ago

160d9cb5

Use xla::TorchGather to lower Gather

asuhan committed 6 years ago

eb769a38

Use xla::TorchIndexSelect to lower IndexSelect

asuhan committed 6 years ago

1080e7f7

Add pending PyTorch patch fixing equal for boolean tensors

asuhan committed 6 years ago

bde41239

Update to latest TF

asuhan committed 6 years ago

73919342

Add step-mark operation, and split sync-tensors into a separate operation.

dlibenzi committed 6 years ago

9e59aa5c

Simplify XLA tensor code generator further, after the recent ATEN changes.

dlibenzi committed 6 years ago

de703fa7

Fix XLA tensor code generation to cope with the PyTorch/ATEN changes.

dlibenzi committed 6 years ago

2ba3696c

Throw when trying to access storage of an XLA tensor

asuhan committed 6 years ago

0ec9b5ac

Bump up the IR graph trim thresholds.

dlibenzi committed 6 years ago

2bde0d0f

Extend the graph-saving debug utility to more places where we enter the XLA tensor world for execution.

dlibenzi committed 6 years ago

25d099bf

Read async status in wait operation.

dlibenzi committed 6 years ago

f3221842

Remove Hash() default to DataHash(), and restrict to arithmetic for it.

dlibenzi committed 6 years ago

b2f6d951

Fix code by coping with the fact that ATEN copy() is no more.

dlibenzi committed 6 years ago

5dd0ea12

Return a torch.device consistently.

dlibenzi committed 6 years ago

d2e239b0

Fix index_put starting with null indices

asuhan committed 6 years ago

64de8e6a

Add base support for ATEN Bool tensor type.

dlibenzi committed 6 years ago

04faa305

Send tensors to XLA device before running tests, in case of tensor-modifying test functions.

dlibenzi committed 6 years ago

9199f36f

Older