pytorch/xla

Pull Requests Commits

Added model_specific defaults to params such as num_epochs, lr etc.

taylanbil committed 6 years ago

fcb61d9d

add a --model flag to use any (supported) torchvision model

taylanbil committed 6 years ago

c7b87dce

Dump the metrics report if the environment variable METRICS_DEBUG is set.

dlibenzi committed 6 years ago

a76e3c65

Added GRPC service to coordinate hosts in a client mesh.

dlibenzi committed 6 years ago

8d67a32b

Ignore layout when computing strides

asuhan committed 6 years ago

0e69fd5c

Add the initial support for local vs. remote devices, to support feeding slices bigger than single TPU host.

dlibenzi committed 6 years ago

7008ac11

Fix u8 dtype cancellation.

dlibenzi committed 6 years ago

21d16c2e

Block deployment on CPU smoke tests (#725)

jysohn23 committed 6 years ago

Verified 697c2adb

Update README with docker distribution (#723)

jysohn23 committed 6 years ago

Verified 882505a0

Allow saving graphs in HLO format as well.

dlibenzi committed 6 years ago

bf585125

Add API to retrieve XLA internal device names.

dlibenzi committed 6 years ago

eb5db322

Tag the docker builds (#721)

jysohn23 committed 6 years ago

Verified 9de32aec

Add local fix for torch.save on XLA tensors

asuhan committed 6 years ago

a3a246d3

Choose optimal TPU layout for all ranks for which makes sense.

dlibenzi committed 6 years ago

f1ba1cf1

Avoid routing special scalar to device data, to give a chance to the XLA compiler to simplify code.

dlibenzi committed 6 years ago

0a1227b5

Re-align code after PT Variable code changes.

dlibenzi committed 6 years ago

98d669ee

Dockerize Builds (#712)

jysohn23 committed 6 years ago

Verified 918e1ce6

Reset the IR node when setting the device data.

dlibenzi committed 6 years ago

1053b6b0

Enabled booting from a slice bigger than a donut.

dlibenzi committed 6 years ago

03323b18

Add script to install torch and torch_xla from source (#708)

jysohn23 committed 6 years ago

Verified b46e2ade

Fix the daily breakage.

dlibenzi committed 6 years ago

1b5c22da

Fix Variable::data() to Variable::tensor_data() API change.

dlibenzi committed 6 years ago

89b06a12

Remove landed 20783 PR.

dlibenzi committed 6 years ago

fbc76442

Mention Nightly builds in README.

dlibenzi committed 6 years ago

03d12b31

Populate HLO metadata if XLA_HLO_DEBUG is set.

dlibenzi committed 6 years ago

ab90f596

Move HLO generation to IR dump util.

dlibenzi committed 6 years ago

90d3c2b4

Update the README file.

dlibenzi committed 6 years ago

84239a2b

Add support for pooling tensors without batch

asuhan committed 6 years ago

15a8dc62

Add incoming PT PR to fix the s_copy premature removal.

dlibenzi committed 6 years ago

7bd0debf

Fix scalar_tensor for XLA backend

asuhan committed 6 years ago

317ee743

Newer Older