pytorch/xla

Pull Requests Commits

Fix build script.

dlibenzi committed 6 years ago

d7b7aafe

Reformat Python code using Google pyformat.

dlibenzi committed 6 years ago

57926630

Import latest TF features and avoid tuple decomposition calls/RPCs.

dlibenzi committed 6 years ago

32ab44c7

Bring new TF commits with the auto-tuple-decompose XRT feature.

dlibenzi committed 6 years ago

8c9cf384

The Python GC does not immediately release the collected objects, so we need to wait for it to happen before heading out.

dlibenzi committed 6 years ago

cbf11819

Make an attempt to release the pending XLA tensors on exit.

dlibenzi committed 6 years ago

2cc427ee

Fix typo: set_inputs_gardients -> set_input_gradients

asuhan committed 6 years ago

b9c2bed1

Added MNIST training test using the PyTorch legacy (or almost) interface.

dlibenzi committed 6 years ago

54a33873

Fix translator.cpp formatting

asuhan committed 6 years ago

452ad483

Use implicit once-like initialization for function static variables and simplify code further.

dlibenzi committed 6 years ago

50cc3f5c

Made the computation client live forever (lie all the static data does in g3).

dlibenzi committed 6 years ago

36c1a1ff

Merge pull request #41 from pytorch/asuhan/remove_mkl_dnn_dep

jysohn23 committed 6 years ago

Verified d2f255bb

Implement condition variable based triggering for handle releases instead of sleeps.

dlibenzi committed 6 years ago

bc4dd1e1

Remove mkldnn dependency

asuhan committed 6 years ago

4afca440

dlibenzi committed 6 years ago

ede9c967

Move handle release in background. (#37)

dlibenzi committed 6 years ago

551c2e89

The compare model was broken ... this is why resnet18 was working.

dlibenzi committed 6 years ago

19cbea9a

Enable Continuous Integration for pytorch/xla

jysohn23 committed 6 years ago

Verified c26880ad

Enable Continuous Integration for pytorch/xla

jysohn23 committed 6 years ago

37fa9059

Move the metric snapshotting after the release handles call, since that already has its metric.

dlibenzi committed 6 years ago

091a334e

Refactored the tests and added resnet18 test.

dlibenzi committed 6 years ago

81317626

Limit the number of differences printed during a test failed dump.

dlibenzi committed 6 years ago

830a16f0

Try to get out of the mess C10 did trying to be Google with logging and CHECK macros.

dlibenzi committed 6 years ago

1a0897de

Enabled use of BF16 in PyTorch computations via the XLA_USE_BF16=1 environment setting.

dlibenzi committed 6 years ago

bc7bc15e

Improved PromoteValues() API.

dlibenzi committed 6 years ago

4b55a904

Added the creation of a local TF service to handle local CPU device (to allow running XRT on local workstation).

dlibenzi committed 6 years ago

0346e9df

Added handling for multiple XRT sessions.

dlibenzi committed 6 years ago

69158c1d

Use common lock for conditions and use lock context manager to simplify code.

dlibenzi committed 6 years ago

3b9d3683

Fix _C -> _XLAC.

dlibenzi committed 6 years ago

d23deefe

Use more threads to to the conversion to XLA tensors, and make sure the order of the samples handed off to the queue is deterministic.

dlibenzi committed 6 years ago

b5836b90

Older