| Commit message (Collapse) | Author | Age |
|
|
|
|
|
| |
Previously, we would schedule a closure for each ResourceHandleOp, because it is erroneously considered to be "expensive". This would cost several microseconds per op, whereas the execution cost of this kernel is as little as 100ns. This change causes these kernels to execute inline at the beginning of a step.
PiperOrigin-RevId: 212712378
|
|
|
|
| |
PiperOrigin-RevId: 212706610
|
|
|
|
|
|
| |
BatchNorm bijector to work with event_ndims > 1.
PiperOrigin-RevId: 212705787
|
|
|
|
| |
PiperOrigin-RevId: 212702577
|
|
|
|
|
|
|
|
| |
Instead of multiple primitive Eigen ops in Conv3DBackpropFilter, call directly into Eigen function.
Modest ~10-25% latency improvement and ~10-20% peak memory reduction.
PiperOrigin-RevId: 212701797
|
|
|
|
| |
PiperOrigin-RevId: 212701024
|
|
|
|
| |
PiperOrigin-RevId: 212698623
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
better:
Currently, static shape inference propagates shapes of tensors, but in some cases, we do need values; for this, we use input_tensors (from Const input tensor) and input_tensors_as_shapes and output_tensors_as_shapes (these are ShapeHandle format, but has values, currently only for 1D vector).
This CL enhances propagation of input_tensors_as_shapes and output_tensors_as_shapes to improve static shape inference.
(1) forward scalar Const as input_tensors_as_shapes (currently, only 1D vector),
(2) export input_tensors_as_shapes, output const tensor, and output_tensors_as_shapes to the values of inferred input/output TensorProperties (currently, only input const tensors are exported as values),
(3) use input_tensors_as_shapes as Const tensor to function input (currently, only Const tensors),
(4) forward input_tensors_as_shapes to output_tensors_as_shapes for Identity op,
(5) when Pack op concats scalar values to form output_tensors_as_shapes, currently it uses only input_tensors (from Const input tensors), but this CL change Pack to use input_tensors_as_shapes as well.
PiperOrigin-RevId: 212696959
|
|
|
|
|
|
|
|
|
|
|
|
| |
These globals were holding onto graphs including FuncGraphs, which
held onto captured tensors leaving garbage around.
This change also adds a test to catch garbage like this in the future.
To make the test work, I needed to manually breakup some reference
cycles caused by OrderedDicts. We should probably have a custom impl
of OrderedDict similar to the one in Python3 and avoid these issues.
PiperOrigin-RevId: 212694290
|
|
|
|
|
|
| |
Re-assigning unique IDs broke serialization of HloSchedule, and keeping IDs stable improves the fidelity of the proto serialization. This change requires that instructions in HLO module protos have valid, module-scope-unique ids so change the XLA builder to hand out module-scope-unique ids. Previously, instruction ids were only unique in the computation scope.
PiperOrigin-RevId: 212692339
|
|
|
|
|
|
| |
returns features and labels as a list instead of dict.
PiperOrigin-RevId: 212685344
|
|
|
|
| |
PiperOrigin-RevId: 212684548
|
|
|
|
|
|
| |
`bucket_by_sequence_length`.
PiperOrigin-RevId: 212684420
|
|
|
|
|
|
| |
https://arxiv.org/pdf/1609.08144.pdf).
PiperOrigin-RevId: 212683753
|
|
|
|
|
|
| |
Gradients annotations.
PiperOrigin-RevId: 212682657
|
|
|
|
| |
PiperOrigin-RevId: 212674212
|
|
|
|
| |
PiperOrigin-RevId: 212668653
|
|
|
|
|
|
|
|
| |
This now matches the definition. I fixed it here rather than in the definition as it seems every call to this function names the variable "num_components".
I also tidied up the comment a little.
PiperOrigin-RevId: 212668416
|
|
|
|
|
|
| |
tests
PiperOrigin-RevId: 212668288
|
|
|
|
|
|
| |
Just for completeness, this isn't being used currently.
PiperOrigin-RevId: 212667528
|
|\
| |
| |
| | |
PiperOrigin-RevId: 212667200
|
| |
| |
| |
| | |
PiperOrigin-RevId: 212658047
|
| |
| |
| |
| |
| |
| | |
optimization pass, instead of a step in XlaCompiler.".
PiperOrigin-RevId: 212657932
|
| |
| |
| |
| |
| |
| |
| | |
will remove this from contrib.
linear_operator_addition is hidden from the public API.
PiperOrigin-RevId: 212655087
|
| |
| |
| |
| |
| |
| |
| |
| | |
copying that type.
This avoids unnecessary string copies and deallocations in the ReadVariableOp, and similar ops.
PiperOrigin-RevId: 212652588
|
| |
| |
| |
| | |
PiperOrigin-RevId: 212651704
|
| |
| |
| |
| |
| |
| | |
The test would fall back to GPU:0 when unable to find a GPU. This should be CPU.
PiperOrigin-RevId: 212649435
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
TFLite has an optimized matmul path for cases where the input and
filter tensors have matching width+height. However, this case doesn't
properly account for multiple *batches*. Account for this and add
an appropriate test.
Credit to zgxnet for the bug and proposed fix.
Fixes #21817
PiperOrigin-RevId: 212645329
|
| |
| |
| |
| | |
PiperOrigin-RevId: 212645190
|
| |
| |
| |
| | |
PiperOrigin-RevId: 212643986
|
| |
| |
| |
| | |
PiperOrigin-RevId: 212642629
|
| |
| |
| |
| | |
PiperOrigin-RevId: 212601523
|
| |
| |
| |
| | |
PiperOrigin-RevId: 212600364
|
| |
| |
| |
| | |
PiperOrigin-RevId: 212595533
|
| |
| |
| |
| |
| |
| | |
correctly. Refactor MapDefunOp for correctness + better C++ hygiene
PiperOrigin-RevId: 212589285
|
| |
| |
| |
| | |
PiperOrigin-RevId: 212582822
|
| |
| |
| |
| | |
PiperOrigin-RevId: 212577288
|
| |
| |
| |
| | |
PiperOrigin-RevId: 212569958
|
| |
| |
| |
| | |
PiperOrigin-RevId: 212568029
|
| |
| |
| |
| | |
PiperOrigin-RevId: 212565231
|
| |
| |
| |
| |
| |
| | |
performance.
PiperOrigin-RevId: 212557406
|
| |
| |
| |
| | |
PiperOrigin-RevId: 212556651
|
| |
| |
| |
| |
| |
| | |
it more accurate.
PiperOrigin-RevId: 212555968
|
| |
| |
| |
| | |
PiperOrigin-RevId: 212555263
|
| |
| |
| |
| | |
PiperOrigin-RevId: 212551965
|
|\ \
| | |
| | |
| | | |
PiperOrigin-RevId: 212551803
|
| | |
| | |
| | |
| | | |
PiperOrigin-RevId: 212548263
|
| | |
| | |
| | |
| | | |
PiperOrigin-RevId: 212545735
|
| | |
| | |
| | |
| | |
| | |
| | | |
the frontend autograph module in contrib for backward compatibility. Certain files, like notebooks and the readme file may be referenced from the outside, so a copy of those is kept as well. In addition, the notebooks subdirectory of examples is also kept in contrib because the extension the build file relies on is not available in the PIP package.
PiperOrigin-RevId: 212543067
|
|\ \ \
| | | |
| | | |
| | | | |
PiperOrigin-RevId: 212543062
|