| Commit message (Collapse) | Author | Age |
|
|
|
|
|
| |
Rename RecvTensorAsync method to GrpcRecvTensorAsync to fix shadowing of method in Worker with a different signature.
PiperOrigin-RevId: 164454701
|
|
|
|
|
|
|
|
| |
This would save a significant amount of disk space.
Also converted a bunch of "LOG(INFO)" to "std::cout" since info logging is not printed in the open-source build while most messages are interesting to users.
PiperOrigin-RevId: 164451472
|
|
|
|
| |
PiperOrigin-RevId: 164409956
|
|
|
|
|
|
| |
Avoid unnecessary transposes in tf.layers.separable_conv2d (and implicitly in tf.contrib.layers.separable_conv2d).
PiperOrigin-RevId: 164408360
|
|
|
|
|
|
| |
Such initializer could be used e.g. for IRNN implementation (https://arxiv.org/pdf/1504.00941.pdf).
PiperOrigin-RevId: 164403770
|
|
|
|
| |
PiperOrigin-RevId: 164369858
|
|
|
|
| |
PiperOrigin-RevId: 164368207
|
|
|
|
|
|
|
|
|
| |
* Mark the evaluator plugin as alwayslink so it doesn't get stripped out by the linker.
* Add a generic LayoutAssignment pass to the pass pipeline; otherwise the entry computation has no layout and Service::Execute CHECK-fails in the AllocationTracker.
* Register the default computation placer for the evaluator backend.
* Add an replay_computation_hlo_evaluator binary that can replay computation snapshots via the HLO evaluator.
PiperOrigin-RevId: 164364780
|
|
|
|
| |
PiperOrigin-RevId: 164352481
|
|
|
|
| |
PiperOrigin-RevId: 164350154
|
|
|
|
| |
PiperOrigin-RevId: 164349082
|
|
|
|
| |
PiperOrigin-RevId: 164343192
|
|
|
|
| |
PiperOrigin-RevId: 164342823
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
maximum age and cache size. Relevant environment variables:
GCS_STAT_CACHE_MAX_AGE (default 0, indicating that the cache is disabled)
GCS_STAT_CACHE_MAX_ENTRIES (default 0, indicating no limit)
GCS_MATCHING_PATHS_CACHE_MAX_AGE (default 0, indicating that the cache is disabled)
GCS_MATCHING_PATHS_CACHE_MAX_ENTRIES (default 0, indicating no limit)
Max age values for these caches are specified in seconds.
Also changed the default GCS read cache block size to 128 MB, so that the block cache is used by default.
PiperOrigin-RevId: 164341936
|
|
|
|
| |
PiperOrigin-RevId: 164336354
|
|
|
|
| |
PiperOrigin-RevId: 164330926
|
|
|
|
| |
PiperOrigin-RevId: 164324195
|
|
|
|
| |
PiperOrigin-RevId: 164323827
|
|
|
|
| |
PiperOrigin-RevId: 164323269
|
|
|
|
|
|
| |
safe to do so.
PiperOrigin-RevId: 164322260
|
|
|
|
| |
PiperOrigin-RevId: 164319817
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
END_PUBLIC
---
Commit cf375f067 authored by A. Unique TensorFlower<gardener@tensorflow.org>
Committed by Benoit Steiner<benoitsteiner@users.noreply.github.com>:
Adds cudnn_rnn_ops_op_lib and cudnn_rnn_kernels to contrib_ops_op_lib and
contrib_kernels respectively.
PiperOrigin-RevId: 164170971
---
Commit 95ec58e27 authored by Asim Shankar<ashankar@google.com>
Committed by Benoit Steiner<benoitsteiner@users.noreply.github.com>:
C API: Make TF_TensorFromTensor return an error instead of just logging it.
PiperOrigin-RevId: 164167582
---
Commit 15175c870 authored by Jonathan Hseu<jhseu@google.com>
Committed by Benoit Steiner<benoitsteiner@users.noreply.github.com>:
Build fixes.
- Allow var_list as a positional argument in CrossShardOptimizer.
- Set the number of shards to 1 when not running on TPU, to allow evaluate() and predict() on CPU/GPU to work.
PiperOrigin-RevId: 164161640
---
Commit bd3e894f7 authored by Yao Zhang<yaozhang@google.com>
Committed by Benoit Steiner<benoitsteiner@users.noreply.github.com>:
Support freeze mode for fused batch norm.
PiperOrigin-RevId: 164149032
---
Commit e6b6b84c0 authored by Asim Shankar<ashankar@google.com>
Committed by Benoit Steiner<benoitsteiner@users.noreply.github.com>:
C API: TF_Tensors will always be in host memory.
This change undoes some experimentation in
commit 22651083406ca01ac9d481e3367a3510d25f88cd
and restores TF_Tensor behavior to what is was prior to that change.
PiperOrigin-RevId: 164146670
---
Commit 8bf3f88f7 authored by Peter Hawkins<phawkins@google.com>
Committed by Benoit Steiner<benoitsteiner@users.noreply.github.com>:
[TF:XLA] Add _XLASend and _XLARecv TF ops that wrap the XLA Send/Recv HLO ops.
PiperOrigin-RevId: 164124764
---
Commit 626d3200f authored by Peter Hawkins<phawkins@google.com>
Committed by Benoit Steiner<benoitsteiner@users.noreply.github.com>:
[XLA] Add test blacklist mechanism for XLA C++ unit tests.
PiperOrigin-RevId: 164124423
---
Commit 359cc5f5e authored by A. Unique TensorFlower<gardener@tensorflow.org>
Committed by Benoit Steiner<benoitsteiner@users.noreply.github.com>:
Document dict ordering in nest and make it consistent with sonnet.
PiperOrigin-RevId: 164114335
---
Commit 05813b531 authored by A. Unique TensorFlower<gardener@tensorflow.org>
Committed by Benoit Steiner<benoitsteiner@users.noreply.github.com>:
Go: Update generated wrapper functions for TensorFlow ops.
PiperOrigin-RevId: 164089206
---
Commit c451f465d authored by Anna R<annarev@google.com>
Committed by Benoit Steiner<benoitsteiner@users.noreply.github.com>:
BEGIN_PUBLIC
Automated g4 rollback of changelist 164078808
PiperOrigin-RevId: 164318935
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
35.6 seconds to 33.3 seconds (average of three runs) (+6.5% improvement).
(1) Avoid extra hash table lookups in HeapSimulator by holding onto the pointer
to the hash table value, rather than looking it up from the same key multiple
times.
(2) In HeapSimulator, reuse operand_buffers_to_free and dead_buffers_to_free
across all instructions, rather than allocating new vectors on every
instructions.
(3) Avoid use of Printf and improve efficiency of string generation for
HloInstruction::ToString and ShapeUtil::HumanStringWithLayout.
PiperOrigin-RevId: 164314222
|
|
|
|
| |
PiperOrigin-RevId: 164313140
|
|
|
|
| |
PiperOrigin-RevId: 164309367
|
|
|
|
|
|
| |
constructors.
PiperOrigin-RevId: 164308929
|
|
|
|
|
|
| |
Set output shapes of RestoreV2 and RestoreSlice the same way with Restore.
PiperOrigin-RevId: 164306383
|
|
|
|
| |
PiperOrigin-RevId: 164305108
|
|
|
|
| |
PiperOrigin-RevId: 164303231
|
|
|
|
| |
PiperOrigin-RevId: 164301697
|
|
|
|
| |
PiperOrigin-RevId: 164295665
|
|
|
|
| |
PiperOrigin-RevId: 164293021
|
|
|
|
|
|
| |
shape (None, None, None) when only one dimension is unknown, the result can have shapes like (40, None, 30).
PiperOrigin-RevId: 164289320
|
|
|
|
|
|
| |
makes grappler shape inference much faster.
PiperOrigin-RevId: 164288667
|
|
|
|
|
|
| |
Add missing op grouping in Cholesky benchmark.
PiperOrigin-RevId: 164281947
|
|
|
|
| |
PiperOrigin-RevId: 164280535
|
|
|
|
|
|
| |
actual data. This makes debugging constant folding easier.
PiperOrigin-RevId: 164279921
|
|
|
|
|
|
|
| |
because const Foo<T>* and Foo<const T>* are not convertible in C++ which broke
the internal machinery of the iterator in the const case.
PiperOrigin-RevId: 164276236
|
|
|
|
|
|
| |
xla::CanShareOperandBufferWithUser.
PiperOrigin-RevId: 164270746
|
|
|
|
| |
PiperOrigin-RevId: 164245854
|
|
|
|
|
|
|
|
| |
Doing conversion in TensorBoard python code is slow, and the conversion
happens every time a trace is requested. It is also much faster to do the
json conversion in C++ code than in python code.
PiperOrigin-RevId: 164237242
|
|
|
|
| |
PiperOrigin-RevId: 164233867
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This change adds a CPU-specific flag: xla_cpu_optimize_for_size
When this flag is passed, it changes the optimizers to run
more or less analogously to LLVM's -Os optimizations.
There are two things that turning on the code size optimization option
controls:
* the internal settings of some optimization passes (which is mostly
controlled through a function attribute)
* the passes that get run (which is decided by the pass manager)
This change also refactors the code by reorganizing the way
that CPU backend specific flags are queried, as well as some
other minor refactoring.
PiperOrigin-RevId: 164218771
|
|
|
|
| |
PiperOrigin-RevId: 164218136
|
|
|
|
|
| |
RELNOTES: TensorFlow Debugger (tfdbg) command-line interface: Support evaluation of arbitrary Python and numpy (np) expressions with debug tensor names enclosed in pairs of backtics. E.g., tfdbg> eval 'np.sum(`Softmax:0`, axis=1)'.
PiperOrigin-RevId: 164217384
|
|
|
|
|
|
|
|
|
| |
(TypedVisitor) are only invoked through HloInstruction->Visit which will not
endend up calling the TypedVisitor::Preprocess.
Also add Postprocess to log evaluated literal value.
PiperOrigin-RevId: 164217266
|
|
|
|
| |
PiperOrigin-RevId: 164215900
|
|
|
|
| |
PiperOrigin-RevId: 164215673
|
|
|
|
| |
PiperOrigin-RevId: 164211593
|
|
|
|
|
|
|
| |
Items are named as tf.initializers.zeros... old aliases like
tf.zeros_initializers still work.
PiperOrigin-RevId: 164211375
|