| Commit message (Collapse) | Author | Age |
|
|
|
|
|
| |
TF clip_ops, also added user input check for clipnorm and clipvalue >= 0 if set
PiperOrigin-RevId: 202516320
|
|
|
|
|
|
| |
ops at runtime.
PiperOrigin-RevId: 202514848
|
|
|
|
|
|
| |
into the other.
PiperOrigin-RevId: 202513508
|
|
|
|
|
|
|
| |
error_format_tag is a helper for building interpolatable strings as part of
a project to improve Python error messages in TensorFlow.
PiperOrigin-RevId: 202509392
|
|
|
|
|
|
| |
np.prod() to use np.int64 in a few places.
PiperOrigin-RevId: 202505308
|
|
|
|
|
|
| |
ops to use the corresponding free functions in namespace xla:: instead.
PiperOrigin-RevId: 202505306
|
|
|
|
| |
PiperOrigin-RevId: 202505228
|
|
|
|
| |
PiperOrigin-RevId: 202504925
|
|
|
|
|
|
| |
and `--input_exprs` headings to match option names.
PiperOrigin-RevId: 202504009
|
|
|
|
|
|
| |
Also enable cloud tpu profiler to detect the TF version for better version compatibility.
PiperOrigin-RevId: 202503162
|
|
|
|
| |
PiperOrigin-RevId: 202501055
|
|
|
|
| |
PiperOrigin-RevId: 202496488
|
|
|
|
| |
PiperOrigin-RevId: 202489637
|
|
|
|
|
|
| |
Only the name of the version is changing here; the version is unchanged.
PiperOrigin-RevId: 202486284
|
|
|
|
| |
PiperOrigin-RevId: 202472329
|
|
|
|
|
|
| |
We returned one-element tensors with uninitialized content, which msan didn't like.
PiperOrigin-RevId: 202463090
|
|
|
|
|
|
|
|
|
| |
The previous implementation recompiled the shape regex at every call
what is an expensive opertaion. The new implementation improves the hlo
text parsing time for very large models for up to 9x by eliminating this
overhead.
PiperOrigin-RevId: 202454354
|
|
|
|
|
|
| |
compatible with hlo string syntax.
PiperOrigin-RevId: 202445509
|
|
|
|
|
|
| |
Without domain propagation in dataflow analysis we end up in inconsistent domain instructions with BF16 as output and F32 as input. In case of tuple shapes these are not fixed by bfloat16_normalization, and later on they cause asserts once the domain instructions are removed.
PiperOrigin-RevId: 202442786
|
|
|
|
| |
PiperOrigin-RevId: 202423156
|
|
|
|
| |
PiperOrigin-RevId: 202419595
|
|
|
|
| |
PiperOrigin-RevId: 202415942
|
|
|
|
| |
PiperOrigin-RevId: 202412660
|
|
|
|
|
|
| |
using auto.
PiperOrigin-RevId: 202409729
|
|
|
|
| |
PiperOrigin-RevId: 202403235
|
|
|
|
| |
PiperOrigin-RevId: 202401460
|
|
|
|
| |
PiperOrigin-RevId: 202401380
|
|
|
|
| |
PiperOrigin-RevId: 202400843
|
|
|
|
| |
PiperOrigin-RevId: 202399218
|
|
|
|
|
|
|
|
|
|
| |
implementation.
Due to an issue with negative StridedSlice indices in TensorFlow, the end indices can specify degenerate slices when negative indices are used to shrink an axis (e.g. for tf.range(4)[-1], start is -1, end is 0, and stride is 1). This fix works around the issue by ignoring stop indices entirely when an axis is shrinking, since in order to be shrunk the length is by definition 1.
Fixes Issue #19260.
PiperOrigin-RevId: 202398678
|
|
|
|
| |
PiperOrigin-RevId: 202397475
|
|
|
|
|
|
|
|
| |
Change XLA lowering of QuantizeAndDequantizeV2/V3 to match the TF kernel much more closely. The main exception is the min_quantized and max_quantized values are calculated as floats to avoid the need for 64-bit integer math, which is not present on all accelerators.
Reformats unary_ops_test.py in passing, but on the whole I don't mind the reformatting.
PiperOrigin-RevId: 202395114
|
|
|
|
| |
PiperOrigin-RevId: 202393642
|
|
|
|
| |
PiperOrigin-RevId: 202392792
|
|
|
|
|
|
| |
Without ',' does not evaluate to a tuple.
PiperOrigin-RevId: 202390939
|
|
|
|
| |
PiperOrigin-RevId: 202388653
|
|
|
|
|
|
|
| |
corresponding RewriteConfig::Toggle as an argument. This fixes
a read-uninit-var error.
PiperOrigin-RevId: 202385487
|
|
|
|
|
|
| |
Useful when the minibatches don't have the same size.
PiperOrigin-RevId: 202381046
|
|
|
|
|
|
| |
to build ops to use the corresponding free functions in namespace xla:: instead.
PiperOrigin-RevId: 202377457
|
|
|
|
|
|
| |
Using simple keys is more efficient.
PiperOrigin-RevId: 202377039
|
|
|
|
| |
PiperOrigin-RevId: 202371689
|
|
|
|
| |
PiperOrigin-RevId: 202370201
|
|
|
|
|
|
|
|
|
|
|
| |
constant value 0 and already have their shape set from the output
activations shape; instead of having it create dummy placeholders
and relying on PropagateFixedSizes to create the constant array.
Rationale: It wasn't PropagateFixedSizes's job to create constant
arrays, and that broke down in a case where the bias vectors not
being constant prevented FuseBinaryIntoPrecedingAffine from running.
PiperOrigin-RevId: 202365030
|
|
|
|
| |
PiperOrigin-RevId: 202363774
|
|
|
|
| |
PiperOrigin-RevId: 202357498
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
more than one device-to-device copy stream per GPU device.
This is an experimental feature that will have no effect unless
copy operations explicitly request a stream other than 0, which
currently does not occur anywhere in a standard build.
Eventually it may be of benefit in the presence of multiple
bi-directional concurrent data copies.
PiperOrigin-RevId: 202354513
|
|
|
|
|
|
| |
Atrous convolutions are often DepthwiseConv2d operations preceded by SpaceToBatchND and followed by BatchToSpaceND operations. This change makes fold_batch_norms.py and quantize.py support handling this pattern.
PiperOrigin-RevId: 202353838
|
|
|
|
|
|
| |
build XlaOps, rather than calling XlaBuilder methods.
PiperOrigin-RevId: 202348891
|
|
|
|
| |
PiperOrigin-RevId: 202347723
|
|
|
|
|
|
| |
to check for None) while doing graph processing.
PiperOrigin-RevId: 202346371
|