1. Only MIN_COMBINED mode is supported;
2. Reshape the output to [d0,..., dn * unpack_size] if input shape is [d0, ..., dn].
3. Only uint32 is supported for the input;
4. Output data type is bfloat16;
5. Only uint8 or uint16 is supported for the original unpacked input.

PiperOrigin-RevId: 225203930

PiperOrigin-RevId: 225202451

This contrain was originally added due to the different weights format issue
between canonical and cudnn (extra input bias). Now since the input bias is
feeded as zeros in cudnn mode, and weights are unified into one format. Having
bias regularizer should not be a issue.

PiperOrigin-RevId: 225193782

PiperOrigin-RevId: 225189182

Removing the LoopCond of a while_loop can cause the partitioner to fail with:
  A cross-device loop must have a pivot predicate

For some reason this only triggers with while_v2 (the lowered while
loop is slightly different than what would be produced by the original
while_loop).

PiperOrigin-RevId: 225188075

In addition, fix a few eval() calls as well as remove some
@test_util.run_v1_only annotations.

PiperOrigin-RevId: 225180248

PiperOrigin-RevId: 225178266

PiperOrigin-RevId: 225140840

Format in the previous state didn't give the timezone.

PiperOrigin-RevId: 225138116

Remove :android_tensorflow_lib_selective_registration* aliases, targets using selective registration can now use the :android_tensorflow_lib_lite* targets.
PiperOrigin-RevId: 225134497

PiperOrigin-RevId: 225131361

PiperOrigin-RevId: 225131218

PiperOrigin-RevId: 225127595

PiperOrigin-RevId: 225125955

No functional change. Rename the proto message Layout to LayoutProto, and Tile to TileProto. Define in-place replacement C++ classes named Layout and Tile with an interface which mirrors the protobuf generated code interface. Having these data structures as C++ classes enables greater flexibility in the interface, enables enforcement of invariants, and potential performance improvements.

PiperOrigin-RevId: 225121052

RELNOTES: Add an ignore_unknown argument to parse_values which suppresses ValueError for unknown hyperparameter types. Such hyperparameter are ignored.

parse_values('a=1,b=foo', {a: int}) Raises a ValueError
parse_values('a=1,b=foo', {a: int}, ignore_unknown=True) does not raise a ValueError, and returns {'a': 1}

PiperOrigin-RevId: 225117666

(1) Able to trace tensors when the model is executed on the CPU.
    (previously, it could only trace when the model is executed on TPU)
(2) Allow the user to specify the op-names and op-types to be excluded or included for tracing via regular expressions.
(3) Two new trace modes: (1) tracing the vector norm of the tensor and (2) tracing the maximum of the absolute values of all elements in the tensor.
(4) Attach the replica-ID to a traced tensor value so that the post-processing tool (Tensor-Inspector) can reconstruct the whole tensor from all replicas.
(5) An API to trace tensors programmatically.
(6) Allow writing the trace to stdout (previously, it must be written to a file).

PiperOrigin-RevId: 225112219

PiperOrigin-RevId: 225110993

PiperOrigin-RevId: 225110815

PiperOrigin-RevId: 225110544

PiperOrigin-RevId: 225109110

PiperOrigin-RevId: 225107801

PiperOrigin-RevId: 225107483

PiperOrigin-RevId: 225107457

PiperOrigin-RevId: 225102983

PiperOrigin-RevId: 225102856

PiperOrigin-RevId: 225102563

PiperOrigin-RevId: 225099426

Under heavy network load, the GRPC watchdog timer can be "stuck" behind other pending RPCs and timeout actively running workers.  Disable this for now.

PiperOrigin-RevId: 225098378

PiperOrigin-RevId: 225098257

PiperOrigin-RevId: 225097826

Caused by an implicit boolean check which should have been an explicit None check

PiperOrigin-RevId: 225096833

PiperOrigin-RevId: 225096446

Small refactor to improve the readability of the Model class for those who use the code as documentation.
General idea: most important methods come first, private utilities are moved to the bottom of the class.
Also use a single method for `_standardize_user_data` (previously split into 2 methods that did not reflect
two separate sets of actions).

PiperOrigin-RevId: 225094903

PiperOrigin-RevId: 225094534

The methods in the parent class should work the same way.

PiperOrigin-RevId: 225094141

PiperOrigin-RevId: 225092500

tensorflow::CudaRoot() now may return multiple possible locations of the CUDA root.
PiperOrigin-RevId: 225091635

PiperOrigin-RevId: 225091038

Now, we optimize any sequence of the form:
AR [Bitcast|Transpose|Reshape|Convert|Multiply|Add|Subtract]* CRS

PiperOrigin-RevId: 225090998