| Commit message (Collapse) | Author | Age |
|
|
|
| |
PiperOrigin-RevId: 198762795
|
|
|
|
|
|
| |
configurable. By default they will not be scoped.
PiperOrigin-RevId: 198759754
|
|
|
|
| |
PiperOrigin-RevId: 198755870
|
|
|
|
| |
PiperOrigin-RevId: 198754463
|
|
|
|
| |
PiperOrigin-RevId: 198743117
|
|
|
|
|
|
| |
This introduces an abstraction for a dispatch context, which allows passing local information through to the specialized operators.
PiperOrigin-RevId: 198742074
|
|
|
|
|
|
| |
with the field name.
PiperOrigin-RevId: 198739727
|
|
|
|
| |
PiperOrigin-RevId: 198739017
|
|
|
|
| |
PiperOrigin-RevId: 198734229
|
|
|
|
|
|
| |
facilities for tracking local state and reindenting node blocks. Rearrange the error handling in base transformer to avoid chained exceptions.
PiperOrigin-RevId: 198727946
|
|
|
|
| |
PiperOrigin-RevId: 198725578
|
|
|
|
|
|
|
|
|
| |
While technically the order of the created nodes doesn't matter, this
makes viewing and debugging graphs more sensible.
Fixes #19594.
PiperOrigin-RevId: 198721173
|
|
|
|
| |
PiperOrigin-RevId: 198715407
|
|
|
|
| |
PiperOrigin-RevId: 198710452
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
- Computation
- ComputeConstant
- Execute
- ExecuteAsync
- ExecuteParallel
- GetComputationStats
- GetComputationShape
- GetLocalShape
- IsConstant
- LoadComputationSnapshot
- Op
- SetReturnValue
- SnapshotComputation
PiperOrigin-RevId: 198669035
|
|
|
|
|
|
|
|
|
|
| |
For example, with this CL, ArithmeticOptimizer can optimize the Reshape below
into a no-op.
s = Shape(t)
Reshape(t, Concat(s[0], s[1], s[2], s[3]))
PiperOrigin-RevId: 198668726
|
|
|
|
| |
PiperOrigin-RevId: 198658074
|
|
|
|
|
|
|
|
|
|
| |
Makes the Variable.trainable property public, which is sensible if we're discouraging use of the global collection (currently eager execution is using ResourceVariable._trainable in a bunch of places anyway). I'm leaving it read-only for now, since we should toggle in and out of the global collection when it changes.
Same change for checkpointable data structures with respect to gathering extra variables. They'll behave like subclassed Models.
I think this makes more sense than trying to have a distinction between "variables" and "weights". It's also more sensible than collecting everything that would get checkpointed, since that will include Optimizer slot variables and metrics. Collecting those is generally pointless, and accidentally adding them to gradient tapes would be horribly confusing.
PiperOrigin-RevId: 198656079
|
|
|
|
| |
PiperOrigin-RevId: 198654780
|
|
|
|
| |
PiperOrigin-RevId: 198654099
|
|
|
|
|
|
|
| |
functions. That should be a programmatic error and we have full control of
internal functions, so it is OK to crash if error happens.
PiperOrigin-RevId: 198651749
|
|
|
|
|
|
|
| |
Also modify relevant ToString functions so we can have the property
Parse(ToString(x)) == x.
PiperOrigin-RevId: 198650340
|
|
|
|
| |
PiperOrigin-RevId: 198649827
|
|
|
|
| |
PiperOrigin-RevId: 198648273
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This lets replay_computation build an executable once and run it
multiple times. This is particularly important because in XLA:GPU, the
first run of an executable does some autotuning and therefore is
unrepresentative.
This change removes --xla_hlo_profile_last_run, because I don't see how
to support it in LocalClient -- LocalClient wants the do-profile bit to
be set when we *compile*. (There may not be an easy fix for this; it
worked with regular Client because we were recompiling every time we
ran.)
PiperOrigin-RevId: 198643577
|
|
|
|
| |
PiperOrigin-RevId: 198642698
|
|
|
|
|
|
| |
assumption that if a real GCS object has file name ending with "/", it is always a directory mark rather than an object carrying actual contents.
PiperOrigin-RevId: 198640604
|
|
|
|
|
|
|
| |
Right now the old while loop can stick around if it had side effects, which is
incorrect.
PiperOrigin-RevId: 198639203
|
|
|
|
|
|
|
|
|
|
|
|
| |
AdaptiveSharedBatchScheduler, but increased reliablility and stability.
ASBS assumes request latency can be minimized at a specific number of batch processing threads. Under reasonable load, this is true and ASBS performs well, but under low load latency is basically unaffected by the number of threads, and ASBS can learn a wide variety of 'optimal' values. If load resumes suddenly, these values can give very poor latencies. In most cases, ASBS will recover, eventually rediscovering the correct value, but we have observed other cases where the latency is so large and noisy that ASBS can't get a good signal to guide its learning and the number of threads remains stuck at the bad value.
In addition, the incremental learning nature of this algorithm means that ASBS is always exploring to some extent, which can give rise to periods of non-optimal latency. This is most significant at high utilization where the wrong number of threads can potentially overload the system.
ASBS uses latency as a proxy for keeping the tensorflow processing pipeline optimally loaded. SDBS, on the other hand, uses a direct measurement of the pipeline fullness, and adjusts its number of batch processing threads accordingly. This solves the exploration problem. SDBS solves the low load problem by not adjusting its thread count when the threads pass some idleness threshold.
PiperOrigin-RevId: 198638918
|
|
|
|
|
|
| |
that takes an Estimator input_fn and returns an input receiver function.
PiperOrigin-RevId: 198638593
|
|
|
|
| |
PiperOrigin-RevId: 198637528
|
|
|
|
| |
PiperOrigin-RevId: 198634986
|
|
|
|
|
|
| |
This is staging for removing the _USE_C_API toggle altogether.
PiperOrigin-RevId: 198634886
|
|
|
|
|
|
| |
This disallows some undefined behavior with unordered reads and writes.
PiperOrigin-RevId: 198633444
|
|
|
|
| |
PiperOrigin-RevId: 198631733
|
|
|
|
| |
PiperOrigin-RevId: 198629366
|
|
|
|
| |
PiperOrigin-RevId: 198625799
|
|
|
|
| |
PiperOrigin-RevId: 198624394
|
|
|
|
| |
PiperOrigin-RevId: 198624285
|
|
|
|
|
|
|
|
|
|
|
| |
such that:
# Lower triangular matrix
x = tf.matrix_band_part(x, -1, 0)
x == fill_triangular(fill_triangular_inverse(x))
Code by srvasude@ which I'm submitting on his behalf.
PiperOrigin-RevId: 198623887
|
|
|
|
| |
PiperOrigin-RevId: 198622727
|
|
|
|
|
|
| |
inverts a lower-triangular matrix.
PiperOrigin-RevId: 198622553
|
|
|
|
| |
PiperOrigin-RevId: 198620715
|
|
|
|
|
|
|
|
| |
error happened.
Currently error can happen if function can't be instantiated as GrapplerFunctionItem.
PiperOrigin-RevId: 198595096
|
|
|
|
| |
PiperOrigin-RevId: 198592202
|
|
|
|
| |
PiperOrigin-RevId: 198588920
|
|
|
|
| |
PiperOrigin-RevId: 198585886
|
|
|
|
| |
PiperOrigin-RevId: 198582954
|
|
|
|
| |
PiperOrigin-RevId: 198582181
|
|
|
|
| |
PiperOrigin-RevId: 198573236
|