- 12 12月, 2018 40 次提交
-
-
由 Scott Zhu 提交于
The methods in the parent class should work the same way. PiperOrigin-RevId: 225094141
-
由 Martin Wicke 提交于
PiperOrigin-RevId: 225092500
-
由 Artem Belevich 提交于
tensorflow::CudaRoot() now may return multiple possible locations of the CUDA root. PiperOrigin-RevId: 225091635
-
由 Anna R 提交于
PiperOrigin-RevId: 225091038
-
由 Dimitris Vardoulakis 提交于
Now, we optimize any sequence of the form: AR [Bitcast|Transpose|Reshape|Convert|Multiply|Add|Subtract]* CRS PiperOrigin-RevId: 225090998
-
由 Eugene Brevdo 提交于
GetCurrentCPU: returns the current CPU of the calling thread. NumTotalCPUs: attempts to get the total number of physical cores on the system When both return non-failing values, we expect 0 <= GetCurrentCPU < NumTotalCPUs. PiperOrigin-RevId: 225088316
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 225085109
-
由 A. Unique TensorFlower 提交于
Small refactor of `thresholds` default value and validation steps. The number of thresholds is used instead of the user specified type of the `thresholds` kwarg to determine the output of the result method: `thresholds` is a scalar or single element list/tuple -> return scalar `thresholds` is a multi element list/tuple -> return list This is functionally equivalent to the previous code except for cases where the user passes in a single element list for the thresholds kwarg. In the previous code, this would cause the result method to return a list whereas now it returns a scalar. PiperOrigin-RevId: 225079221
-
由 Andrew Selle 提交于
PiperOrigin-RevId: 225074738
-
由 A. Unique TensorFlower 提交于
(1) Combine tensors from multiple replicas into a single tensor; each replica may have its own trace file. (2) Accept two tensor traces and report their difference. (3) Summarize and print the value of a tensor in terms of: (a) full tensor value, (b) partial tensor value, (c) any NaN/Inf, (d) the vector Norm, (e) Max-absolute value across all elements. (4) Various print order: topological, numerical, alphabetical. (5) Many more unit tests. PiperOrigin-RevId: 225072821
-
由 Allen Lavoie 提交于
Missing things like variables, function/argument names, support for arguments that aren't a flat list of Tensors, and many other things. But it does manage to save, restore, and call a function. Starts saving a bit of extra metadata when a new function trace is created. Since this does not have to be computed each time the function is called, I expect the performance impact to be minimal. PiperOrigin-RevId: 225072712
-
由 Yuefeng Zhou 提交于
PiperOrigin-RevId: 225069740
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 225064608
-
由 James Ring 提交于
PiperOrigin-RevId: 225063980
-
由 Anna R 提交于
names set. The former would behave correctly for subclasses. Also, moving get_v1_names|constants and get_v2_names|constants functions to tf_export.py to reduce code duplication. PiperOrigin-RevId: 225063242
-
由 Tim Shen 提交于
PiperOrigin-RevId: 225054204
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 225051897
-
由 Sergei Lebedev 提交于
This caused hard to diagnose failures in enable_eager_execution calls. PiperOrigin-RevId: 225050519
-
由 Alexandre Passos 提交于
PiperOrigin-RevId: 225050185
-
由 Gaurav Jain 提交于
PiperOrigin-RevId: 225049315
-
由 Mihai Maruseac 提交于
PiperOrigin-RevId: 225048682
-
由 Peter Hawkins 提交于
Populate the tuple index table of the return value; the callee cannot do this since it does not know the buffer assignments. Explicitly enable custom_call_test only for cpu in the BUILD file, rather than disabling it on non-CPU backends. These tests would not work on any non-CPU backend. PiperOrigin-RevId: 225048065
-
由 Allen Lavoie 提交于
Having the warning on little-used 1.x utilities isn't super important, but the 2.x utilities shouldn't print them during normal use for sure. PiperOrigin-RevId: 225047956
-
由 Peter Hawkins 提交于
This allows non-TF Python clients to reuse the TensorFlow implementations of these ops (and any future improvements to be shared between users). PiperOrigin-RevId: 225047881
-
由 Billy Lamberta 提交于
PiperOrigin-RevId: 225045442
-
由 Peter Hawkins 提交于
Add a couple of simple C++ tests. PiperOrigin-RevId: 225044584
-
由 Michael Kuperstein 提交于
The proto field is int64, but the class field is int. PiperOrigin-RevId: 225044350
-
由 Michael Case 提交于
PiperOrigin-RevId: 225041387
-
由 Peter Hawkins 提交于
Move loop helpers used by Cholesky decomposition into xla/client/lib/loops.*. PiperOrigin-RevId: 225037112
-
由 Michael Kuperstein 提交于
This doesn't have any benefit in terms of sizeof(HloInstruction), but it's awkward to have a sublcass for DS and not DUS. Also adds an intermediate class in the hierarchy that avoids having to hard-code the index operand's number. PiperOrigin-RevId: 225033893
-
由 Peter Buchlovsky 提交于
PiperOrigin-RevId: 225028975
-
由 Scott Zhu 提交于
Also added unit test to cover that. PiperOrigin-RevId: 225028823
-
由 Mark Heffernan 提交于
Also clean up type error reporting in evaluator. PiperOrigin-RevId: 225028144
-
由 Mihai Maruseac 提交于
- io.PaddingFIFOQueue -> queue.PaddingFIFOQueue - io.PriorityQueue -> queue.PriorityQueue - io.QueueBase -> queue.QueueBase - io.RandomShuffleQueue -> queue.RandomShuffleQueue - FIFOQueue -> queue.FIFOQueue - train.Server -> distribute.Server PiperOrigin-RevId: 225025906
-
由 James Keeling 提交于
Somehow this comment seems to have been repeated. I delete the second copy and re-flow the existing one. PiperOrigin-RevId: 225022682
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 225018765
-
由 Gaurav Jain 提交于
PiperOrigin-RevId: 225017976
-
由 Bixia Zheng 提交于
Convert the implementation of scalar reduction, row reduction and column reduction to use EmitTiledKernel, which is a more general kernel tiling implementation that is based on the information defined by an object of KernelMappingScheme. For scalar reduction and row reduction, the new implementation should generate the similar optimized code as the old implementation. For column reduction, the new implementation is not exactly the same as the old implementation for a few reasons. First, in the old implementation, routine IrEmitterUnnested::EmitColumnReduction uses kTileWidth to control the number of output elements for which each thread computes a partial result and set the value of kTileWidth to 2. The new implementation is equivalent to the old implementation with kTileWidth=1. Supporting kTileWidth=1 in the new implementation will complicate the implementation and our experiment didn't show much benefit of kTileWidth=2. Second, the old implementation tries to maximize the hardware thread blocks. The new implementation currently only uses one hardware thread block to precess one block of tiles because it uses the hardware block ID as the index for the block of tiles and uses the hardware thread ID as the index for the elements within a tile. PiperOrigin-RevId: 225013188
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 225011350
-
由 A. Unique TensorFlower 提交于
Using XRTAllocate to register device memory, a user gets back int64 handles which needs to be explicitly deleted in order to avoid memory leaks. If a client crashes (or has bugs in its handle release logic), a remote TF server will be leaking memory with no possibility of recover. Since clients always run a ConfigureDistributedTPU at boot time, we clear the XRT allocated resource manager container at that time. Also add a new XRTReleaseAllAllocations operation, to clear all the XRT memory on the target host. PiperOrigin-RevId: 225006277
-