| Commit message (Collapse) | Author | Age |
... | |
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
A plugin name shouldn't contain slashes or be empty, as these would lead
to confusing routes: a plugin with name "eggs/ham" and route "/data"
would handle the route "/data/plugin/eggs/ham/data", but this route
could also be handled by the plugin "eggs" along route "/ham/data".
Furthermore, it could also be handled by the plugin with name "" along
route "eggs/ham/data", or the plugin with name "eggs/h" along route
"am/data". None of these is desirable, so I'm forbidding
- plugins with characters not in [A-Za-z0-9_.-], and
- routes without a leading slash.
PiperOrigin-RevId: 157007010
|
|
|
|
| |
PiperOrigin-RevId: 157006018
|
|
|
|
| |
PiperOrigin-RevId: 157005668
|
|
|
|
|
|
| |
that throws TypeError for tensors.
PiperOrigin-RevId: 157000504
|
|
|
|
| |
PiperOrigin-RevId: 156990440
|
|
|
|
|
|
|
| |
This change demonstrates my proposed "best practice" for having a documentation
page, demo page, and a test package.
PiperOrigin-RevId: 156945855
|
|
|
|
|
|
|
|
| |
To avoid test timeout in OSS builds such as:
http://ci.tensorflow.org/view/Tensorflow%20Jenkins%20Monitored%20builds/job/nightly-matrix-cpu/TF_BUILD_IS_OPT=NO_OPT,TF_BUILD_IS_PIP=NO_PIP,TF_BUILD_PYTHON_VERSION=PYTHON3,label=mac-slave/494/console
PiperOrigin-RevId: 156945796
|
|
|
|
| |
PiperOrigin-RevId: 156944397
|
|
|
|
|
|
| |
location. The problem was that the new translateBy method introduced by d3 v4 applied scaling before translation. This change manually counters that with a division.
PiperOrigin-RevId: 156943470
|
|
|
|
| |
PiperOrigin-RevId: 156939644
|
|
|
|
| |
PiperOrigin-RevId: 156939364
|
|
|
|
| |
PiperOrigin-RevId: 156939333
|
|
|
|
| |
PiperOrigin-RevId: 156938197
|
|
|
|
|
|
|
| |
1. Handle the case when node_name in RunMetadata is suffixed with :op_type.
2. Properly distinguish cpu/gpu time
PiperOrigin-RevId: 156937421
|
|
|
|
| |
PiperOrigin-RevId: 156936949
|
|
|
|
|
|
| |
This will be useful if we want to run TensorBoard through the Closure Compiler.
PiperOrigin-RevId: 156936819
|
|
|
|
|
|
|
| |
number of non-trivial reshape/transpose operands is non-zero. (Otherwise there's
no benefit to sink Reshape/Transpose).
PiperOrigin-RevId: 156936778
|
|
|
|
|
|
|
|
| |
Before, Graph._add_function() call was used to import function library from
GraphDef. This private function lacks hash value comparison and importing
the inner function library used by the function gets improted.
PiperOrigin-RevId: 156933480
|
|
|
|
| |
PiperOrigin-RevId: 156933321
|
|
|
|
| |
PiperOrigin-RevId: 156927675
|
|
|
|
|
|
| |
numbers, and drops some unnecessary tests.
PiperOrigin-RevId: 156925729
|
|
|
|
|
|
| |
This makes the ordering more reverse topological.
PiperOrigin-RevId: 156924652
|
|
|
|
| |
PiperOrigin-RevId: 156923445
|
|
|
|
| |
PiperOrigin-RevId: 156919832
|
|
|
|
| |
PiperOrigin-RevId: 156917877
|
|
|
|
| |
PiperOrigin-RevId: 156917717
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
gumbel noise. -log(-log(U)) gives infinity if U draws a hard 0. Adds a tiny
offset to U (2e-30) to avoid log(U) = -inf.
The CPU sampling algorithm depends on the order of the logits which is
undesirable and can also oversample the first logit if it is smaller than the
smallest random float larger than 0 (~1e-7). Switching to double precision
internally mitigates these problems, although it doesn't fix them. Slowdown
is ~35% in the worst case.
Also adds various tests that we would like the sampling to pass.
CPU Benchmark before:
32 10000 1 0.060 0.069 0.87
32 10000 4 0.229 0.074 3.10
32 10000 32 2.180 0.059 37.09
32 100000 1 0.430 0.480 0.90
32 100000 4 2.322 0.449 5.17
32 100000 32 31.508 0.471 66.96
128 10000 1 0.168 0.235 0.71
128 10000 4 0.965 0.246 3.93
128 10000 32 7.989 0.225 35.51
128 100000 1 1.681 1.539 1.09
128 100000 4 9.012 1.57 35.73
128 100000 32 126.222 1.626 77.60
CPU Benchmark after:
32 10000 1 0.054 0.112 0.48
32 10000 4 0.206 0.093 2.21
32 10000 32 1.826 0.091 20.12
32 100000 1 0.292 0.636 0.46
32 100000 4 2.086 0.606 3.44
32 100000 32 28.496 0.633 45.03
128 10000 1 0.125 0.266 0.47
128 10000 4 0.759 0.258 2.94
128 10000 32 7.362 0.254 29.03
128 100000 1 1.550 2.18 10.71
128 100000 4 8.712 2.22 23.92
128 100000 32 122.585 2.213 55.39
PiperOrigin-RevId: 156917266
|
|
|
|
|
|
|
| |
defined in the header, instead of strcmp. strcmp is not a constexpr on all
platforms.
PiperOrigin-RevId: 156914421
|
|
|
|
| |
PiperOrigin-RevId: 156913869
|
|
|
|
| |
PiperOrigin-RevId: 156909830
|
|
|
|
| |
PiperOrigin-RevId: 156908714
|
|
|
|
| |
PiperOrigin-RevId: 156908010
|
|
|
|
| |
PiperOrigin-RevId: 156907935
|
|
|
|
| |
PiperOrigin-RevId: 156907723
|
|
|
|
| |
PiperOrigin-RevId: 156906671
|
|
|
|
| |
PiperOrigin-RevId: 156901957
|
|
|
|
| |
PiperOrigin-RevId: 156895715
|
|
|
|
| |
PiperOrigin-RevId: 156895551
|
|
|
|
| |
PiperOrigin-RevId: 156879382
|
|
|
|
| |
PiperOrigin-RevId: 156876795
|
|
|
|
| |
PiperOrigin-RevId: 156875936
|
|
|
|
| |
PiperOrigin-RevId: 156870745
|
|
|
|
| |
PiperOrigin-RevId: 156870702
|
|
|
|
| |
PiperOrigin-RevId: 156870444
|
|
|
|
| |
PiperOrigin-RevId: 156869789
|
|
|
|
|
|
| |
Tested by enabling the fast path by default and ensuring unit tests still pass.
PiperOrigin-RevId: 156866900
|
|
|
|
| |
PiperOrigin-RevId: 156843392
|
|
|
|
|
|
|
|
| |
This completes the sequence by moving ExecuteOnStreamWrapper from xla::Service onto xla::Executable, where it conceptually belongs. The fact that it belongs on Executable is illustrated by the fact that it doesn't access any elements of the Service object, but (in the original form) it takes an Executable object as its first parameter. We can trivially replace this first parameter with an implicit "this".
This also removes the (now-unused) "backend" parameter from ExecuteOnStreamWrapper, and moves DumpExecutedHlo from xla::Service to xla::Executable. As with ExecuteOnStreamWrapper, DumpExecutedHlo has no conceptual ties to xla::Service, and the name suggests a clear fit on xla::Executable.
PiperOrigin-RevId: 156834459
|
|
|
|
| |
PiperOrigin-RevId: 156827000
|
|
|
|
|
|
|
|
| |
This follows patch 1 by removing the ShapeSizeBytes function from xla::Compiler, and changing its callers to use Executable::ShapeSizeBytesFunction.
Many of the callers are reproducing the same pattern to make a lambda to computes the size of a buffer, and so we condense this into xla::Compiler::BufferSizeBytesFunction.
PiperOrigin-RevId: 156814087
|