pull #107
Merged
pull #107
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This fix tries to address the issue raised in 22793 where an invalid axis (outside of `[-dim, dim)`) still returns value. This behavior is different from most other ops in tf like `tf.argmax`/etc. This fix add the validation of axis so that an error will be returned in case of invalid axis. This fix fixes 22793. Signed-off-by: Yong Tang <yong.tang.github@outlook.com>
Signed-off-by: Yong Tang <yong.tang.github@outlook.com>
Signed-off-by: Yong Tang <yong.tang.github@outlook.com>
Signed-off-by: Yong Tang <yong.tang.github@outlook.com>
Signed-off-by: Yong Tang <yong.tang.github@outlook.com>
My recent change modified how `Executor::Args` was populated and missed the fact that we rewrite the `Options` to add in a created `Rendezvous` object in some cases. This change correctly uses the rewritten `Options` in both cases. PiperOrigin-RevId: 220218606
Fixes #23059 Also adds "Warning:" to the deprecation warning, which is rendered in many doc tools as a red callout box. PiperOrigin-RevId: 220228838
PiperOrigin-RevId: 220229772
…ich takes `tf.data.experimental.StatsOptions` object. `StatsOptions` can configure options for collecting `dataset` stats using `StatsAggregator`, and it has aggregator as an argument which attaches the given aggregator to the dataset. (this will also replace `set_stats_aggregator()` dataset transformation.) PiperOrigin-RevId: 220230269
With this change, tensorflow/python/summary/summary.py contains all the existing python op definitions. Tests remain split with summary_test.py in the same directory testing the API and separate (newly v1-named) tests in tensorflow/python/kernel_tests to test the serialized proto formats. PiperOrigin-RevId: 220230494
PiperOrigin-RevId: 220231014
* absl_int128 is referenced by str_format_internal
PiperOrigin-RevId: 220239919
We now parallelize not only the loop through the dimension that should be sorted, but also the other loops. Also combine several compare loops if the comparisons happen within a small block of memory. PiperOrigin-RevId: 220245099
PiperOrigin-RevId: 220260494
… self. Use this mechanism in defun to convey that reference to AutoGraph. PiperOrigin-RevId: 220260542
Copying the sharding from one instruction to an other can create invalid sharding information for tiled or tuple sharded instructions. We can just ignore the sharding for this case and rely on the domains to preserve the data instead. PiperOrigin-RevId: 220272535
…memory. [tf.data] Optimize the creation of tensors in `tf.data.Dataset.range()`. This change improves the range benchmark from 148.7 ns/element to 122.4 ns/element. PiperOrigin-RevId: 220279090
…rcular dependencies. Override the self argument with the results of inspect_utils.getmethodclass for consistency. Clean up the module cleanup mechanism. PiperOrigin-RevId: 220280056
PiperOrigin-RevId: 220292732
PiperOrigin-RevId: 220292905
changes to model_analyser.analyse(...): - Swapped context.in_eager_execution() to the currently supported context.executing_eagerly(). - Added negation to eager check. In all likelihood, the negation was always supposed to be there since getting default graph in eager mode does not make sense. The current `if` condition is likely a bug. The proposed fix is consistent with other functions in this module, e.g., `profile(...)`, line 339.
…stop()`. PiperOrigin-RevId: 220296822
…ier to SUM_OVER_BATCH_SIZE for TF V2. It's SUM in V1. PiperOrigin-RevId: 220346638
PiperOrigin-RevId: 220346878
PiperOrigin-RevId: 220348163
PiperOrigin-RevId: 220349112
…achinery. PiperOrigin-RevId: 220351155
- Add DeviceOp (e.g. GPU stream) logging to distributed execution. This is on a best-effort basis. If device tracing is Unavailable, continue without it. - Add timeline logging of RecvBuf transfers over gRPC. - Add bandwidth consumed measurement to click tile for RecvTensor and RecvBuf. PiperOrigin-RevId: 220352522
…h bazel file name. #22167 PiperOrigin-RevId: 220353261
PiperOrigin-RevId: 220356488
PiperOrigin-RevId: 220356963
PiperOrigin-RevId: 220359861
PiperOrigin-RevId: 220368991
PiperOrigin-RevId: 220371736
PiperOrigin-RevId: 220372969
…in most cases rather than parsing an OpKernelContext. Build the compilation cache signature from the XlaCompiler::Arguments, rather than building it inside the cache. Add a helper method XlaComputationLaunchContext::BuildXlaCompilerArguments that builds the necessary XlaCompiler::Arguments. Fix a bug in XlaCompilationCache::Signature::operator==, where only the contents of a tensor were compared for equality, not the shapes. Remove code that reasons about OpKernelContexts in XlaCompiler::CompileSingleOp, instead passing the specific information required. PiperOrigin-RevId: 220373606
… rather than unconditionally registering one. Also set a GpuDeviceContext, since the commented-out code works fine. PiperOrigin-RevId: 220373651
PiperOrigin-RevId: 220378109
PiperOrigin-RevId: 220378263
…logical_or_and PiperOrigin-RevId: 220378489
…_profile. Previously we'd say that the "cost" of an instruction with subcomputations (e.g. a while loop) was equal to the total cost of its instructions. Fine (I mean, not actually fine, a loop can run multiple times, but at least it's something), but we also profile *the inside* of a subcomputation, and we weren't saving any information about *those* instructions. PiperOrigin-RevId: 220381684
This is a limitation of how we use deadness analysis. The right fix is to change how we use deadness analysis, but this CL prevents miscompiles in the meantime. PiperOrigin-RevId: 220383866
… above optimum" total. In other words, for the purposes of calculating totals, let the "true optimum" be the smaller of the estimated optimum and the actual runtime. PiperOrigin-RevId: 220383899
PiperOrigin-RevId: 220385991
PiperOrigin-RevId: 220387985
CloneWithReplacements assumed that the `extras` instructions came before all other instructions in the postorder traversal. This is not true in general. In the course of fixing this, I ended up removing the `extras` parameter entirely, which makes the function easier to call. I also augmented the interface a bit to make CloneWithReplacements easier to call when you just have one or two replacements to make. I will use these overloads in a future patch. PiperOrigin-RevId: 220390779
…hat XLA can know maximum size to allocate for the list at compile time. PiperOrigin-RevId: 220396133
PiperOrigin-RevId: 220398192
PiperOrigin-RevId: 220399204
PiperOrigin-RevId: 220400279
PiperOrigin-RevId: 220421900
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
No description provided.