646570256 by A. Unique TensorFlower<gardener@tensorflow.org>:
Put back type aliases for some 3p projects until they're migrated off of xla::Status.
--
646567067 by A. Unique TensorFlower<gardener@tensorflow.org>:
[xla:cpu] Optimize KernelThunk alignment checks
--
646562233 by A. Unique TensorFlower<gardener@tensorflow.org>:
Automated rollback of changelist 609005660.
646560125 by A. Unique TensorFlower<gardener@tensorflow.org>:
[XLA:CollectivePipeliner] Add more execution tests (using the HLOs in collective_pipeliner_test.cc).
--
646554714 by A. Unique TensorFlower<gardener@tensorflow.org>:
Instead of copybara rules, use `if_google` to remove extra proto deps
Followup will do the same for TSL
--
646551061 by A. Unique TensorFlower<gardener@tensorflow.org>:
Remove `Array::Reshard`
This CL removes the deprecated `Array::Reshard` API. All existing users have been manually migrated to use `Client::CopyArrays`.
IFRT Proxy is updated such that the client no longer issues `Array::Reshard` and the server emulates the reshard behavior by using `Client::CopyArrays`. Since this does not actually change the wire format, we do not need to update the version number. Once the reshard API passes the compatibility window, we can remove its proto message and handler altogether.
--
646545951 by A. Unique TensorFlower<gardener@tensorflow.org>:
Add license header to `dependabot.yml`
--
646545541 by A. Unique TensorFlower<gardener@tensorflow.org>:
Remove force_synchronous attribute from ParallelMap op in map_parallelization optimizer.
The code reuses the attributes/inputs of the original Map op but just changes it to a ParallelMap op. But the force_synchronous attribute is not supported in ParallelMap and causes log warnings.
The issue was introduced in cl/642418430
--
646534280 by A. Unique TensorFlower<gardener@tensorflow.org>:
Use absl::StatusOr instead of xla::StatusOr.
--
646517068 by A. Unique TensorFlower<gardener@tensorflow.org>:
Add more pattern to HloUnstacker pass + some refactoring.
Added a support for handling slicing fusion pattern:
fusion(stacked_operand, loop_iteration_var), calls=fusion_computation
fusion_computation {
p0 = parameter(0)
p1 = parameter(1)
slice = dynamic_slice(p0, p1, zero, ...)
ROOT bitcast = bitcast(slice)
}
Add "xla_enable_hlo_unstacker" flag to the compiler.
--
646513305 by A. Unique TensorFlower<gardener@tensorflow.org>:
Remove unused deps.
--
646513101 by A. Unique TensorFlower<gardener@tensorflow.org>:
[xla:cpu] Add a fast path for executing thunks sequentially
--
646507520 by A. Unique TensorFlower<gardener@tensorflow.org>:
Added a fingerprint field to PjRtStreamExecutorLoadedExecutable to avoid recalculating fingerprints when FingerprintExecutable() is called. This change significantly reduces idle time before execution when the GPU load tracker enqueues an executable.
--
646505763 by A. Unique TensorFlower<gardener@tensorflow.org>:
Change visibility rules.
--
646505592 by A. Unique TensorFlower<gardener@tensorflow.org>:
[XLA:GPU] Parse block-level parameters from backend config when available.
If block-level parameters are not available, fall back to the SoftMax heuristic.
The original plan was to parse block-level parameters from the config and remove the heuristic, but it turned out that we don't support all "valid" tiling. With this change it will be easier to write tests and verify that we don't have problem, before we could remove the heuristic and fully migrate to fusion backend config.
Also fix strides in ir_emitter_triton.cc. This was not a problem before, because SoftMax heuristic only produces tiles that are contiguous in memory.
--
646505352 by A. Unique TensorFlower<gardener@tensorflow.org>:
[xla:cpu] Add dynamic-update-slice fusion optimization to IrEmitter2
+ enable select-and-scatter test that used to time out without DUS optimization
--
646504512 by A. Unique TensorFlower<gardener@tensorflow.org>:
PR #62472: Hash Pin docker images
Imported from GitHub PR https://github.com/tensorflow/tensorflow/pull/62472
Also related to https://github.com/tensorflow/tensorflow/pull/62471, would you consider hash pin the docker images?
The security benefit of doing so is that it mitigates the risk of typosquatting attacks since the images are public. If there is a need for them to be updated regularly, I can also submit a .github/dependabot file to update the docker images regularly (weekly or monthly for example).
Besides, AFAIUC, the dockerfiles are used for build and tests, which lead to another benefit of hash pinning: reliability and stability.
Let me know your thoughts about i.
Thanks!
Copybara import of the project:
--
8f4589fe583518d3099c98215e5e6bf3858fa24e by Joyce Brum <joycebrum@google.com>:
feat: create dependabot
Signed-off-by: Joyce Brum <joycebrum@google.com>
Merging this change closes#62472
--
PiperOrigin-RevId: 646570256
This folder is still under construction. It is part of an ongoing effort to improve the structure of CI and build related files within the TensorFlow repo.
PiperOrigin-RevId: 538283635
