[pull] master from tensorflow:master #338
Open
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
![pull[bot]](https://avatars.githubusercontent.com/in/12910?s=60&v=4){kind=small}
…s BF16 arguments and algorithm=BF16_BF16_F32 We have a wide range of algorithms for dot and the majority of them require F32 arguments. But the BF16_BF16_F32 one actually could accept BF16 arguments because it does not affect the final precision of the dot. Lets relax the check. PiperOrigin-RevId: 717795772
For ProducerConsumer fusions, we can have the case that a multi-output fusion is created even if the producer fusion was not a multi-output fusion. This happens if the fusion root of the producer is also used outside of the created ProducerConsumer fusion, and we don't want to duplicate the producer. This change adds the possibility to create a HloFusionAdaptor for a ProducerConsumer fusion with extra outputs created for producer roots that are used outside the ProducerConsumer fusion. PiperOrigin-RevId: 717796118
Copies are considered unary elementwise ops. We need to make sure we don't try to move a copy over a copy, otherwise the MoveCopyToUsers pass will not converge to a fixed point. PiperOrigin-RevId: 717824897
Follow up to '[XLA:GPU] Add RewritePatterns for binary elementwise ops in SimplifyAffinePass.'. Fixed pass not finding the parent module (as it was a module op to begin with) and defaulting to i64. PiperOrigin-RevId: 717825878
Updates LLVM usage to match [e2402615a5a7](llvm/llvm-project@e2402615a5a7) PiperOrigin-RevId: 717835256
PiperOrigin-RevId: 717836550
We don't yet support scalable NCCL communicator initialization and should not be generating more than 1 unique id PiperOrigin-RevId: 717840923
Added a benchmark for popping a task from a queue ----------------------------------------------------- Benchmark Time CPU Iterations ----------------------------------------------------- BM_PopTask 2.63 ns 2.63 ns 265317289 PiperOrigin-RevId: 717846665
PiperOrigin-RevId: 717856803
PiperOrigin-RevId: 717858442
We always use intra-op device to run XLA:CPU kernels, stop pretending that we might have some other option (until we have a real alternative). PiperOrigin-RevId: 717858535
A recent change to pass order changed the HLO and check started to fail. Remove the check completely as we don't really care which dimensions are used but whether non-major batch works at all. PiperOrigin-RevId: 717871376
…mits/632bfc342d3a7d63ce8b21209355139ee070d392) Reverts 1893c10 PiperOrigin-RevId: 717875236
…xla.proto`. PiperOrigin-RevId: 717887032
This paves the way for sharing these passes with CPU. Note that lower_tensors requires more work before CPU tests can pass. PiperOrigin-RevId: 717896623
PiperOrigin-RevId: 717901856
Still missing direct atomics (vs. cmpxchg), but we can add those later. Note that the added test for non-gep loads would be an error in non-CPU, hence the separate test file for CPU. PiperOrigin-RevId: 717914337
PiperOrigin-RevId: 717939551
The `cilent_polling_for_error_` variable previously claimed the following: > Once set to true, the value will never change back to false, so no mutex is > needed. This is not true. Concurrent access to the boolean is a data race and must be avoided with a mutex. Thankfully, the code was already protecting access to the variable with a mutex. This CL adds some thread annotations to double check this. PiperOrigin-RevId: 717957732
Embedding pipelining requires a least two steps. However, because the inclusion of summary ops is expensive and only needed periodically, most users run a single training step when enabling summaries. This change detects when summaries are active and automatically disables pipelining (under the assumption that the user will only be running a single step). PiperOrigin-RevId: 717961226
PiperOrigin-RevId: 717963233
PiperOrigin-RevId: 717971669
There are much fewer of these than expected! PiperOrigin-RevId: 717978053
PiperOrigin-RevId: 717991433
… custom suffix (e.g. 2.19.0-rc1). The old implementation didn't delete `-` symbol and expected the wheel filename to be `tensorflow-2.19.0.rc1-cp310-cp310-linux_x86_64.whl` instead of `tensorflow-2.19.0rc1-cp310-cp310-linux_x86_64.whl`. PiperOrigin-RevId: 718020442
PiperOrigin-RevId: 718021261
…rflow:tf_quantization_passes from //third_party/tensorflow/compiler/mlir:passes PiperOrigin-RevId: 718028638
PiperOrigin-RevId: 718030758
PiperOrigin-RevId: 718035528
PiperOrigin-RevId: 719376107
PiperOrigin-RevId: 719389902
…ocate. All callers are migrated to StreamExecutor::CreateMemoryAllocator(MemoryType::kUnified). PiperOrigin-RevId: 719391711
This is now no longer used by JAX, and never matter on xla_extension.so because we no longer build it in a CUDA-specific configuration, instead using CUDA plugins. PiperOrigin-RevId: 719399140
- Add a C++ wrapper type for easier management. - Make the compile options mandatory in `CompiledModel::Create`. This aligns the default value for hardware accelerator selection (depending on how the options were specified, you would either get `None` or `Cpu`). PiperOrigin-RevId: 719402464
PiperOrigin-RevId: 719410517
Reported in: jax-ml/jax#26062 PiperOrigin-RevId: 719416696
…_all PiperOrigin-RevId: 719422910
PiperOrigin-RevId: 719437823
…ilable) PiperOrigin-RevId: 719441192
This was considered legacy under the XLA Runtime transition, but XLA Runtime is no more, so this isn't legacy any more. While at it, remove a stale declaration of an XlaRuntime method whose definition was removed long ago. PiperOrigin-RevId: 719451829
Updating: - `env.h` - `env_time.h` - `errors.h` - `file_statistics.h` - `file_system.h` - `file_system_helper.h` - `logging.h` - `macros.h` - `status.h` - `status_matchers.h` - `status_to_from_proto.h` - `statusor.h` - `test.h` - `test_benchmark.h` - `threadpool.h` - `threadpool_async_executor.h` - `threadpool_interface.h` - `threadpool_options.h` - `types.h` and associated targets. PiperOrigin-RevId: 719457663
PiperOrigin-RevId: 719465588
PiperOrigin-RevId: 719477321
PiperOrigin-RevId: 719525655
PiperOrigin-RevId: 719527644
… within the internal model api. PiperOrigin-RevId: 719535624
PiperOrigin-RevId: 719553967
PiperOrigin-RevId: 719560240
PiperOrigin-RevId: 719567198
PiperOrigin-RevId: 719569599
PiperOrigin-RevId: 719570890
PiperOrigin-RevId: 719581836
PiperOrigin-RevId: 719581903
PiperOrigin-RevId: 719608871
PiperOrigin-RevId: 719659554
PiperOrigin-RevId: 719659988
PiperOrigin-RevId: 719660012
PiperOrigin-RevId: 719676659
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.
See Commits and Changes for more details.
Created by
pull[bot] (v2.0.0-alpha.1)
Can you help keep this open source service alive? 💖 Please sponsor : )