215 Pull requests merged by 6 people

• [tf] Use template parameter to pass async callback

  #97440 merged Jul 23, 2025

• Migrate uses of XLA_TEST_BACKEND macros to use utilities in xla_test_backend_predicates.h

  #97306 merged Jul 23, 2025

• PR #28540: [XLA:GPU] Add support for pre-padded scales for block scaled dot custom call

  #97347 merged Jul 23, 2025

• Parse and ignore 'mode' attribute on collectives.

  #97392 merged Jul 23, 2025

• Remove a redundant condition in mlir_hlo_to_hlo.cc.

  #97405 merged Jul 23, 2025

• Update keras-nightly and typing-extensions versions in the lock files.

  #97445 merged Jul 23, 2025

• [xla:cpu] Make dot thunk support int8 x int8 -> int32 matmul.

  #97435 merged Jul 23, 2025

• Avoid registering definition events until after the allocation succeeds

  #97418 merged Jul 23, 2025

• Add Hermetic C++ Toolchains for Linux x86_64 builds.

  #96820 merged Jul 23, 2025

• Unbreak HloPassPipeline silent/noop change debugging.

  #97431 merged Jul 23, 2025

• [xla] Optimize ShapeTree construction time for arrays

  #97394 merged Jul 23, 2025

• Correct the required gpu to test_core_h100

  #97352 merged Jul 23, 2025

• Use mdformat on the XNNPack delegate readme.

  #97182 merged Jul 23, 2025

• This is an automatic update to a device compatibility allowlist.

  #97423 merged Jul 23, 2025

• [tfcompile] Prefix xla generated symbols with tfcompile_xla_generated.

  #97427 merged Jul 23, 2025

• [XLA] Add stack trace breakdown to HloLiveRange::ToString for peak memory usage

  #94954 merged Jul 23, 2025

• [XLA:GPU] Do not fail when profiling executables, instead ignore failing configurations.

  #97344 merged Jul 23, 2025

• Reverts e4e87da4c3ed8a529f67bb1e3485101c37a4e1f9

  #97424 merged Jul 23, 2025

• Guard against possible null pointer dereferences.

  #97426 merged Jul 23, 2025

• [XLA:Autotuner] Allow gpu_profiler to collect output buffers.

  #97355 merged Jul 23, 2025

• Automated Code Change

  #97401 merged Jul 23, 2025

• [XLA:GPU][host offloading] Implement host offloading thunks.

  #97059 merged Jul 23, 2025

• PR #28735: [XLA:GPU] Enabling cuda graph concurrent mode by default

  #97045 merged Jul 23, 2025

• [XLA:GPU] Early return if there are no elements to compare with.

  #97402 merged Jul 23, 2025

• [XLA:CPU] Run CSE after inlining in fusion compiler.

  #97114 merged Jul 23, 2025

• Reverts 1b7ced5a66c8948c8b6dfa984ddea56e99c3cba3

  #97404 merged Jul 23, 2025

• [triton] Avoid propagating slice layouts across broadcast.

  #97350 merged Jul 23, 2025

• Prevent XLA from crashing when a Literal is too big to fit in memory.

  #97400 merged Jul 23, 2025

• Refactor shardy_xla_pass_test with IsCustomCall. No behavior change.

  #97388 merged Jul 23, 2025

• [XLA][Numerics][HLO Value Tracking] Support original values during HLO and stableHLO round trip

  #97148 merged Jul 23, 2025

• Automated Code Change

  #97342 merged Jul 23, 2025

• [xla] ShapeUtil: optimize shape traversal

  #97381 merged Jul 23, 2025

• [Hlo Diff] Use full constant values when generating fingerpints and simplify the ExactSubgraphMatcher.

  #96312 merged Jul 23, 2025

• Fix a bug where doing a CPU jax.Array -> TPU jax.Array transfer was not setting the layout correctly on the output TPU array.

  #97393 merged Jul 23, 2025

• Change PjRtClient::LazyToLiteral to take a generator that returns a future of the literal

  #97162 merged Jul 23, 2025

• [HLO Diff] Fix bug in instruction users matching to use fingerprints instead of pointers.

  #97292 merged Jul 23, 2025

• Migrate away from ArrayRef(std::nullopt_t)

  #97382 merged Jul 23, 2025

• [tf] Migrate from std::map to absl::flat_hash_map

  #97376 merged Jul 23, 2025

• [IFRT] Define user_context() in Value and LoadedExecutable

  #97153 merged Jul 23, 2025

• [HLO Diff] Fix the order of operands in ?: operator in TopDownMatcher.

  #97291 merged Jul 23, 2025

• [HLO Diff] Fix existence check in TopDownMatcher.

  #97290 merged Jul 23, 2025

• XLA: Fix method ambiguity

  #97359 merged Jul 23, 2025

• [xla] ShapeUtil: avoid WithStatus callbacks on a hot path

  #97380 merged Jul 23, 2025

• Fix resource calculation for annotation groups that do not contain full pairs of start-done ops.

  #97366 merged Jul 23, 2025

• WithIndex: an iterator wrapper that adds iteration index to yielded value.

  #97370 merged Jul 23, 2025

• [IFRT] Remove the error message mentioning an invalid output target

  #97390 merged Jul 23, 2025

• [xla:tsl] Add Eigen contraction kernel template specialization for int8 LHS x int8 RHS -> int32 result.

  #97383 merged Jul 23, 2025

• Refactor StreamExecutorGpuTopologyDescription::DeviceDescriptions to share SetupDeviceDescription

  #97378 merged Jul 23, 2025

• Increase max stack depth of GPU compile callstack.

  #97384 merged Jul 22, 2025

• Update zlib/zstd patch after

  #97369 merged Jul 22, 2025

• Reverts c197694417be40d121b8b8ba9bb50efd81c4f227

  #97375 merged Jul 22, 2025

• #sdy update hasGspmdAttrsOrOps to look at the body of the main function.

  #97358 merged Jul 22, 2025

• [tf] Do not keep XlaComputation alive after it's compiled to Xla executable

  #97364 merged Jul 22, 2025

• Integrate LLVM at llvm/llvm-project@0a343098b0ea

  #97327 merged Jul 22, 2025

• [xla:cpu] Do not serialize HloProto if user didn't ask for it

  #97363 merged Jul 22, 2025

• Add reshard cache when the sharding of reshape is tile maximal (replicated or maximal).

  #97317 merged Jul 22, 2025

• Removed optimized batch matmul implementation from tflite in favor of XNNPACK. Also performed tiny refactoring along the way.

  #97184 merged Jul 22, 2025

• Internal change only.

  #97195 merged Jul 22, 2025

• [XLA:GPU] enable emitting dots through the generic emitter

  #96720 merged Jul 22, 2025

• Enable N-dimensional sparse tensor in tpu_embedding_v3.py

  #97151 merged Jul 22, 2025

• PR #29115: Bump github/codeql-action from 3.29.2 to 3.29.3

  #97337 merged Jul 22, 2025

• [XLA][host offloading] Open source annotate host compute offload.

  #97362 merged Jul 22, 2025

• [XLA:GPU] Replace "agent" memory sync scope with "device" for nVidia backend.

  #97360 merged Jul 22, 2025

• [XLA:CPU][XLA:GPU] Move concat fusion emitter to shared directory

  #97050 merged Jul 22, 2025

• Automated Code Change

  #97339 merged Jul 22, 2025

• This is an automatic update to a device compatibility allowlist.

  #97346 merged Jul 22, 2025

• [XLA:GPU] Remove horizontal fusion.

  #96241 merged Jul 22, 2025

• Automated Code Change

  #97247 merged Jul 22, 2025

• [XLA:GPU] Choose a tile size according to hardware constraints.

  #97335 merged Jul 22, 2025

• Use CONCURRENT_KERNEL tracing in KernelNameTracer

  #97280 merged Jul 22, 2025

• Fix bf16 propagation for host memory space

  #97310 merged Jul 22, 2025

• Avoid crash in RequestActivityBuffer

  #97279 merged Jul 22, 2025

• Automated Code Change

  #97241 merged Jul 22, 2025

• Add TPU v4i to GetDeviceMemoryInBytes()

  #97316 merged Jul 22, 2025

• [xla] ShapedBuffer: Don't compute on_host_shape if it's the same as device shape

  #97323 merged Jul 22, 2025

• Automated Code Change

  #97248 merged Jul 22, 2025

• Automated Code Change

  #97025 merged Jul 22, 2025

• Fixes minor bug with low memory limits in Peak Priority remat allowing them to be zero.

  #97298 merged Jul 22, 2025

• [xla:cpu] Refactor a few helper functions DotLibraryRewriter for readablility.

  #97324 merged Jul 22, 2025

• Remove AbstractCpuBuffer. All subclasses can be replaced with CommonPjRtBufferImpl and removed.

  #97194 merged Jul 22, 2025

• [xla:cpu] Add useful TraceMe annotations to CpuExecutable

  #97318 merged Jul 22, 2025

• [xla:cpu] Pass --xla_cpu_experimental_onednn_fusion_types to OneDnnMatcher.

  #97320 merged Jul 22, 2025

• [xla] Annotate likely branches with ABSL_PREDICT_TRUE

  #97311 merged Jul 22, 2025

• Add missing includes.

  #96385 merged Jul 22, 2025

• Add a warning log in xla::ShardingPropagation that it will be deprecated and replaced with Shardy.

  #97299 merged Jul 22, 2025

• Introduce AllGatherDynamicSliceShuffledOffsetSimplifier, a new HLO pass that collapse dynamic-slice(all-gather) with shuffled offset into collective-permute.

  #97287 merged Jul 22, 2025

• Updating necessary visibilities ahead of sa-train-google in-place migrations.

  #97304 merged Jul 22, 2025

• Introduce PickUnusedPort() that doesn't terminate when failing to find a port. The new function returns -1 when failing to find an unused port. The original PickUnusedPortOrDie() still terminates when a port can't be found.

  #97308 merged Jul 21, 2025

• [xla] Make Shape::Equal faster

  #97296 merged Jul 21, 2025

• lite: Add config option to enable benchmark_model

  #96999 merged Jul 21, 2025

• Add a RequiresQuantizedBiasInterface

  #96580 merged Jul 21, 2025

• Remove unused API

  #97303 merged Jul 21, 2025

• Update BUILD files with dependencies

  #97305 merged Jul 21, 2025

• Optimize BatchMatmul to Fully Connected when RHS is reshaped after dequantization.

  #97205 merged Jul 21, 2025

• Add SparseCore documentation

  #97200 merged Jul 21, 2025

• Pin python dependencies and update lock files

  #97302 merged Jul 21, 2025

• Update version numbers for TensorFlow 2.20.0-rc0

  #97295 merged Jul 21, 2025

• Use stablehlo precision config conversion for stablehlo ops

  #97233 merged Jul 21, 2025

• Replace implementation of TSL status matchers with Abseil.

  #97283 merged Jul 21, 2025

• [xla:cpu] Make se::HostStream fully synchronous

  #97288 merged Jul 21, 2025

• Skip flaky test on Windows.

  #97286 merged Jul 21, 2025

• Added WatchJobState RPC to coordination service.

  #97077 merged Jul 21, 2025

• Add individual test targets for tensorflow/core:legacy_lib_core_tests.

  #97285 merged Jul 21, 2025

• Updating XNNPACK readme to include the correct weight cache file path field.

  #97289 merged Jul 21, 2025

• Add 10 Maxtext-derived HLO-based benchmarks

  #97132 merged Jul 21, 2025

• [xla] Add benchmarks for Shape::Equal

  #97293 merged Jul 21, 2025

• Avoid heap allocation for the sub buffer address

  #97230 merged Jul 21, 2025

• Avoid checking captured_tensors' usage when deciding if

  #97141 merged Jul 21, 2025

• LatencyHidingScheduler: Only recalculate when we've touched an already-scheduled computation, and use computation-specific peak rather than module peak in statistics.

  #97192 merged Jul 21, 2025

• Reverts 0c0a40e9e8dd2f04ab9a66643ca628acdef5840d

  #97281 merged Jul 21, 2025

• #sdy delete round_trip_import dir since JAX export can now export a model with StableHLO + Shardy dialects.

  #96423 merged Jul 21, 2025

• #sdy use new API to pass dump-index and use MeshAttr::getMaximal

  #97275 merged Jul 21, 2025

• Avoid recomputation of pjrt_buffer->memory_space() in MakeMemoryKindFromPjRtBuffer.

  #97068 merged Jul 21, 2025

• [XLA:GPU] Use correct unroll factor in indexing map

  #97276 merged Jul 21, 2025

• Add missing dependencies.

  #97261 merged Jul 21, 2025

• [XLA] NFC: remove duplication in diag message

  #97274 merged Jul 21, 2025

• [XLA:GPU] Add Get method to GpuPerformanceModelOwning.

  #97277 merged Jul 21, 2025

• Reverts b96ee234c76e3f76341d0acec46e2e7d067cd5a5

  #97273 merged Jul 21, 2025

• [XLA:GPU][host offloading] Implement gpu host offloading allocator.

  #97051 merged Jul 21, 2025

• [XLA:GPU] Refactor tests of IndexingMap

  #97138 merged Jul 21, 2025

• Fix cost analysis on for output byte accessed when result is tuple

  #96927 merged Jul 21, 2025

• Remove LLVM dependency from KernelThunk

  #97178 merged Jul 21, 2025

• Fix a problem with a variable being used out of scope.

  #97272 merged Jul 21, 2025

• [XLA:GPU][Tiling] Use SmallVector<OneDimTile> to store tiling info.

  #97229 merged Jul 21, 2025

• Automated Code Change

  #97268 merged Jul 21, 2025

• Give better error in run_hlo_module if HLO has collectives.

  #97208 merged Jul 21, 2025

• Introduces a new utility function, MatchPermutedSliceAndPartitionOffset, to detect a pattern where a DynamicSlice consumes the output of an AllGather with a permuted set of offsets. This pattern is equivalent to a CollectivePermute and can be optimized accordingly.

  #97189 merged Jul 20, 2025

• Automated Code Change

  #97259 merged Jul 20, 2025

• Add sdy shardings in frontend_attributes alongside hlo shardings for extra wrapper main added in tf2xla bridge.

  #96919 merged Jul 19, 2025

• [XLA:GPU] Support control flow thunks in command buffer conversion pass. We only convert kWhile and kConditional thunks if all thunks in all brunches are convertible.

  #97183 merged Jul 19, 2025

• #sdy Fix forward of making XLA C++ changes so we can fall back to GSPMD in JAX export if the loaded module was lowered for GSPMD.

  #96368 merged Jul 19, 2025

• [XLA] Add helper function GetIndicesSpecForDynamicSlice to get indices spec for dynamic slice fed by all-gather, the spec includes the mapping from slice offsets to corresponding partition IDs(flattened-id).

  #96805 merged Jul 19, 2025

• Internal, visibility only changes to public code.

  #97207 merged Jul 19, 2025

• Add visibility to hlo_input_output_format

  #96758 merged Jul 19, 2025

• Use a literal sentinel value for kernel init failure

  #97193 merged Jul 19, 2025

• Reduce redundancy between StringTo* enum functions.

  #97201 merged Jul 19, 2025

• [XLA:CPU] Refactor Intrinsic and use it in all math intrinsics.

  #97000 merged Jul 19, 2025

• Integrate LLVM at llvm/llvm-project@06ae0c2a1086

  #97154 merged Jul 19, 2025

• Update nccl_archive BUILD file to fix TF GPU wheel build.

  #97206 merged Jul 19, 2025

• [XLA:GPU] Add a verifier to the GPU compiler before post-scheduling pipeline.

  #97150 merged Jul 19, 2025

• Use host callback in the CopyToHostFuture method in Async PjRt.

  #97203 merged Jul 18, 2025

• Add function ExtractDynamicSliceFromCollectiveUser to extract a dynamic slice user from a collective.

  #96802 merged Jul 18, 2025

• no external change

  #97100 merged Jul 18, 2025

• Reverts 849435a30d0487e415126507953575358ed3c4eb

  #97190 merged Jul 18, 2025

• Reverts 2a45c5b0c326e20eafe833df055326b39edadcf2

  #97071 merged Jul 18, 2025

• Bump sqlite to 3.50.3

  #97191 merged Jul 18, 2025

• Typo fix "perferred" -> "preferred".

  #97198 merged Jul 18, 2025

• PR #28257: [XLA:GPU] Update ONEAPI crosstool compiler wrapper

  #97149 merged Jul 18, 2025

• Use ASSERT_THAT to check pass.Run() result

  #97164 merged Jul 18, 2025

• Update the XNNPack delegate README.

  #97181 merged Jul 18, 2025

• Annotate some XLA:GPU flags as stable i.e. they should provide 6 month deprecation notice.

  #97134 merged Jul 18, 2025

• [XLA:GPU] Add a test for DotForInt4vsIdentityBF16ReturnsCorrectResult.

  #97064 merged Jul 18, 2025

• PR #28985: [XLA:GPU] Add shared_memory_per_block_optin device info member

  #97140 merged Jul 18, 2025

• Update README.md

  #96902 merged Jul 18, 2025

• Update dependencies to XNNPACK.

  #97177 merged Jul 18, 2025

• [XLA:GPU] Move Dot strength reduction out of algebraic simplifier

  #97166 merged Jul 18, 2025

• [XLA:GPU] Remove CHECK-CSE since it is not used.

  #97129 merged Jul 18, 2025

• #sdy improve the error messaging when importing and exporting sharding custom calls.

  #97041 merged Jul 18, 2025

• Introduce new helper function that produces device lists for iota tile assignment. Apply it in xla_sharding_util.cc.

  #97176 merged Jul 18, 2025

• Introduce stable flags and associated deprecation policy for XLA debug options.

  #97049 merged Jul 18, 2025

• Use GetInPlaceInputOutputPairs from AliasInfo instead of HloDataflowAnalysis.

  #97170 merged Jul 18, 2025

• Remove ifdef from ir_emitter_unnested and fix various clang-tidy warnings

  #97127 merged Jul 18, 2025

• Add TmaMetadata serialization support

  #97103 merged Jul 18, 2025

• Automated Code Change

  #97109 merged Jul 18, 2025

• Move GetInPlaceInputOutputPairs and related code to AliasInfo class (NFC).

  #97119 merged Jul 18, 2025

• Automated Code Change

  #97123 merged Jul 18, 2025

• Fix tests paths and visibility issue for tflite/converter

  #97147 merged Jul 18, 2025

• Remove leftover logging

  #97145 merged Jul 18, 2025

• Automated Code Change

  #97033 merged Jul 18, 2025

• Propagate context to the waiter destruction sequence, so that all contained operations execute with the correct context.

  #97143 merged Jul 18, 2025

• Update PjRtCpuExecutable to not rely on any internals of PjRtCpuBuffer.

  #97146 merged Jul 18, 2025

• Handle V2 xla::OpSharding in ExtractInputsForLogicalDevices and ParseAndValidateOutputSharding.

  #97136 merged Jul 18, 2025

• Exclude tensorflow/lite/mlir/lite protos definitions when compiling under LiteRT repo and enable LiteRT disbale_tf_lite_py by default

  #97137 merged Jul 18, 2025

• Update version to 2.21.0

  #97079 merged Jul 18, 2025

• [XLA:TPU] In MSA, when removing instructions, we need to remove their scoped allocations from PresetAssignments.

  #96945 merged Jul 17, 2025

• Modified python bindings to enable passing a probe_instrumentation_dir to support interpreter ops in eval_module. Consistent with StableHLO interpreter usage from command line

  #97091 merged Jul 17, 2025

• [XLA][host offloading] Return AsyncValue from HostOffloadingExecutable.

  #96915 merged Jul 17, 2025

• #sdy update dump names and add index as prefix so they would be clearer for users

  #97117 merged Jul 17, 2025

• [Autotuner] Add block level emitter backend for Triton fusion (3).

  #96798 merged Jul 17, 2025

• [IFRT] Add UserContextScope

  #97012 merged Jul 17, 2025

• Add ReleaseDeviceMemoryOwnership implementation based on

  #97144 merged Jul 17, 2025

• Migrate uses of XLA_TEST_BACKEND macros to use utilities in xla_test_backend_predicates.h

  #97135 merged Jul 17, 2025

• Correctly identify async start and done ops in latency hiding scheduler.

  #97089 merged Jul 17, 2025

• Close output shardings to respect allow_spmd_sharding_propagation_to_output flag set to default {false} value. Added multiple test variants to test shardy, use_compile_options_from_model.

  #97126 merged Jul 17, 2025

• [xla:cpu] Make DotLibraryRewriter support greedy fusion mode.

  #96319 merged Jul 17, 2025

• Internal change only

  #97065 merged Jul 17, 2025

• Optimize BM_GlobalDecreasingSizeBestFitHeap benchmark by up to 3%.

  #97075 merged Jul 17, 2025

• Update release notes for TensorFlow 2.20.0

  #97080 merged Jul 17, 2025

• Relax the folding size threshold to 200 MiB.

  #97078 merged Jul 17, 2025

• Update CommonPjRtBufferImpl to have specialized versions for both cpu->device

  #97085 merged Jul 17, 2025

• #sdy define the utils that JAX jaxlib will use to allow for falling back to GSPMD when loading an old checkpoint.

  #97130 merged Jul 17, 2025

• [Autotuner] Add block level emitter backend for Triton fusion (2).

  #96796 merged Jul 17, 2025

• Use ASSERT_THAT(..., IsOkAndHolds(true)) for consistency and correctness

  #97005 merged Jul 17, 2025

• fix(dtensor): guard against nullptr from TF_TensorData in ExtractSmallTensorValue

  #96866 merged Jul 17, 2025

• Reverts 812bb86d50b1cee5cf32ccb1629a49687e924ea5

  #97098 merged Jul 17, 2025

• Simplify ShouldSkipForSideEffect function in zero_sized_hlo_elimination.

  #97101 merged Jul 17, 2025

• [XLA:GPU] Remove unused DotSparsityRewriter.

  #97128 merged Jul 17, 2025

• Automated Code Change

  #97122 merged Jul 17, 2025

• [XLA:GPU] additional logging in triton fusion numeric verifier

  #97056 merged Jul 17, 2025

• [xla:gpu][triton] triton-xla-squeeze-dims pass improvements.

  #97099 merged Jul 17, 2025

• Automated Code Change

  #96959 merged Jul 17, 2025

• PR #28073: [XLA:GPU][oneAPI] Enable Level_zero support

  #97022 merged Jul 17, 2025

• Remove deprecated HloAliasAnalysis::Run method

  #97044 merged Jul 17, 2025

• Add serialization and deserialization for the cuDNN thunk

  #96914 merged Jul 17, 2025

• no external change

  #96942 merged Jul 17, 2025

• [xla] Optimize ShapeUtil::ForEach traverals

  #97063 merged Jul 17, 2025

• Support INT16 for PRelu op

  #96899 merged Jul 17, 2025

• [xla:tf] Check if device shape is already a host shape

  #97018 merged Jul 17, 2025

• Add int16 kernel support for DIV op

  #96934 merged Jul 17, 2025

• Rollback https://github.com/openxla/xla/commit/cf3dfa9723c4cd4e2b25a606207a201a95fe71db

  #97074 merged Jul 17, 2025

• Fix //tflite/converter/tests/... MLIR tests by fixing .bzl rules and redirecting tensorflow submodule

  #97003 merged Jul 16, 2025

• Update release notes at HEAD

  #97073 merged Jul 16, 2025

• Enable --flaky_test_attempts in release branch

  #97076 merged Jul 16, 2025

• Move op name longest prefix logic from annotation.cc to somewhere upper level

  #93906 merged Jul 16, 2025

• Internal change only

  #96928 merged Jul 16, 2025

116 Pull requests opened by 4 people

• Remove `local_config_nvshmem` repository from XLA and Tensorflow WORKSPACE files.

  #97082 opened Jul 17, 2025

• Cache device on `PJRT_Buffer`.

  #97083 opened Jul 17, 2025

• [XLA:MSA] Add block allocations for program weights that are not aliased and single use.

  #97084 opened Jul 17, 2025

• PR #28883: [XLA:CPU][oneDNN] Add build flag to enable asynchronous support in oneDNN

  #97115 opened Jul 17, 2025

• Introduce --dump_tflite_model_dir to dump TFLite models in Delegate Test Suite (DTS)

  #97118 opened Jul 17, 2025

• Integrate LLVM at llvm/llvm-project@06ae0c2a1086

  #97139 opened Jul 17, 2025

• Add Metal LiteRt Tensor Buffer support

  #97142 opened Jul 17, 2025

• Remove redundant string conversion.

  #97152 opened Jul 17, 2025

• Optimize `HasCombinableReplicaGroup` and `xla::CheckReplicaGroups`.

  #97155 opened Jul 18, 2025

• [IFRT] Support XLA GPU flag overrides.

  #97156 opened Jul 18, 2025

• [XLA:MSA] Reduce available memory bandwidth for instruction that are overlapped with bandwidth limiting asynchronous instructions.

  #97163 opened Jul 18, 2025

• Remove HLO and Autotuner dependency from CublasLtMatmulThunk

  #97180 opened Jul 18, 2025

• [XLA] Use sort instead of btree in MakeFreeChunks.

  #97185 opened Jul 18, 2025

• Automated Code Change

  #97188 opened Jul 18, 2025

• IFRT proxy logging fix: Do not log error when Executable is destroyed before its metadata is queried by the server (and sent over to the client).

  #97197 opened Jul 18, 2025

• Handle negative permutations in IsTransposeTrivial.

  #97199 opened Jul 18, 2025

• Also create reduce-scatters when SPMD-partitioning is disabled.

  #97211 opened Jul 19, 2025

• [xla:gpu][triton] In squeeze-dims pass, keep at least two dimensions.

  #97223 opened Jul 19, 2025

• Add a scratch implemention of muon

  #97231 opened Jul 19, 2025

• Automated Code Change

  #97236 opened Jul 19, 2025

• Automated Code Change

  #97237 opened Jul 19, 2025

• Automated Code Change

  #97238 opened Jul 19, 2025

• Add HloInstruction extractor for custom-call kernel metadata.

  #97239 opened Jul 19, 2025

• Automated Code Change

  #97240 opened Jul 19, 2025

• Automated Code Change

  #97249 opened Jul 20, 2025

• Automated Code Change

  #97254 opened Jul 20, 2025

• Fix stride overflow, and solution to issue number #97165

  #97260 opened Jul 20, 2025

• Have HloModule::Clone copy layout_canonicalization_callback.

  #97262 opened Jul 20, 2025

• Automated Code Change

  #97265 opened Jul 21, 2025

• Automated Code Change

  #97267 opened Jul 21, 2025

• Integrate LLVM at llvm/llvm-project@13f7786f72d1

  #97271 opened Jul 21, 2025

• Add an xla flag to lower ragged dots through the ragged dot fusion emitter

  #97278 opened Jul 21, 2025

• Integrate Triton up to [6af1c4b5](https://github.com/openai/triton/commits/6af1c4b507dfddd0d62cdc8613839f46bdf6acb0)

  #97282 opened Jul 21, 2025

• Integrate LLVM at llvm/llvm-project@13f7786f72d1

  #97284 opened Jul 21, 2025

• [HLO Diff] Use full constant values when generating fingerpints.

  #97294 opened Jul 21, 2025

• Update release notes for TensorFlow 2.19.1

  #97297 opened Jul 21, 2025

• [XLA:GPU] Remove horizontal fusion.

  #97301 opened Jul 21, 2025

• PR #28919: [NVIDIA GPU] [XLA_GPU_MS_COLLECTIVE] Round-robin stream assignment for async communications

  #97307 opened Jul 21, 2025

• Adding Metal delegate kernel

  #97312 opened Jul 21, 2025

• Integrate LLVM at llvm/llvm-project@c384ec431dd7

  #97314 opened Jul 21, 2025

• Add gemm_config for xnnpack gemms.

  #97315 opened Jul 22, 2025

• Fixed dependency cleaner support for `tf_kernel_library` so that it will update both `deps` and `gpu_deps`.

  #97319 opened Jul 22, 2025

• [SDY] dump source sharding info when output shardyDir is non-empty. This occurs when `xla_dump_to` flag is set and `xla_dump_hlo_pass_re` regex matches the shardy pass (see `getShardyDirIfShouldDump`).

  #97321 opened Jul 22, 2025

• Integrate LLVM at llvm/llvm-project@c384ec431dd7

  #97322 opened Jul 22, 2025

• Automated Code Change

  #97326 opened Jul 22, 2025

• Automated Code Change

  #97330 opened Jul 22, 2025

• Enable dynamic-slice offset based on multiply for `AllGatherPermutedDsSimplifierTest`, more specifically,

  #97332 opened Jul 22, 2025

• PR #27904: [XLA:GPU][oneAPI] Enable Clang compiler as the host compiler

  #97333 opened Jul 22, 2025

• Add utility function `MatchDsPadAllGather` to match the pattern of `DynamicSlice(Pad(AllGather))`.

  #97336 opened Jul 22, 2025

• [XLA:GPU] Migrate GemmAlgorithmPicker to new autotuner.

  #97340 opened Jul 22, 2025

• Add `_XlaShardingV2` in `TPUPartitionedOps` and use if present for all op's as opposed to just `XlaSharding` op.

  #97348 opened Jul 22, 2025

• Improvements to KernelNameTracer

  #97351 opened Jul 22, 2025

• [XLA][host offloading] Correctly handle queueing tasks on stream.

  #97353 opened Jul 22, 2025

• [XLA:GPU] Enable strength reduction for s32xs32->s32 dots

  #97356 opened Jul 22, 2025

• [XLA:MGPU][Experimental] HLO -> Pallas.

  #97357 opened Jul 22, 2025

• Internal change only.

  #97365 opened Jul 22, 2025

• [ #HLODiff ]Fix a bug in hlo_diff_summary.cc

  #97367 opened Jul 22, 2025

• Turn DCHECK into runtime error. Because literal is user provided, it should

  #97368 opened Jul 22, 2025

• [XLA:CPU] Lower to xla.rsqrt for all rsqrt ops.

  #97371 opened Jul 22, 2025

• Add float data generation within given range to test utils

  #97372 opened Jul 22, 2025

• Add int8/int16 support for SQRT op to AEQ

  #97373 opened Jul 22, 2025

• [XLA:GPU] verify no collective deadlocks regardless if module has schedule

  #97374 opened Jul 22, 2025

• avoid removing some copy kernels.

  #97377 opened Jul 22, 2025

• changes to the build process for litert

  #97385 opened Jul 23, 2025

• Add the skeleton for PropagateQsvPass in the tflite converter.

  #97386 opened Jul 23, 2025

• [xla] Change Shape::IsArray and Shape::IsArrayOrBuffer.

  #97389 opened Jul 23, 2025

• Automated Code Change

  #97391 opened Jul 23, 2025

• Integrate LLVM at llvm/llvm-project@13f7786f72d1

  #97395 opened Jul 23, 2025

• Add an optimization pattern that tranform const<[a, 1]> @ <[1, b]> to <[1, b]> * const<[a, 1]>.

  #97396 opened Jul 23, 2025

• PR #29204: [XLA:GPU] Remove Stream Id from command buffer

  #97397 opened Jul 23, 2025

• Integrate LLVM at llvm/llvm-project@22b083539051

  #97398 opened Jul 23, 2025

• Automated Code Change

  #97399 opened Jul 23, 2025

• PR #26187: Simplify HloComputation::IsFusionComputation semantics.

  #97406 opened Jul 23, 2025

• Replace lite/tools/versioning:versioning with compiler/mlir/lite/tools/versioning:versioning

  #97407 opened Jul 23, 2025

• Automated Code Change

  #97408 opened Jul 23, 2025

• Automated Code Change

  #97409 opened Jul 23, 2025

• Automated Code Change

  #97410 opened Jul 23, 2025

• Automated Code Change

  #97411 opened Jul 23, 2025

• Automated Code Change

  #97412 opened Jul 23, 2025

• Simplify `get-tuple-element(tuple(A))` as `copy(A)` in mhlo->hlo conversion.

  #97413 opened Jul 23, 2025

• Automated Code Change

  #97414 opened Jul 23, 2025

• Automated Code Change

  #97415 opened Jul 23, 2025

• Automated Code Change

  #97416 opened Jul 23, 2025

• Automated Code Change

  #97417 opened Jul 23, 2025

• PR #28883: [XLA:CPU][oneDNN] Add build flag to enable asynchronous support in oneDNN

  #97419 opened Jul 23, 2025

• Automated Code Change

  #97420 opened Jul 23, 2025

• Automated Code Change

  #97421 opened Jul 23, 2025

• [XLA:GPU] Support different unroll factors than just 2.

  #97422 opened Jul 23, 2025

• [XLA] Improve hlo_live_range_test to use HLO textual form

  #97425 opened Jul 23, 2025

• Expose GPU tracing knobs for 3P

  #97428 opened Jul 23, 2025

• Make CustomCallThunk own the associated HloComputation

  #97429 opened Jul 23, 2025

• [XLA:GPU] Enable deviceless AutotunerPass.

  #97430 opened Jul 23, 2025

• Use file mapping on Windows for the XNNPack weight cache.

  #97433 opened Jul 23, 2025

• [XLA:GPU] Adjust indexing cost model heuristic for register usage.

  #97434 opened Jul 23, 2025

• [xla] Change Shape::element_type and Shape::array_or_buffer_element_type.

  #97436 opened Jul 23, 2025

• Extend the Profiler Interface in Autotuner.

  #97437 opened Jul 23, 2025

• Add Multi output reduce test

  #97438 opened Jul 23, 2025

• RaggedDot in HloEvaluator now tolerates inputs with different precision, like conv.

  #97439 opened Jul 23, 2025

• Remove redundant conversion to std::string

  #97441 opened Jul 23, 2025

• [XLA] Don't require flattened control-flow in InfeedTokenPropagation.

  #97442 opened Jul 23, 2025

• [StableHLO] Fix boolean converters

  #97443 opened Jul 23, 2025

• Implement PjRtClient::MakeCrossHostReceiveBuffers and PjRtBuffer:CopyToRemoteDevice in the PjRt C API.

  #97444 opened Jul 23, 2025

• Add an option to swallow errors in HloRunnerAgnosticTestBase and friends.

  #97446 opened Jul 23, 2025

• Add a new constructor for HloRunnerAgnosticTestBase to accept an options struct.

  #97447 opened Jul 23, 2025

• Add bug id in the warning message for involuntary full rematerialization.

  #97448 opened Jul 23, 2025

• Allow MHLO passes to operate on `mlir::FunctionOpInterface` instead of `func::FuncOp`.

  #97449 opened Jul 23, 2025

• [xla] Switch from std::set to absl::flat_hash_set

  #97450 opened Jul 23, 2025

• [StableHLO<->HLO] Remove special handling for AddOp to Xor for booleans

  #97451 opened Jul 23, 2025

• [XLA][Numerics][HLO Value Tracking] Support MLIR dumps from Shardy passes in get_hlo_original_value_coverage.sh

  #97452 opened Jul 23, 2025

• Fix GPU Zero Occupancy Stats

  #97453 opened Jul 23, 2025

• [xla] Optimize ShapeTree construction time for non-tuples

  #97454 opened Jul 23, 2025

• [xla] Remove NOINLINE from ShapeIndex constructor

  #97455 opened Jul 23, 2025

• [xla] Switch from std::unordered_map to absl::flat_hash_map

  #97456 opened Jul 23, 2025

• MPMD XLA bump

  #97457 opened Jul 23, 2025

• Use constexpr absl::string_view for hlo string instead of const char*

  #97458 opened Jul 23, 2025

• Add JAX tests for deadlock verifier

  #97459 opened Jul 23, 2025

5 Issues closed by 3 people

• tensorflow load failed

  #97432 closed Jul 23, 2025

• how to use libtensorflowlite_c.so C API and delegate gpu opencl correctly?

  #95795 closed Jul 23, 2025

• Error in loading Tensorflow in python

  #97345 closed Jul 22, 2025

• Inconsistent NotEqual broadcasting behavior between CPU and GPU (CPU fails silently, GPU raises error)

  #97227 closed Jul 19, 2025

• graph execution error bug with tfm.nlp.layers.MultiHeadRelativeAttention

  #94599 closed Jul 18, 2025

10 Issues opened by 5 people

• tf.keras.Model have not attribute : submodules but the document still use the error attribute

  #97403 opened Jul 23, 2025

• TensorFlow was not built with CUDA kernel binaries compatible with compute capability 12.0 CUDA_ERROR_INVALID_HANDLE

  #97387 opened Jul 23, 2025

• Request for Python 3.13 support in TensorFlow

  #97361 opened Jul 22, 2025

• Inconsistent behavior for `tf.raw_ops.NotEqual` between CPU and GPU with non-broadcastable shapes

  #97204 opened Jul 18, 2025

• could you add support of the new optimizer: Muon

  #97187 opened Jul 18, 2025

• `tf.nn.depthwise_conv2d` crashes with large `strides` values when ONEDNN is enabled

  #97165 opened Jul 18, 2025

• `tf.pow` returns inconsistent value on CPU vs GPU

  #97125 opened Jul 17, 2025

• `tf.nn.local_response_normalization` returns incorrect output

  #97105 opened Jul 17, 2025

• `tf.linalg.matrix_rank` produces inconsistent output on CPU vs GPU with `tol=6`

  #97102 opened Jul 17, 2025

• `tf.math.argmax` throws `InvalidArgumentError` with valid `axis` of `int16` dtype

  #97096 opened Jul 17, 2025

47 Unresolved conversations

Sometimes conversations happen on old items that aren’t yet closed. Here is a list of all the Issues and Pull Requests with unresolved conversations.

• Add a Reflection Map to `emitc` class

  #96263 commented on Jul 23, 2025 • 11 new comments

• TensorFlow DLL failed to load with newer version of TF

  #91656 commented on Jul 17, 2025 • 0 new comments

• Fix compile error in tensorflow/python/tfcompile_wrapper.cc on s390x

  #87676 commented on Jul 21, 2025 • 0 new comments

• Move duplicate CUDA/XLA registration logs from INFO to VLOG

  #89808 commented on Jul 21, 2025 • 0 new comments

• Fix comparison functions and add unit tests

  #94484 commented on Jul 21, 2025 • 0 new comments

• Fix reading mapped names in LayerNameMapper

  #94651 commented on Jul 21, 2025 • 0 new comments

• Update Protobuf to 6.31.1

  #95873 commented on Jul 23, 2025 • 0 new comments

• Remove LiteRT modules from TF python deps.

  #95991 commented on Jul 22, 2025 • 0 new comments

• Add support for deserializing xplanes to Jaxlib

  #96282 commented on Jul 23, 2025 • 0 new comments

• Add validation for label probability distribution in softmax_cross_entropy_with_logits

  #96387 commented on Jul 21, 2025 • 0 new comments

• Call `UpdateGlobalProcessInfo` from PjRt IFRT client.

  #96494 commented on Jul 23, 2025 • 0 new comments

• Fixes L2Pool implementation to not average pooling region squares

  #96599 commented on Jul 23, 2025 • 0 new comments

• PR #19067: [XLA:CPU][oneDNN] Move simplification pass before oneDNN pass

  #96617 commented on Jul 19, 2025 • 0 new comments

• #sdy Remove MHLO shardings from round-trip export

  #96640 commented on Jul 17, 2025 • 0 new comments

• Don't export Shardy in MPMD before going to IFRT, since IFRT can support (and export) a module with StableHLO + Shardy.

  #96648 commented on Jul 23, 2025 • 0 new comments

• Use RPG's solution as a hint to CP-SAT

  #96674 commented on Jul 17, 2025 • 0 new comments

• [TFLite] Support 16KB page sizes alignment for libtensorflowlite_gpu_gl.so

  #96702 commented on Jul 21, 2025 • 0 new comments

• Automated Code Change

  #96929 commented on Jul 23, 2025 • 0 new comments

• Avoid crashing when LRU cache keys change.

  #96930 commented on Jul 23, 2025 • 0 new comments

• Update deps:

  #97035 commented on Jul 23, 2025 • 0 new comments

• [XLA:GPU] Move the s4 unpacking sequence from llvm pass to int4->int8 pass

  #97047 commented on Jul 23, 2025 • 0 new comments

• Allow the chaining of state across MetricHookInterface instantiations for multiple compilations.

  #97054 commented on Jul 17, 2025 • 0 new comments

• [ #HLODiff ] Add support for manual node matching.

  #97060 commented on Jul 22, 2025 • 0 new comments

• No changes to 3rd party.

  #97070 commented on Jul 17, 2025 • 0 new comments

• `tf.experimental.numpy.cumsum` handles overflow inconsistently on CPU and GPU

  #97042 commented on Jul 17, 2025 • 0 new comments

• Core Dump When Training

  #97016 commented on Jul 17, 2025 • 0 new comments

• YoloX different Model Output for Python and Android

  #95489 commented on Jul 17, 2025 • 0 new comments

• Tensorflow 2.19 fails to load after Pyside6

  #97058 commented on Jul 17, 2025 • 0 new comments

• Fail to build libtensorflow_framework.so.2.20.0

  #96569 commented on Jul 18, 2025 • 0 new comments

• `tf.nn.conv2d_transpose` crashes with "Illegal instruction (core dumped)"

  #93733 commented on Jul 18, 2025 • 0 new comments

• [Compatibility][Upgrade] TensorFlow 2.x to 2.15.0: Dependency Conflict and Version Downgrade Issue

  #96694 commented on Jul 19, 2025 • 0 new comments

• Build Error While Compiling TensorFlow Lite Using CMake

  #96654 commented on Jul 19, 2025 • 0 new comments

• Memory leak in tf.data when iterating over Dataset.from_generator

  #65675 commented on Jul 19, 2025 • 0 new comments

• TensorFlow on RTX 5090

  #89272 commented on Jul 19, 2025 • 0 new comments

• GPU Not Detected by TensorFlow Despite Proper System Setup

  #96707 commented on Jul 22, 2025 • 0 new comments

• Some sorting related ops produce results inconsistent with NumPy when tensor contains NaN

  #95235 commented on Jul 22, 2025 • 0 new comments

• tf.data.experimental.prefetch_to_device has no effect inside tf.distribute.Strategy.distribute_datasets_from_function.

  #94735 commented on Jul 22, 2025 • 0 new comments

• `tf.linalg.solve` behaves inconsistently on GPU for singular matrices depending on shape

  #94657 commented on Jul 22, 2025 • 0 new comments

• lib new version not support 16kb pages in android

  #96602 commented on Jul 22, 2025 • 0 new comments

• Convolution: CPU memory increase with growing number of different sequence lengths

  #62441 commented on Jul 23, 2025 • 0 new comments

• Remove or update zh-cn translation from installation instructions

  #62245 commented on Jul 23, 2025 • 0 new comments

• tf.strings.to_number cannot convert positive integers prefixed with "+" when out_type is tf.int32 or tf.int64

  #62191 commented on Jul 23, 2025 • 0 new comments

• failed to build branch r2.13

  #60716 commented on Jul 23, 2025 • 0 new comments

• Incorrect gradient in divide_no_nan and reciprocal_no_nan when divide by 0

  #60715 commented on Jul 23, 2025 • 0 new comments

• `tf.split` or `tf.transpose` cause errors for quantize-aware training with `quantize_apply`

  #60714 commented on Jul 23, 2025 • 0 new comments

• Dataset.ragged_batch does not produce correct specs with tf.py_function and tf.numpy_function

  #60710 commented on Jul 23, 2025 • 0 new comments

• Will TF supprot triton at future

  #96876 commented on Jul 23, 2025 • 0 new comments