1 Release published by 1 person

• v2.20.0-rc0 TensorFlow 2.20.0-rc0

  published Jul 28, 2025

341 Pull requests merged by 4 people

• Automated Code Change

  #98166 merged Aug 3, 2025

• Fix logic error around initialization of DefaultSchedulerCore::ScheduleCandidate

  #98178 merged Aug 3, 2025

• Speed up LatencyHidingScheduler by ~2X for large graphs by improving

  #97718 merged Aug 3, 2025

• [XLA] Simplify and speed up HloModule::ReplaceComputations

  #98111 merged Aug 2, 2025

• PR #29622: [XLA:CPU][BugFix] Fix bug in populating oneDNN post-op arguments

  #98078 merged Aug 2, 2025

• PR #29392: [XLA:CPU][oneDNN][BugFix] Fix a failing FP16 matmul test

  #98079 merged Aug 2, 2025

• Add a bias quantizer pass

  #98122 merged Aug 2, 2025

• Skip i64 when SRQ quantizing

  #98121 merged Aug 2, 2025

• [tsl:concurrency] Silence offsetof compile time warning

  #98104 merged Aug 2, 2025

• [XLA:MSA] Speed up SimplifyGraph by replacing c_find with a lookup in a hashmap.

  #98096 merged Aug 2, 2025

• Automated Code Change

  #98067 merged Aug 2, 2025

• In some cases, we may only want to do strictly legality checks instead of profitability checks when making fusion decisions. This CL adds that ability to InstructionFusion pass.

  #97595 merged Aug 2, 2025

• [XLA][Numerics][Hlo Value Tracking] Do not add a recovery computation for trivial cases

  #98116 merged Aug 2, 2025

• Automated Code Change

  #98126 merged Aug 2, 2025

• [IFRT] Track UserContext in LoadedExecutable implementations

  #98113 merged Aug 2, 2025

• Add xnn_gemm_config with a learnt classifier for zenver2/zenver3.

  #97852 merged Aug 2, 2025

• [XLA] Add Unreduced HLOSharding.

  #97843 merged Aug 2, 2025

• Ensure that the axis tensor is not a nullptr in reduce.

  #94300 merged Aug 2, 2025

• Update rules_ml_toolchain version to enable mirror URLs for the hermetic llvm_linux_x86_64 repository archive.

  #98112 merged Aug 2, 2025

• Add PropagateQsvPass to the TFLite converter.

  #97591 merged Aug 2, 2025

• Remove uses of XLA_BACKEND_DOES_NOT_SUPPORT_FLOAT64 from exhaustive tests

  #98107 merged Aug 1, 2025

• Update rules_ml_toolchain_version to enable usage of newer CUDA/CUDNN versions.

  #98102 merged Aug 1, 2025

• Remove unused variable.

  #98101 merged Aug 1, 2025

• Take serialized PJRT executable as an input for hlo runner.

  #94695 merged Aug 1, 2025

• Cleanup: Remove device_ids from GpuTopology.

  #98012 merged Aug 1, 2025

• Fix python build standalone url for python 3.14-ft on aarch64-unknown-linux-gnu

  #98099 merged Aug 1, 2025

• Build python 3.14.0rc1 and 3.14-nogil from source in linux arm64 docker image

  #98100 merged Aug 1, 2025

• Bump Python Protobuf to 6.31.1.

  #98046 merged Aug 1, 2025

• Move the initializers of the embedding table to be performed in batch.

  #96712 merged Aug 1, 2025

• Improve XNNPack weight cache nameless cache file handling.

  #98080 merged Aug 1, 2025

• Delete uses of XLA_BACKEND_DOES_NOT_SUPPORT_FLOAT64 and replace them with xla::test::BackendSupportsFloat64

  #98032 merged Aug 1, 2025

• [tsl:concurrency] Use absl::NoDestructor to keep static TypeInfo table

  #98094 merged Aug 1, 2025

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

  #97371 merged Aug 1, 2025

• Added ForEach function to LRUCache.

  #98014 merged Aug 1, 2025

• Expose the device list in additional to addressable devices in the prjt executable.

  #97694 merged Aug 1, 2025

• [xla:cpu] Use absl::NoDestructor to keep OkExecuteEvent in static storage

  #98031 merged Aug 1, 2025

• [XLA:ffi] Silence a warning:

  #98091 merged Aug 1, 2025

• [tsl] Silence some deprecation warnings from tsl/platform/file_system.h

  #98089 merged Aug 1, 2025

• [PJRT] Migrate to absl errors, fixes warnings.

  #98090 merged Aug 1, 2025

• Fix msan use-of-uninitialized-value / asan stack-use-after-return error in command_buffer_conversion_pass_test.

  #98086 merged Aug 1, 2025

• Add tests for addSdyShardingsToEntryComputation util.

  #97996 merged Aug 1, 2025

• Prepare for removal of vector.ExtractElement and vector.InsertElement.

  #98081 merged Aug 1, 2025

• Explicitly initialize the new create_arg fields in pjrt_c_api_gpu_test.cc

  #98077 merged Aug 1, 2025

• Fix autotuner test target.

  #98075 merged Aug 1, 2025

• PR #29653: Changed the heading level for the TPU XLA flags section

  #98076 merged Aug 1, 2025

• Automated Code Change

  #98045 merged Aug 1, 2025

• Update TFRT dependency to use revision

  #98048 merged Aug 1, 2025

• [XLA:GPU] log which backend is selected in autotuner

  #97990 merged Aug 1, 2025

• Automated Code Change

  #98059 merged Aug 1, 2025

• Automated Code Change

  #97987 merged Aug 1, 2025

• Automated Code Change

  #97939 merged Aug 1, 2025

• Automated Code Change

  #97982 merged Aug 1, 2025

• Automated Code Change

  #98061 merged Aug 1, 2025

• Automated Code Change

  #98060 merged Aug 1, 2025

• Automated Code Change

  #98055 merged Aug 1, 2025

• Automated Code Change

  #98058 merged Aug 1, 2025

• Automated Code Change

  #98057 merged Aug 1, 2025

• Automated Code Change

  #98056 merged Aug 1, 2025

• Automated Code Change

  #98050 merged Aug 1, 2025

• Automated Code Change

  #98053 merged Aug 1, 2025

• Automated Code Change

  #98051 merged Aug 1, 2025

• Automated Code Change

  #98052 merged Aug 1, 2025

• Automated Code Change

  #98049 merged Aug 1, 2025

• Automated Code Change

  #97985 merged Aug 1, 2025

• Automated Code Change

  #97984 merged Aug 1, 2025

• Automated Code Change

  #97981 merged Aug 1, 2025

• Automated Code Change

  #97983 merged Aug 1, 2025

• Automated Code Change

  #97623 merged Aug 1, 2025

• Automated Code Change

  #97978 merged Aug 1, 2025

• Automated Code Change

  #97942 merged Aug 1, 2025

• Automated Code Change

  #97975 merged Aug 1, 2025

• Automated Code Change

  #97974 merged Aug 1, 2025

• [xla:cpu] Remove unneeded check for the number of computations

  #98047 merged Aug 1, 2025

• Automated Code Change

  #97957 merged Aug 1, 2025

• Update grpc and abseil-cpp deps.

  #97838 merged Aug 1, 2025

• Automated Code Change

  #98042 merged Aug 1, 2025

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

  #97163 merged Aug 1, 2025

• Automated Code Change

  #97986 merged Aug 1, 2025

• Reverts changelist 789470746

  #98034 merged Aug 1, 2025

• [XLA][Numerics][HLO Value Tracking] Add a test for keeping value recovery table through Shardy passes

  #97850 merged Aug 1, 2025

• [xla:pjrt] Optimize PjRtFuture<>(absl::OkStatus()) construction

  #98016 merged Aug 1, 2025

• Fix a typo in the cross-host transfer attribute in the C API client and add a test that would have caught it.

  #98035 merged Aug 1, 2025

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

  #97084 merged Aug 1, 2025

• Add Metal LiteRt Tensor Buffer support

  #97142 merged Aug 1, 2025

• Upgrade python 3.14.0rc1 and install python 3.14-nogil in linux docker images

  #98018 merged Aug 1, 2025

• Removed tflite from _pywrap_tensorflow and generated from tf_python_api_gen_v2

  #97534 merged Aug 1, 2025

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

  #97060 merged Aug 1, 2025

• OSS-ing the generic pjrtTopologyDescription proto

  #97543 merged Aug 1, 2025

• Call UpdateGlobalProcessInfo from PjRt IFRT client.

  #96494 merged Aug 1, 2025

• Remove unused proto file diff_option.proto

  #98027 merged Aug 1, 2025

• Improve PJRT logging

  #98000 merged Aug 1, 2025

• [XLA:GPU] remove execution_thread="explicit" annotation for GPU stream annotation passes

  #98022 merged Aug 1, 2025

• Delete all uses of XLA_BACKEND_DOES_NOT_SUPPORT_FLOAT16

  #98019 merged Aug 1, 2025

• Delete all uses of XLA_BACKEND_DOES_NOT_SUPPORT_BFLOAT16

  #98008 merged Aug 1, 2025

• [XLA:Python] Remove Squash64BitTypes, which is no longer used by JAX.

  #98010 merged Aug 1, 2025

• Fix bug uncovered by setting very low transfer sizes on an internal test.

  #98011 merged Aug 1, 2025

• Refactor h2d_copy in TfrtGpuClient::BufferFromHostBuffer to make the logic more clear

  #98007 merged Aug 1, 2025

• [XLA:Python] Split xla::Literal type casters into their own module.

  #98009 merged Aug 1, 2025

• PR #28782: [XLA:GPU] Annotate cuBLAS/cuDNN outputs to avoid initcheck failures

  #97831 merged Aug 1, 2025

• Automated Code Change

  #97988 merged Jul 31, 2025

• [xla:cpu:nanort] Cleanup ifrt_client + sprinkle ABSL_PREDICT_FALSE on unlikely branches

  #98006 merged Jul 31, 2025

• Reverts cc792635229ab2ba91d8a420fde0a86f58042c53

  #98003 merged Jul 31, 2025

• Automated Code Change

  #97945 merged Jul 31, 2025

• Automated Code Change

  #97947 merged Jul 31, 2025

• Call on_done_with_host_buffer only after data being copied to staging buffer or device.

  #97893 merged Jul 31, 2025

• Automated Code Change

  #97829 merged Jul 31, 2025

• Minor refactor to remove some verbosity in SymbolicExprTest

  #97977 merged Jul 31, 2025

• Remove the handling of MemRef in MHLO to HLO type converter.

  #97851 merged Jul 31, 2025

• Automated Code Change

  #97958 merged Jul 31, 2025

• PR #29517: Fix the presubmit lld_failure

  #97953 merged Jul 31, 2025

• setup configuration for shardy support to follow

  #97813 merged Jul 31, 2025

• [XLA] Support DimensionVector return type in math_util, permutation_util, util.

  #97858 merged Jul 31, 2025

• Change test from xla_test to xla_cc_test.

  #97979 merged Jul 31, 2025

• Move SymbolicExprStorage to the .cc file

  #97980 merged Jul 31, 2025

• [xla:cpu:nanort] Cache single-device device lists to avoid heap allocations on a hot path

  #97927 merged Jul 31, 2025

• Migrate away from mlir::OpPrintingFlags(std::nullopt_t)

  #97859 merged Jul 31, 2025

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

  #97830 merged Jul 31, 2025

• Add operators and overloading functions to SymbolicExpr

  #97837 merged Jul 31, 2025

• Automated Code Change

  #96864 merged Jul 31, 2025

• Automated Code Change

  #97973 merged Jul 31, 2025

• Automated Code Change

  #96923 merged Jul 31, 2025

• Add uniquing of SymbolicExpr based on StorageUniquer

  #97826 merged Jul 31, 2025

• Automated Code Change

  #96853 merged Jul 31, 2025

• Automated Code Change

  #96850 merged Jul 31, 2025

• Automated Code Change

  #96795 merged Jul 31, 2025

• Automated Code Change

  #97970 merged Jul 31, 2025

• Automated Code Change

  #97967 merged Jul 31, 2025

• Automated Code Change

  #97968 merged Jul 31, 2025

• Automated Code Change

  #97971 merged Jul 31, 2025

• Automated Code Change

  #97972 merged Jul 31, 2025

• Automated Code Change

  #96929 merged Jul 31, 2025

• [XLA:GPU][Triton] Handle elementwise ops with multiple results in triton-xla-squeeze-dims.

  #97948 merged Jul 31, 2025

• Automated Code Change

  #97963 merged Jul 31, 2025

• Automated Code Change

  #97962 merged Jul 31, 2025

• Automated Code Change

  #96863 merged Jul 31, 2025

• Automated Code Change

  #97961 merged Jul 31, 2025

• Automated Code Change

  #97969 merged Jul 31, 2025

• Automated Code Change

  #97965 merged Jul 31, 2025

• Automated Code Change

  #97964 merged Jul 31, 2025

• [XLA:CPU] Merge entry arguments before wrapping in call frame.

  #97955 merged Jul 31, 2025

• Automated Code Change

  #97959 merged Jul 31, 2025

• Automated Code Change

  #97952 merged Jul 31, 2025

• Automated Code Change

  #97954 merged Jul 31, 2025

• Automated Code Change

  #97951 merged Jul 31, 2025

• Automated Code Change

  #97943 merged Jul 31, 2025

• Automated Code Change

  #97937 merged Jul 31, 2025

• Automated Code Change

  #97936 merged Jul 31, 2025

• [XLA:GPU] Avoid non-determinism in ComputationPartitioner.

  #97950 merged Jul 31, 2025

• Automated Code Change

  #97933 merged Jul 31, 2025

• Automated Code Change

  #97934 merged Jul 31, 2025

• Automated Code Change

  #97931 merged Jul 31, 2025

• Automated Code Change

  #97926 merged Jul 31, 2025

• Automated Code Change

  #97932 merged Jul 31, 2025

• Automated Code Change

  #97930 merged Jul 31, 2025

• Automated Code Change

  #97925 merged Jul 31, 2025

• Automated Code Change

  #96857 merged Jul 31, 2025

• Automated Code Change

  #97949 merged Jul 31, 2025

• Automated Code Change

  #97922 merged Jul 31, 2025

• Automated Code Change

  #97919 merged Jul 31, 2025

• Automated Code Change

  #97915 merged Jul 31, 2025

• Automated Code Change

  #97917 merged Jul 31, 2025

• Automated Code Change

  #97921 merged Jul 31, 2025

• [XLA:CPU][XLA:GPU] Print compilation backend in slow compilation alarm.

  #97807 merged Jul 31, 2025

• Automated Code Change

  #97920 merged Jul 31, 2025

• Automated Code Change

  #97918 merged Jul 31, 2025

• Automated Code Change

  #97916 merged Jul 31, 2025

• Automated Code Change

  #97913 merged Jul 31, 2025

• Automated Code Change

  #97907 merged Jul 31, 2025

• Automated Code Change

  #97914 merged Jul 31, 2025

• Automated Code Change

  #97912 merged Jul 31, 2025

• Automated Code Change

  #97911 merged Jul 31, 2025

• Automated Code Change

  #97908 merged Jul 31, 2025

• Automated Code Change

  #97910 merged Jul 31, 2025

• Automated Code Change

  #97906 merged Jul 31, 2025

• Automated Code Change

  #97905 merged Jul 31, 2025

• Automated Code Change

  #97900 merged Jul 31, 2025

• Automated Code Change

  #97901 merged Jul 31, 2025

• Automated Code Change

  #97902 merged Jul 31, 2025

• Automated Code Change

  #97899 merged Jul 31, 2025

• [XLA:GPU] Nit: Update cuDNN backend description.

  #97766 merged Jul 31, 2025

• Automated Code Change

  #97897 merged Jul 31, 2025

• Automated Code Change

  #97898 merged Jul 31, 2025

• Automated Code Change

  #97894 merged Jul 31, 2025

• [XLA] Add F8E4M3FN and F8E5M2 to F32 conversion.

  #97884 merged Jul 31, 2025

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

  #97282 merged Jul 31, 2025

• [XLA:GPU] Store constraints in the tiling space.

  #97904 merged Jul 31, 2025

• Automated Code Change

  #97891 merged Jul 31, 2025

• Automated Code Change

  #97889 merged Jul 31, 2025

• Automated Code Change

  #97887 merged Jul 31, 2025

• Automated Code Change

  #97879 merged Jul 31, 2025

• Automated Code Change

  #97886 merged Jul 31, 2025

• Automated Code Change

  #97883 merged Jul 31, 2025

• Automated Code Change

  #97882 merged Jul 31, 2025

• Automated Code Change

  #97888 merged Jul 31, 2025

• Automated Code Change

  #97881 merged Jul 31, 2025

• Automated Code Change

  #97616 merged Jul 31, 2025

• [IFRT] Remove temporary redirection methods that takes user_context

  #97860 merged Jul 31, 2025

• [XLA][Numerics][HLO Value Tracking] Add recovery modules when skipping degenerate reshapes before broadcasting

  #97596 merged Jul 31, 2025

• Automated Code Change

  #97878 merged Jul 31, 2025

• Automated Code Change

  #97876 merged Jul 31, 2025

• Automated Code Change

  #97877 merged Jul 31, 2025

• Automated Code Change

  #97874 merged Jul 31, 2025

• Automated Code Change

  #97873 merged Jul 31, 2025

• Automated Code Change

  #97875 merged Jul 31, 2025

• Automated Code Change

  #97863 merged Jul 31, 2025

• Automated Code Change

  #97872 merged Jul 31, 2025

• Automated Code Change

  #97871 merged Jul 31, 2025

• Automated Code Change

  #97868 merged Jul 31, 2025

• [IFRT] Add tests for Array::user_context()

  #97774 merged Jul 31, 2025

• Automated Code Change

  #97862 merged Jul 31, 2025

• Automated Code Change

  #97861 merged Jul 31, 2025

• Automated Code Change

  #97867 merged Jul 31, 2025

• Automated Code Change

  #97866 merged Jul 31, 2025

• Automated Code Change

  #97869 merged Jul 31, 2025

• [XLA][Numerics][HLO Value Tracking] Add recovery modules when removing no-op transposes

  #97604 merged Jul 31, 2025

• [IFRT] Add basic support for Array::user_context() to NanoRt

  #97779 merged Jul 31, 2025

• [IFRT] Add basic support for Array::user_context() to IFRT Proxy client

  #97777 merged Jul 31, 2025

• [IFRT] Add basic support for Array::user_context() to PjRt-IFRT

  #97776 merged Jul 31, 2025

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

  #97444 merged Jul 31, 2025

• Add support for deserializing xplanes to Jaxlib

  #97518 merged Jul 31, 2025

• [JAX] Experimental: Use device for the default memory of CPU devices

  #97787 merged Jul 31, 2025

• [IFRT] Remove an explicit user_context argument from APIs

  #97781 merged Jul 31, 2025

• Add BackendSupports{Float64,Complex128} to xla_test_backend_predicates.

  #97849 merged Jul 31, 2025

• Add location fields to OpMetadata: source_end_line, source_column, and source_end_column

  #97775 merged Jul 31, 2025

• Internal tooling change

  #97598 merged Jul 31, 2025

• Port recent MHLO changes to StableHLO optimization path.

  #95210 merged Jul 30, 2025

• fix a bug

  #97845 merged Jul 30, 2025

• Update Protobuf to 6.31.1

  #95873 merged Jul 30, 2025

• Fix problem where a double-poison could happen in certain circumstances which

  #97842 merged Jul 30, 2025

• Move assignment after early exit in SimplifyConvToDot

  #97707 merged Jul 30, 2025

• [XLA:MSA] Add explicit pinning mode in MSA, sort buffer intervals and allocate pinned buffers first.

  #97773 merged Jul 30, 2025

• Reshard should not modify the base shape. PartitionedHlo::TryComplexReshardHandling may modify the base shape and trigger other issues.

  #97783 merged Jul 30, 2025

• Add support for GetHostPointer() to the raw-buffer c-api. This will allow zero-copy from buffers which

  #97788 merged Jul 30, 2025

• [pjrt gpu] Improve naming of lambdas to make things consistent

  #97836 merged Jul 30, 2025

• Use Slinky by default for XNNPACK ops/fusions

  #97839 merged Jul 30, 2025

• [XLA:GPU] Store symbolic tiles for the roots of the fusion in TilingSpace.

  #97835 merged Jul 30, 2025

• Update units tests to work on v6

  #97784 merged Jul 30, 2025

• Add a heuristic to choose the quantization type for same-scale ops.

  #97537 merged Jul 30, 2025

• Add proto for serialized executble metadata to IFRT

  #97782 merged Jul 30, 2025

• Automated Code Change

  #97828 merged Jul 30, 2025

• [XLA][host offloading] Make writes to events map thread safe.

  #97827 merged Jul 30, 2025

• [XLA:GPU] Wrap kernel args into LLVM IrArray in IrEmitterUnnested.

  #97823 merged Jul 30, 2025

• Reverts f727033f09806479ff8e9dbf9c960824c67ab980

  #97811 merged Jul 30, 2025

• [XLA:GPU] Do not remove duplicate arguments from kernel interface.

  #97752 merged Jul 30, 2025

• Add NOLINT to suppress redundant string conversion warnings

  #97803 merged Jul 30, 2025

• PR #28936: clean device description for rocm

  #97825 merged Jul 30, 2025

• [XLA:CPU] Fix unrolling pass.

  #97757 merged Jul 30, 2025

• Update XNNPACK to enable SME2 by default

  #97789 merged Jul 30, 2025

• [XLA:GPU] ignore "invalid argument" status in gemm autotuning

  #97755 merged Jul 30, 2025

• Update Maxtext-derived HLOs with large constants

  #97536 merged Jul 30, 2025

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

  #96702 merged Jul 30, 2025

• [XLA:CPU] Don't fuse non-scalar constants.

  #97759 merged Jul 30, 2025

• [XLA:GPU] Fix: Check for bf16x2 enabled when using inline asm. Otherwise int4 A100 tests are failing.

  #97747 merged Jul 30, 2025

• Add unit test with non-flat call graph on shardy xla pass.

  #97692 merged Jul 30, 2025

• PR #29332: Fix the xla/service/gpu:gpu_aot_compilation_test

  #97821 merged Jul 30, 2025

• PR #29342: Fix the oneAPI Presubmit failure

  #97820 merged Jul 30, 2025

• Add a missing symbol to unbreak Windows.

  #97818 merged Jul 30, 2025

• #sdy if we need to use tuples, tuplify entry computation layout even if no layout has been set.

  #97686 merged Jul 30, 2025

• [XLA:CPU] Add logging for LLVM compilation time.

  #97763 merged Jul 30, 2025

• Remove PJRT v70 checks since the 3 week forward compatibility window has passed.

  #97690 merged Jul 30, 2025

• [XLA:CPU] Run inline pass multiple times.

  #97758 merged Jul 30, 2025

• [XLA:GPU] Do not call ComposePermutations for non-permutations (NFC).

  #97814 merged Jul 30, 2025

• Add a variadic scatter test with repeated operand.

  #97804 merged Jul 30, 2025

• [XLA:GPU/TMA] Restrict TMA for use-cases where the offsets don't meet the constraints. The original issue manifested in the form of CUDA_ERROR_ILLEGAL_INSTRUCTION on microbenchmarks that used TMA.

  #97751 merged Jul 30, 2025

• [XLA:GPU] Enable Triton TMA tuning in new autotuner. Added support in both backends that support Triton. Also refactored a common tma utility and updated the usages.

  #97685 merged Jul 30, 2025

• [XLA] Support DimensionVector return type in math_util, permutation_util, util.

  #97726 merged Jul 30, 2025

• #sdy set PropagationOptions::enableAutoPartitioning in shardy_xla_pass and cleanup legacy auto partitioner code, as it's no longer needed.

  #97506 merged Jul 30, 2025

• [XLA:GPU] Code style improvements for matmul_indexing_utils

  #97801 merged Jul 30, 2025

• [XLA:GPU] Use MapVector instead of DenseMap for storing constraints.

  #97753 merged Jul 30, 2025

• Change return value of IFRT MakeDeviceList from DeviceListRef to absl::StatusOr<DeviceListRef> so that upcoming CLs can do validation that the devices match requirements.

  #97715 merged Jul 30, 2025

• Allow CPU embedding lookup inputs to be a subset of embedding feature configs.

  #97719 merged Jul 29, 2025

• PR #29390: pjrt/ffi: add FFI handler registration API to the FFI PjRt extension

  #97770 merged Jul 29, 2025

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

  #97767 merged Jul 29, 2025

• [XLA:GPU] Use XLA_VLOG_LINES instead of VLogModule.

  #97771 merged Jul 29, 2025

• Fix crash on reentrant call to LRUCache::Clear().

  #97681 merged Jul 29, 2025

• Disable de-Morgan's law clang-tidy check on macro. Users of the macro aren't aware, don't control, and shouldn't care that the macro expands to "!(x)", and it's reasonable for users to pass in "a || b" as "x" and shouldn't cause clang-tidy findings.

  #97768 merged Jul 29, 2025

• [XLA:GPU] move matmul_indexing_utils one dir up

  #97762 merged Jul 29, 2025

• [XLA:GPU] Refactor KernelArguments::Create.

  #97754 merged Jul 29, 2025

• [XLA:GPU] Remove MaterializeOp (NFC).

  #97756 merged Jul 29, 2025

• Update the coords of PjRtStreamExecutorDeviceDescription.

  #97465 merged Jul 29, 2025

• [xla:cpu] Register runtime symbols as weak

  #97765 merged Jul 29, 2025

• [xla:cpu] Fix use-after-free on error handling path

  #97741 merged Jul 29, 2025

• [XLA:GPU] Remove redundant needed_operands argument from KernelArguments::Create.

  #97746 merged Jul 29, 2025

• PR #29386: Bump github/codeql-action from 3.29.3 to 3.29.4

  #97749 merged Jul 29, 2025

• [XLA:GPU] Remove FuseLoops pass

  #97748 merged Jul 29, 2025

• [XLA:GPU] Pass operands to BuildKernelThunkForNonFusionOp for RNG and sort kernels.

  #97745 merged Jul 29, 2025

• [XLA:GPU][Tiling] Store ptr to TilingSpace in SymbolicTile.

  #97710 merged Jul 29, 2025

• [XLA:CPU] [XLA:GPU] Remove a source of non-determinism in emitters.

  #97732 merged Jul 29, 2025

• [XLA:CPU][XLA:GPU] Add has_no_compute to epilogue.

  #97696 merged Jul 29, 2025

• Automated Code Change

  #97659 merged Jul 29, 2025

• Don't make a copy of xla_dump_hlo_pass_re

  #97740 merged Jul 29, 2025

• [XLA:GPU] Remove distinction between inputs and outputs in BuildKernelPrototype.

  #97704 merged Jul 29, 2025

• [XLA:GPU] Add SymbolicTiledHlo class.

  #97709 merged Jul 29, 2025

• [XLA:CPU][XLA:GPU] Add missing instructions from has_no_compute list.

  #97695 merged Jul 29, 2025

• Automated Code Change

  #97641 merged Jul 29, 2025

• [XLA:GPU] Wrap ragged dot into a custom fusion.

  #97691 merged Jul 29, 2025

• Automated Code Change

  #97622 merged Jul 29, 2025

• Automated Code Change

  #97675 merged Jul 29, 2025

• Automated Code Change

  #97652 merged Jul 29, 2025

• [shlo] Add module to avoid depending on position of split (NFC).

  #97703 merged Jul 29, 2025

• r2.20 cherry-pick: b0b73e1d3b4 "Fix release wheels build."

  #97728 merged Jul 29, 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 contains shardy-verbose.

  #97321 merged Jul 29, 2025

• Remove XLA_TEST_BACKEND defines from xla_test implementation now that there are no users

  #97713 merged Jul 29, 2025

• Fix release wheels build.

  #97727 merged Jul 29, 2025

• [xla:cpu] Rename Thunk::async_resume() to ExecutesOnExternalThreadPool()

  #97702 merged Jul 29, 2025

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

  #97712 merged Jul 28, 2025

• Added benchmark for LatencyHidingScheduler

  #97711 merged Jul 28, 2025

• Always pass TargetMachineFeatures to IsDotSupportedByXnn.

  #97682 merged Jul 28, 2025

• Remove duplicate patch files.

  #97572 merged Jul 28, 2025

• Use LazyRE2 to avoid repeated compilation of regexes as recommended in https://abseil.io/fast/21

  #97689 merged Jul 28, 2025

• Migrate away from llvm::Intrinsic::getDeclaration

  #97646 merged Jul 28, 2025

• Fix nccl for libc++.

  #97527 merged Jul 28, 2025

• Support bfloat16 type when serializing to numpy

  #97557 merged Jul 28, 2025

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

  #97396 merged Jul 28, 2025

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

  #97698 merged Jul 28, 2025

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

  #97597 merged Jul 28, 2025

• xla_compile: Link the GPU plugin as well as the CPU plugin.

  #97594 merged Jul 28, 2025

• Add documentation for pywrap rules

  #97636 merged Jul 28, 2025

• OptimizeSparseTensor::Split() by not calling Tensor::matrix() for each element.

  #97647 merged Jul 28, 2025

• [IFRT Proxy] Bump the minimum version of the IFRT Proxy client to kSerDesVersioning.

  #97472 merged Jul 28, 2025

• Remove unused variable in ResizeBilinear888Uint8

  #97484 merged Jul 28, 2025

• Share the logic to check if a module has an unapproved dialect for versioned serialization.

  #97504 merged Jul 28, 2025

• [XLA] WhileLoopUnroller should only inline newly created calls.

  #97526 merged Jul 28, 2025

• [XLA:GPU] Add the s4 unpacking sequence similar to the one from llvm pass to int4->int8 pass and add two more improvements to the triton lowering passes.

  #97047 merged Jul 28, 2025

• PR #24406: Initial Cupti PM sampling support for gpu profiler

  #97583 merged Jul 28, 2025

• [XLA] Don't flatten the CFG inside WhileLoopUnroller.

  #97520 merged Jul 28, 2025

• Move SymbolicExpr to xla experimental folder

  #97683 merged Jul 28, 2025

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

  #97278 merged Jul 28, 2025

• Automated Code Change

  #97679 merged Jul 28, 2025

• [XLA:GPU] Do not dump hlo for autotuning compilations.

  #97578 merged Jul 28, 2025

• cupti_buffer_events_test cleanup

  #97678 merged Jul 28, 2025

• [XLA:GPU] fusion emitter: dump hlo as a module with a name matching the fusion

  #97575 merged Jul 28, 2025

• [XLA:GPU] Add input->output tiling for cwise-transpose-broadcast.

  #97619 merged Jul 28, 2025

• Automated Code Change

  #97665 merged Jul 28, 2025

• [XLA:GPU] Increase unroll factor for sort to 4.

  #97571 merged Jul 28, 2025

• Automated Code Change

  #97667 merged Jul 28, 2025

• Replace ElementsAreArray(<float>) to Pointwise(FloatingPointEq(), <float>) for StablehloCaseTestFloat

  #97491 merged Jul 28, 2025

• Automated Code Change

  #97660 merged Jul 28, 2025

• Only add interpreter op after stablehlo ops.

  #97589 merged Jul 28, 2025

• Start support litert cmake build by fixing tflite

  #97650 merged Jul 28, 2025

• Automated Code Change

  #97648 merged Jul 28, 2025

168 Pull requests opened by 9 people

• [FIX] multi_process_runner_test times out after 2400 seconds on gfx11xx and gfx12xx GPUs

  #97680 opened Jul 28, 2025

• GPU: Fix overflow inconsistency in cumsum on CPU/GPU

  #97693 opened Jul 28, 2025

• Reset the RBE cache on Windows.

  #97699 opened Jul 28, 2025

• Don't construct GPU cliques with stale incarnations.

  #97700 opened Jul 28, 2025

• [oneDNN] Refactor oneDNN matmul op integration to avoid unnecessary thread pool object creation

  #97701 opened Jul 28, 2025

• Add a profiler for subprocesses.

  #97705 opened Jul 28, 2025

• Another test here

  #97708 opened Jul 28, 2025

• Use LazyRE2 to avoid repeated compilation of regexes in tensorflow/lite/experimental/acceleration/compatibility/canonicalize_value.cc

  #97714 opened Jul 28, 2025

• Introduce subclassed versions of HostOffloader for different backends.

  #97716 opened Jul 28, 2025

• [tosa] updated scatter_nd to support duplicate indices

  #97717 opened Jul 28, 2025

• Remove metal_memory deps from tensor buffer header.

  #97721 opened Jul 28, 2025

• Test: Release 2 of Rules ML Toolchain testing

  #97722 opened Jul 28, 2025

• Fix: Prevent Duplicate Plugin Factory Registration (cuDNN/cuFFT/cuBLAS)

  #97724 opened Jul 28, 2025

• Check `pjrt_apis` is initialized in `GetRegisteredPjrtApis` and `IsPjrtPluginInitialized`.

  #97725 opened Jul 29, 2025

• Expose interface to fingerprint individual HLO computations or instructions.

  #97730 opened Jul 29, 2025

• Automated Code Change

  #97731 opened Jul 29, 2025

• [XLA:CPU] Don't fuse dynamic-update-slice in new emitters.

  #97760 opened Jul 29, 2025

• [XLA:CPU] Add caching for new fusion kernels.

  #97761 opened Jul 29, 2025

• [XLA:GPU] Add scaled dot HLO instruction

  #97769 opened Jul 29, 2025

• Add `AllGatherPadDsSimplifier` into GPU compiler pipeline,

  #97772 opened Jul 29, 2025

• [stablehlo] Fold SetDimensionSize ops if dimension size being set matches static result size.

  #97778 opened Jul 29, 2025

• [XLA] Add some missing includes and remove unused ones.

  #97786 opened Jul 30, 2025

• [XLA:GPU] Migrate ConvAlgorithmPicker to new autotuner backend.

  #97806 opened Jul 30, 2025

• PR #28782: [XLA:GPU] Annotate cuBLAS/cuDNN outputs to avoid initcheck failures

  #97808 opened Jul 30, 2025

• Unify OSS and internal logging libraries.

  #97809 opened Jul 30, 2025

• [XLA:CPU] Enable new fusions by default.

  #97815 opened Jul 30, 2025

• [XLA] Add scale shape verifier for ScaledDot to hlo_verifier.

  #97819 opened Jul 30, 2025

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

  #97822 opened Jul 30, 2025

• Reverts f727033f09806479ff8e9dbf9c960824c67ab980

  #97824 opened Jul 30, 2025

• [XLA:CPU] Don't inline after lowering to LLVM.

  #97832 opened Jul 30, 2025

• Support extracting `xla::OpSharding`s from sdy frontend attributes

  #97833 opened Jul 30, 2025

• Avoid undefined behavior when `LRUCache` keys change.

  #97834 opened Jul 30, 2025

• [GPU] Cost model to evaluate if converting dot to reduce is profitable.

  #97841 opened Jul 30, 2025

• [xla:sdy] Export unreduced as HLOSharding instead of frontend attributes.

  #97844 opened Jul 30, 2025

• [XLA] Simplify orthogonal AG->reduce to reduce->AG.

  #97846 opened Jul 30, 2025

• [XLA] Allow removing dynamism from the module

  #97847 opened Jul 30, 2025

• Integrate LLVM at llvm/llvm-project@fa6965f722e0

  #97848 opened Jul 30, 2025

• Add a peek method to AncestorStack.

  #97853 opened Jul 30, 2025

• Cleanup `InterpreterBuilder::operator()`.

  #97854 opened Jul 31, 2025

• Remove unused proto file `diff_option.proto`

  #97855 opened Jul 31, 2025

• Do not erase location info when moving an op within the same block.

  #97856 opened Jul 31, 2025

• Fix typo in XlaSparseGradientsStack shape inference

  #97857 opened Jul 31, 2025

• Integrate LLVM at llvm/llvm-project@b9a627e6fba4

  #97880 opened Jul 31, 2025

• Add CHECK before accessing possibly-empty vector

  #97895 opened Jul 31, 2025

• Automated Code Change

  #97896 opened Jul 31, 2025

• Automated Code Change

  #97938 opened Jul 31, 2025

• PR #27943: [XLA:GPU] Add support for intel gpu backend for tests

  #97956 opened Jul 31, 2025

• Automated Code Change

  #97976 opened Jul 31, 2025

• Replace sdy::kShardingRoundTripAttr with HloSharding::kShardingFrontendAttrName

  #97989 opened Jul 31, 2025

• Update the tuple sharding scope for `RecvOp` to avoid copying over the entire sharding to `GetTupleElement`.

  #97991 opened Jul 31, 2025

• #sdy fix JAX shmap returning pure_callback.

  #97992 opened Jul 31, 2025

• [XLA:GPU] Remove SparseDot op from the MHLO dialect

  #97993 opened Jul 31, 2025

• [XLA:GPU] Consolidate backend config to protobuf::Any.

  #97995 opened Jul 31, 2025

• [XLA:GPU] Remove sparsity support from HloDotInstruction.

  #97997 opened Jul 31, 2025

• Enable TMA by default. This will allow TMA to be considered as an option for codegen that goes through the generic Triton emitter.

  #97998 opened Jul 31, 2025

• [XLA:GPU] Make debugging easier when the all-reduce test fails by logging FP32 values.

  #97999 opened Jul 31, 2025

• Fix typos in documentation strings

  #98001 opened Jul 31, 2025

• Create SymbolicExpr canonicalization tests

  #98004 opened Jul 31, 2025

• Automated Code Change

  #98005 opened Jul 31, 2025

• Use `LRUCache::ForEach` in jaxlib.

  #98013 opened Jul 31, 2025

• Adding HBM derate curve for H100 in dot fusion cost model

  #98015 opened Jul 31, 2025

• Set shardy for tests using jax models

  #98017 opened Jul 31, 2025

• [XLA] fix handling of infeed/outfeed sharding when converting mhlo to hlo

  #98020 opened Jul 31, 2025

• Fix msan use-of-uninitialized-value / asan stack-use-after-return error in command_buffer_conversion_pass_test.

  #98021 opened Jul 31, 2025

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

  #98023 opened Jul 31, 2025

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

  #98024 opened Jul 31, 2025

• Only call on_done_with_host_buffer after the buffer is copied to device.

  #98025 opened Jul 31, 2025

• Make sure that we permit all semantically correct and lowerable conv fusions without caring for costs or env variables set for must-fuse cases.

  #98026 opened Jul 31, 2025

• Add `inline_shardy_manual_computation` in `CallInliner`.

  #98028 opened Jul 31, 2025

• Rename split test utils and flags to be more general.

  #98029 opened Jul 31, 2025

• Ensure that we only enable checks related to lowering for nested fusion cases of user directed (e.g. must fuse) fusions.

  #98030 opened Jul 31, 2025

• Integrate LLVM at llvm/llvm-project@652048ad2578

  #98033 opened Jul 31, 2025

• Enable lowering/legality checks for user-guided fusions in TPU instruction fusion decision process.

  #98036 opened Aug 1, 2025

• Update fusion pipeline in deepsea compiler to prioritize user-guided fusion above all other types of fusions and ensure that they form.

  #98037 opened Aug 1, 2025

• Automated Code Change

  #98038 opened Aug 1, 2025

• Automated Code Change

  #98039 opened Aug 1, 2025

• Automated Code Change

  #98040 opened Aug 1, 2025

• Automated Code Change

  #98041 opened Aug 1, 2025

• Automated Code Change

  #98043 opened Aug 1, 2025

• Automated Code Change

  #98044 opened Aug 1, 2025

• Automated Code Change

  #98054 opened Aug 1, 2025

• XLA: fix include path

  #98062 opened Aug 1, 2025

• Fix windows compilation

  #98063 opened Aug 1, 2025

• Bump ubuntu from `01a3ee0` to `1ec65b2` in /tensorflow/tools/tf_sig_build_dockerfiles

  #98064 opened Aug 1, 2025

• Bump ubuntu from `b59d215` to `a08e551` in /tensorflow/tools/gcs_test

  #98065 opened Aug 1, 2025

• Bump the github-actions group with 2 updates

  #98066 opened Aug 1, 2025

• Automated Code Change

  #98068 opened Aug 1, 2025

• Upgrade the default cuDNN version to v9.10.0

  #98069 opened Aug 1, 2025

• Automated Code Change

  #98070 opened Aug 1, 2025

• Automated Code Change

  #98071 opened Aug 1, 2025

• Automated Code Change

  #98072 opened Aug 1, 2025

• Automated Code Change

  #98073 opened Aug 1, 2025

• Automated Code Change

  #98074 opened Aug 1, 2025

• [XLA:CPU] Support dumping snapshots with unoptimized HLO.

  #98082 opened Aug 1, 2025

• Build system cleanup for xla/backends/profiler/gpu

  #98083 opened Aug 1, 2025

• [XLA] Add a HLO pass, `HloIdentityComputationRemover`, to remove identify computations.

  #98084 opened Aug 1, 2025

• Silence a particularly noisy warning from clang:

  #98085 opened Aug 1, 2025

• Automated Code Change

  #98087 opened Aug 1, 2025

• PR #26391: [NVIDIA GPU] Dynamic SPMD iteration limit for larger fast-interconnect domain

  #98088 opened Aug 1, 2025

• [XLA:GPU] Enable TMA for softmax.

  #98092 opened Aug 1, 2025

• [Autotuner] Extend profiler to check buffers used for profiling.

  #98093 opened Aug 1, 2025

• Adding test for non-flat while computations and shared bodies and conditions.

  #98095 opened Aug 1, 2025

• [XLA:MSA] Speed up FixSchedule by iterating over instructions once.

  #98097 opened Aug 1, 2025

• [XLA:MSA] Speed up Evict() by combining computations across retries.

  #98098 opened Aug 1, 2025

• [MLIR<->HLO] Split off XLA Dialect Registration

  #98105 opened Aug 1, 2025

• Update rules_ml_toolchain dependency to latest version.

  #98106 opened Aug 1, 2025

• Fix: Critical Python Bugs, Type Safety, and Named Tuple Issue

  #98108 opened Aug 1, 2025

• [HLO Translate] Add sdy to registered MLIR->HLO dialects.

  #98109 opened Aug 1, 2025

• [XLA] Moving helper methods out from reduce window rewriter

  #98110 opened Aug 1, 2025

• debug_options_parsers_test: Don't assume the same object is reused.

  #98114 opened Aug 1, 2025

• Integrate LLVM at llvm/llvm-project@fa6965f722e0

  #98117 opened Aug 1, 2025

• remove dequantize from ReorderBroadcastToAndUnaryOp

  #98118 opened Aug 2, 2025

• [XLA] Defer computation removal in HloModule to avoid iterator invalidation issues.

  #98119 opened Aug 2, 2025

• Set L2Norm sum reduce axis to true

  #98120 opened Aug 2, 2025

• rename slice_index in gpu topology to partition_index

  #98123 opened Aug 2, 2025

• Integrate LLVM at llvm/llvm-project@fa6965f722e0

  #98124 opened Aug 2, 2025

• Integrate LLVM at llvm/llvm-project@3d9944680980

  #98125 opened Aug 2, 2025

• Automated Code Change

  #98127 opened Aug 2, 2025

• Allow propagation of per-tensor QSV across broadcast

  #98128 opened Aug 2, 2025

• Add new CachingHloEvaluator decorator.

  #98130 opened Aug 2, 2025

• Add split-phase compilation support to interpreter reference mixin.

  #98131 opened Aug 2, 2025

• Automated Code Change

  #98132 opened Aug 2, 2025

• Automated Code Change

  #98133 opened Aug 2, 2025

• Automated Code Change

  #98134 opened Aug 2, 2025

• rename `slice_index` in `partition_index` gpu topology

  #98135 opened Aug 2, 2025

• Automated Code Change

  #98136 opened Aug 2, 2025

• Automated Code Change

  #98137 opened Aug 2, 2025

• Automated Code Change

  #98138 opened Aug 2, 2025

• Automated Code Change

  #98139 opened Aug 2, 2025

• Automated Code Change

  #98140 opened Aug 2, 2025

• Automated Code Change

  #98141 opened Aug 2, 2025

• Automated Code Change

  #98142 opened Aug 2, 2025

• Automated Code Change

  #98143 opened Aug 2, 2025

• Automated Code Change

  #98144 opened Aug 2, 2025

• Automated Code Change

  #98145 opened Aug 2, 2025

• Automated Code Change

  #98146 opened Aug 2, 2025

• Automated Code Change

  #98147 opened Aug 2, 2025

• Automated Code Change

  #98148 opened Aug 2, 2025

• Automated Code Change

  #98149 opened Aug 2, 2025

• Automated Code Change

  #98150 opened Aug 2, 2025

• Automated Code Change

  #98151 opened Aug 2, 2025

• Automated Code Change

  #98152 opened Aug 2, 2025

• Automated Code Change

  #98153 opened Aug 2, 2025

• Automated Code Change

  #98154 opened Aug 2, 2025

• Automated Code Change

  #98155 opened Aug 2, 2025

• [XLA:CPU] Begin conversion of F8 float conversions to xla intrinsics.

  #98156 opened Aug 2, 2025

• Automated Code Change

  #98157 opened Aug 2, 2025

• [XLA] Add a HLO pass, HloIdentityComputationRemover, to remove identify computations.

  #98158 opened Aug 2, 2025

• Automated Code Change

  #98159 opened Aug 3, 2025

• Automated Code Change

  #98160 opened Aug 3, 2025

• Automated Code Change

  #98161 opened Aug 3, 2025

• Automated Code Change

  #98162 opened Aug 3, 2025

• Automated Code Change

  #98163 opened Aug 3, 2025

• Automated Code Change

  #98164 opened Aug 3, 2025

• Automated Code Change

  #98165 opened Aug 3, 2025

• Automated Code Change

  #98167 opened Aug 3, 2025

• Automated Code Change

  #98168 opened Aug 3, 2025

• Automated Code Change

  #98169 opened Aug 3, 2025

• Automated Code Change

  #98170 opened Aug 3, 2025

• Automated Code Change

  #98171 opened Aug 3, 2025

• Automated Code Change

  #98172 opened Aug 3, 2025

• Make dot_decomposer handle zero dimensions properly

  #98173 opened Aug 3, 2025

• Automated Code Change

  #98174 opened Aug 3, 2025

• Clean up xnn_fusion_test.

  #98175 opened Aug 3, 2025

• Allow de-annotation of unsupported scheduling groups

  #98176 opened Aug 3, 2025

• Skip call_library_for_dot for non-dot ops.

  #98177 opened Aug 3, 2025

• [XLA:GPU][Tiling] Port symbolic_tile_analysis to the new tile format.

  #98179 opened Aug 3, 2025

• Add a couple of flag descriptions

  #98180 opened Aug 3, 2025

15 Issues closed by 5 people

• Core Dump When Training

  #97016 closed Aug 1, 2025

• TFLite Interpreter `experimental_preserve_all_tensors` yields different output

  #79355 closed Jul 31, 2025

• TFLite Libraries Do Not Yet Support 16KB Page Sizes

  #97935 closed Jul 31, 2025

• DLL load failed while importing _pywrap_tensorflow_internal: A dynamic link library (DLL) initialization routine failed.

  #97812 closed Jul 30, 2025

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

  #60710 closed Jul 30, 2025

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

  #60714 closed Jul 30, 2025

• Incorrect gradient in divide_no_nan and reciprocal_no_nan when divide by 0

  #60715 closed Jul 30, 2025

• failed to build branch r2.13

  #60716 closed Jul 30, 2025

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

  #62191 closed Jul 30, 2025

• Remove or update zh-cn translation from installation instructions

  #62245 closed Jul 30, 2025

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

  #62441 closed Jul 30, 2025

• Error: 'class tensorflow::tools::proto_splitter::SavedModelSplitter' has no member named 'WriteToCord'

  #72108 closed Jul 29, 2025

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

  #95235 closed Jul 29, 2025

• GPU Not Detected by TensorFlow Despite Proper System Setup

  #96707 closed Jul 29, 2025

• Tried downgrading to other versions of Tensorflow such as 2.18.0, 2.17.0, 2.16.2 etc since the latest 2.19.0 was not working, though these versions dont seem to have the Keras module in them , infact it should for all

  #97620 closed Jul 28, 2025

10 Issues opened by 10 people

• Installation on Grace Hopper H200 server is not GPU-enabled

  #98115 opened Aug 1, 2025

• Bug in GPU delegate output tensor extraction when Split node is included

  #98103 opened Aug 1, 2025

• EXC_BAD_ACCESS when using GemmImplUsingEigen in MacOS

  #98002 opened Jul 31, 2025

• Issue while converting a model with GRU to .tflite

  #97941 opened Jul 31, 2025

• Precision issues occur when using tf.segment_maxwith floating-point tensors of certain dimensions.

  #97805 opened Jul 30, 2025

• `tf.linalg.eigh` faces OOM on CUDA with 6D tensor and raises SIGABRT

  #97780 opened Jul 29, 2025

• Memory leak in custom_gradient decorator

  #97697 opened Jul 28, 2025

• Add option for deterministic tie-breaking in tf.math.reduce_min

  #97688 opened Jul 28, 2025

• Mismatch Between Quantized TFLite Layer Outputs and Expected Mathematical Values

  #97677 opened Jul 28, 2025

• tf.image.combined_non_max_suppression crashed with unmatched scores

  #97672 opened Jul 28, 2025

52 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 type-safe Go session config API, docs, and tests

  #97479 commented on Jul 29, 2025 • 11 new comments

• Fix: Clarify GPU compatibility check in TFLite Model Analyzer

  #97513 commented on Jul 29, 2025 • 2 new comments

• Automated Code Change

  #96862 commented on Jul 31, 2025 • 0 new comments

• Automated Code Change

  #96905 commented on Jul 31, 2025 • 0 new comments

• Automated Code Change

  #96912 commented on Jul 31, 2025 • 0 new comments

• Automated Code Change

  #96924 commented on Jul 31, 2025 • 0 new comments

• Automated Code Change

  #96931 commented on Jul 31, 2025 • 0 new comments

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

  #97118 commented on Jul 29, 2025 • 0 new comments

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

  #97260 commented on Jul 28, 2025 • 0 new comments

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

  #97340 commented on Jul 31, 2025 • 0 new comments

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

  #97348 commented on Jul 31, 2025 • 0 new comments

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

  #97406 commented on Jul 28, 2025 • 0 new comments

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

  #97407 commented on Jul 29, 2025 • 0 new comments

• MPMD XLA bump

  #97457 commented on Jul 29, 2025 • 0 new comments

• Add option to enable coordination service client to recover.

  #97460 commented on Aug 1, 2025 • 0 new comments

• Add environment variable to enable TFRT GPU client in GPU plugin.

  #97461 commented on Aug 1, 2025 • 0 new comments

• Add option to enable TFRT GPU client in C API plugin.

  #97462 commented on Aug 1, 2025 • 0 new comments

• Add option to enable coordination service client to recover.

  #97463 commented on Aug 1, 2025 • 0 new comments

• Fix: Remove dangling reference to ExperimentalTFLiteToTosaBytecode

  #97508 commented on Jul 28, 2025 • 0 new comments

• Pipe incarnations through C plugin.

  #97540 commented on Jul 28, 2025 • 0 new comments

• Pallas/Mosaic support host output.

  #97542 commented on Jul 28, 2025 • 0 new comments

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

  #97581 commented on Jul 30, 2025 • 0 new comments

• Use sdy front end attributes when comparing instruction identicalness

  #97585 commented on Jul 30, 2025 • 0 new comments

• [IFRT] Use compile options to check if sdy partitioned

  #97586 commented on Jul 28, 2025 • 0 new comments

• Integrate StableHLO at openxla/stablehlo@e07debd5

  #97588 commented on Aug 1, 2025 • 0 new comments

• Add a quantized folder for reshape.

  #97592 commented on Jul 30, 2025 • 0 new comments

• The model does not have any trainable weights.

  #97599 commented on Jul 28, 2025 • 0 new comments

• `tf.raw_ops.QuantizedConv2D` crashes with assertion failure instead of proper error handling

  #94045 commented on Jul 28, 2025 • 0 new comments

• xprof compilation fails with gcc 14.2

  #94035 commented on Jul 28, 2025 • 0 new comments

• [RNN] Conversion of model containing GRU layer to quantized TFLite causes Segmentation Fault

  #87468 commented on Jul 28, 2025 • 0 new comments

• tf.no_op DAG cycle issue

  #97512 commented on Jul 28, 2025 • 0 new comments

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

  #97096 commented on Jul 28, 2025 • 0 new comments

• TensorFlow Java documentation is outdated

  #96799 commented on Jul 29, 2025 • 0 new comments

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

  #97042 commented on Jul 29, 2025 • 0 new comments

• Mixed precision results in lower performance on AMD GPUs

  #97645 commented on Jul 29, 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 30, 2025 • 0 new comments

• TensorFlow disables SwiftUI Previews

  #95106 commented on Jul 30, 2025 • 0 new comments

• It doesn't support on python3.13

  #78774 commented on Jul 30, 2025 • 0 new comments

• Create a trainable tensorflow or LiteRT (with signatures) graph from a frozen tensorflow model

  #79377 commented on Jul 31, 2025 • 0 new comments

• Cannot Convert 51 to a shape - Movenet pose classification tutorial

  #79435 commented on Jul 31, 2025 • 0 new comments

• lib new version not support 16kb pages in android

  #96602 commented on Jul 31, 2025 • 0 new comments

• Inconsistent Error Handling in `tf.raw_ops.SparseSegmentSumGradV2` Between CPU and GPU Implementations

  #94151 commented on Jul 31, 2025 • 0 new comments

• Enhance Memory Optimizer with Dynamic Cost Model for Operation Recomputation

  #94653 commented on Aug 3, 2025 • 0 new comments

• Always lower ragged dot for cpu, gpu, and tpu

  #96631 commented on Jul 28, 2025 • 0 new comments

• #sdy Remove MHLO shardings from round-trip export

  #96640 commented on Jul 31, 2025 • 0 new comments

• Automated Code Change

  #96811 commented on Jul 31, 2025 • 0 new comments

• Automated Code Change

  #96838 commented on Jul 31, 2025 • 0 new comments

• Automated Code Change

  #96839 commented on Jul 31, 2025 • 0 new comments

• Automated Code Change

  #96840 commented on Aug 1, 2025 • 0 new comments

• Automated Code Change

  #96841 commented on Aug 1, 2025 • 0 new comments

• Automated Code Change

  #96851 commented on Jul 31, 2025 • 0 new comments

• Automated Code Change

  #96856 commented on Jul 31, 2025 • 0 new comments