[pull] master from tensorflow:master#2374

Open

pull[bot] wants to merge 10000 commits intoxiaohanhuang:masterfrom

tensorflow:master

This pull request is big! We're only showing the most recent 250 commits

Commits on Apr 17, 2026

[XLA:GPU] Remove Split-K from Triton Autotuner and Search Space.

mooskagh
authored and
tensorflower-gardener
committed
Fix handling of auto layouts in PJRT Compiler

beckerhe
authored and
tensorflower-gardener
committed
[XLA:GPU] Remove Split-K GEMM rewriter

mooskagh
authored and
tensorflower-gardener
committed
[XLA:GPU] Only run python_hlo_runner_test on GPU backends where it passes.

akuegel
authored and
tensorflower-gardener
committed
Tokens are buffer-less, and so should be skipped in nanort where we operate on buffers.
tensorflower-gardener
committed
Update log statement to clearly specify intent if we receive `HloShardingV2` instead of V3 during Shardy import when flag is enabled.

KanishAnand
authored and
tensorflower-gardener
committed
[XLA:GPU]: Reattempt switching on one-shot all-reduce.

sohaibiftikhar
authored and
tensorflower-gardener
committed
[XLA] Serialize BufferAllocation peak buffers in buffer assignment
tensorflower-gardener
committed
Integrate Triton up to [6a0b5461eb](https://github.com/triton-lang/triton/commits/6a0b5461eb439bdb3ef0c721837c64dc2c345fcb)
tensorflower-gardener
committed
Rename rms_norm_benchmark_test to norm_benchmark_test

dsharletg
authored and
tensorflower-gardener
committed
Fix `DebugOptions` flag parsing to respect initial proto values

KanishAnand
authored and
tensorflower-gardener
committed
Move EarlyExitCompilationResult to its own files.

beckerhe
authored and
tensorflower-gardener
committed
GHA: profile upload stability improvements

dmiltr3
authored and
tensorflower-gardener
committed
[XLA][cleanup]: Use absl::string_view in xla_builder.cc

Varcho
authored and
tensorflower-gardener
committed
[XLA:TPU] Speedup UpdateComputationSchedule.
tensorflower-gardener
committed
[XLA] Use HloCloneContext suffix for cloned computation when available
tensorflower-gardener
committed
convert FusionInterface to return AsyncThunkSequence

ermilovmaxim
authored and
tensorflower-gardener
committed
[XLA:GPU] Log ptx contents that produce a warning at VLOG(2)
tensorflower-gardener
committed
Plug a memory leak.
tensorflower-gardener
committed
This change introduces a `SetEnviron` method to the `tsl::SubProcess` class, enabling users to specify a custom environment map for the child process.
tensorflower-gardener
committed
In `ReadFileToString`, allow files to report length longer than their actual contents

ddunl
authored and
tensorflower-gardener
committed
[IFRT IR] Add a method that the IFRT IR compiler that only creates a compiled program, but not an executable.

ICGog
authored and
tensorflower-gardener
committed
PR #41127: [xla:gpu] Extract nccl_types library

ezhulenev
authored and
tensorflower-gardener
committed

Commits on Apr 18, 2026

Commits on Apr 19, 2026

Remove a typo

amitsabne1
authored and
tensorflower-gardener
committed

Commits on Apr 20, 2026

Commits on Apr 21, 2026

Make LLVMCompiler hooks private and thread-safe

ermilovmaxim
authored and
tensorflower-gardener
committed
[XLA] Improve conditional code motion for broadcasts to avoid performance and memory issues.

blakehechtman
authored and
tensorflower-gardener
committed
[IFRT IR] Keep delete_options and device_list for IFRT IR programs

ICGog
authored and
tensorflower-gardener
committed
Add axis count check in Mean op.

lynne0326
authored and
tensorflower-gardener
committed
Remove redundant A - A => 0 rule

dsharletg
authored and
tensorflower-gardener
committed
Automated Code Change
tensorflower-gardener
committed
Automated Code Change
tensorflower-gardener
committed
Automated Code Change
tensorflower-gardener
committed
Automated Code Change
tensorflower-gardener
committed
Automated Code Change
tensorflower-gardener
committed
Automated Code Change
tensorflower-gardener
committed
Automated Code Change
tensorflower-gardener
committed
Enable layering_check in XLA builds.

beckerhe
authored and
tensorflower-gardener
committed
PR #41174: Limit GetSymbolicTileAnalysis to low-rank tensors

dignissimus
authored and
tensorflower-gardener
committed
PR #41131: [xla:gpu] Pass non-const collective cliques and memory to FFI

ezhulenev
authored and
tensorflower-gardener
committed
PR #41055: Cross Host Data Transfers: Add CrossHostTransferBuffers; modify CrossHostSendBuffers to call into it

rao-ashish
authored and
tensorflower-gardener
committed
Add test to generate the CUB sort scratch size table data

EusebioDM
authored and
tensorflower-gardener
committed
Cherry-pick https://github.com/triton-lang/triton/pull/10032 which reverts a change that causes a crash in getTMABlockShape

vwbaker
authored and
tensorflower-gardener
committed
Create CLI tool to generate CUB scratch size data

EusebioDM
authored and
tensorflower-gardener
committed
Make sure we don't re-allocate a `scale` that is already the correct size (e.g. channel-wise quantization with `output_channels == 1`).

gonnet
authored and
tensorflower-gardener
committed
[XLA:GPU] Minor improvements

dvadym
authored and
tensorflower-gardener
committed
[XLA:GPU] Update default Triton configs for A100.

mooskagh
authored and
tensorflower-gardener
committed
PR #41225: [xla:gpu] Delete dynamic slice commands

ezhulenev
authored and
tensorflower-gardener
committed
When considering warp specialization for Triton dot kernels, do not enable it if num_stages is equal to 2. This is currently causing CUDA_ERROR_MISALIGNED_ADDRESS on kernels that have a broadcast c…

Moerafaat
authored and
tensorflower-gardener
committed
Drop dependencies on xla::sdy from shardy flattenner.

ekayaaslan
authored and
tensorflower-gardener
committed
[XLA:GPU] Remove Split-K from Triton Fusion Analysis and Tiling Propagation

mooskagh
authored and
tensorflower-gardener
committed
[XLA:GPU] Fix duplicate main function in `cudnn_fused_conv_rewriter_test`.

penpornk
authored and
tensorflower-gardener
committed
PR #40202: [XLA:GPU][oneAPI][Bugfix] Fix unused globals for SPIR-V backend

mdfaijul
authored and
tensorflower-gardener
committed
PR #41293: [ROCm] Add cuda-only tags for cuda specific tests

hsharsha
authored and
tensorflower-gardener
committed
[XLA:GPU] Move buffers and buffer_uses to CollectiveThunk base class.

PatriosTheGreat
authored and
tensorflower-gardener
committed
[XLA:GPU] Fix the offset evaluation when there are no dims.

pifon2a
authored and
tensorflower-gardener
committed
Move shardy flattener from xla::sdy to the mlir::sdy.

ekayaaslan
authored and
tensorflower-gardener
committed
[XLA:GPU] Move two-side NCCL collectives to symmetric memory.

PatriosTheGreat
authored and
tensorflower-gardener
committed
PR #41194: [xla:gpu] Migrate CustomKernelLaunchCmd into CustomKernelThunk

shawnwang18
authored and
tensorflower-gardener
committed
Run shardy/google/integrate_latest.sh to integrate.

ekayaaslan
authored and
tensorflower-gardener
committed
[CublasLt] Enable __cublas$lt$matmul in dynamic slice fusion rewriter.

derdrdirk
authored and
tensorflower-gardener
committed
PR #41229: [ROCm] Use rocprim::segmented_inclusive_scan for batched row scan

magaonka-amd
authored and
tensorflower-gardener
committed
[XLA:BUILD]: Add a clang-tidy config for OSS

sohaibiftikhar
authored and
tensorflower-gardener
committed
Add --force_auto_layout to xla_compile.

khasanovaa
authored and
tensorflower-gardener
committed
PR #41205: [XLA:GPU] Migrate CuDnnCmd into CuDnnThunk

shawnwang18
authored and
tensorflower-gardener
committed
Reverts 97f1b3fb5c228db2398e518e1f434db832e11b75

terryysun
authored and
tensorflower-gardener
committed
[XLA] Rename fields in CudnnVersion and ComputeCapability to avoid clashes with system macros.

derdrdirk
authored and
tensorflower-gardener
committed
PR #41304: [ROCm] Fix sorting.hlo lit test on GCN

hsharsha
authored and
tensorflower-gardener
committed
[XLA] Allow "textproto" as an alias for "pbtxt" in HLO module loader.

mkuperst
authored and
tensorflower-gardener
committed
Only set replica_count and num_partitions in HLO config if they are > 1.

khasanovaa
authored and
tensorflower-gardener
committed
[XLA:GPU] Add test-cases which are using symmetric buffers.

PatriosTheGreat
authored and
tensorflower-gardener
committed
[NFC] Use `std::array` for fixed-size buffer in `GetDeviceName`.

allanrenucci
authored and
tensorflower-gardener
committed
Add overflow checks to convolution kernels' Prepare functions.

snnn
authored and
tensorflower-gardener
committed
Update tests to use __cublas$lt$matmul.

derdrdirk
authored and
tensorflower-gardener
committed
Add rules to cancel the same value from subtractions when there is another add/subtract that confounds the `x - x` => `0` rule

dsharletg
authored and
tensorflower-gardener
committed
[XLA:CPU] Do not hoist small while loop with custom fusions.

penpornk
authored and
tensorflower-gardener
committed
PR #41260: [xla:gpu] Migrate LegacyCustomCallCmd into LegacyCustomCallThunk

shawnwang18
authored and
tensorflower-gardener
committed
PR #36806: [XLA:CPU][oneDNN] Refactor oneDNN Resources in oneDNN Thunk

othakkar
authored and
tensorflower-gardener
committed
implement optional threadpool support in CubinCustomKernelCompiler

ermilovmaxim
authored and
tensorflower-gardener
committed
Automated Code Change

derekmauro
authored and
tensorflower-gardener
committed
[XLA::CPU] Update Gemma Keras end-to-end benchmarks to support multiple model versions.

vksnk
authored and
tensorflower-gardener
committed
PR #38198: [oneDNN] Upgrade to oneDNN v3.11

othakkar
authored and
tensorflower-gardener
committed
[XLA:GPU] Move one-side collectives to symmetric memory and remove MaybeRegisterBuffers API

PatriosTheGreat
authored and
tensorflower-gardener
committed
Reverts 1222cfb07c7faccfd2ce184f1ba2f700258fff2d
tensorflower-gardener
committed
Propagate `eval_literal_handler` during HloEvaluator cloning.
tensorflower-gardener
committed
PR #40316: [Nvidia GPU] Do aliasing only when collective permute is in-place

Tixxx
authored and
tensorflower-gardener
committed
Always enable `strict_shape_checking` for HloRunnerPjRt.

nvgrw
authored and
tensorflower-gardener
committed
Add checks for unsupported sharding types in ToNamedSharding.

Kingston Mandisodza
authored and
tensorflower-gardener
committed
Integrate SelectKThunk into the ThunkProto deserialization

apivovarov
authored and
tensorflower-gardener
committed
Update custom_gradient library from eager to lazy importing.
tensorflower-gardener
committed
[XLA:GPU] Temporarily skip testing conversions to/from F8E4M3FN on NVIDIA GPUs older than Ada.

penpornk
authored and
tensorflower-gardener
committed
Always enable `strict_shape_checking` for HloRunnerPjRt.
tensorflower-gardener
committed
Reverts 4498b8721ff17a3b2cffd9e551141949c41baadd

mwhittaker
authored and
tensorflower-gardener
committed
Fix typo in comment in kernel_fallback_compat_request_state.h (falllback -> fallback).

ch1nlu
authored and
tensorflower-gardener
committed
Remove unused `TransactionToken` and associated methods from `tsl::FileSystem`

ddunl
authored and
tensorflower-gardener
committed
Integrate StableHLO at openxla/stablehlo@9b278496

mrguenther
authored and
tensorflower-gardener
committed
#tf-data Add VLOG to the take dataset counter.

yangustc07
authored and
tensorflower-gardener
committed
Update to batch task split.
tensorflower-gardener
committed
PR #40750: Add overflow checks in shape inference dimension computations

mohammadmseet-hue
authored and
tensorflower-gardener
committed
[XLA:BufferAssignment] Add FastMerge and FastSplit buffer assignment algorithms.
tensorflower-gardener
committed

Commits on Apr 22, 2026

Commits on Apr 23, 2026