[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 24, 2026

Rollback CublasLt deprecation.

derdrdirk
authored and
tensorflower-gardener
committed
Automated Code Change
tensorflower-gardener
committed
PR #41484: [XLA:GPU][oneAPI] Skip event wait for completed event in SyclEvent::WaitStreamOnEvent

bhavani-subramanian
authored and
tensorflower-gardener
committed
PR #41380: Handle X.Y CUDA driver versions.

olupton
authored and
tensorflower-gardener
committed
PR #41327: Remove no_oss tags

mmakevic-amd
authored and
tensorflower-gardener
committed
Move CublasPadForGemms to run after GemmFusion (and just before the actual cublas rewrite)

vwbaker
authored and
tensorflower-gardener
committed
Reverts 5000c713c57501c14074ea93eb1174eff883117b

beckerhe
authored and
tensorflower-gardener
committed
Add check to not call `SubgroupReplicationDim` with `HloShardingV3`

KanishAnand
authored and
tensorflower-gardener
committed
PR #41483: [ROCm] Fix leading comma in AMDGPU feature string

Ruturaj4
authored and
tensorflower-gardener
committed
Rollback CublasLt deprecation.

derdrdirk
authored and
tensorflower-gardener
committed
[XLA:GPU] Fix collective permute tests.

PatriosTheGreat
authored and
tensorflower-gardener
committed
PR #41175: [PJRT & GPU] implement `PJRT_Device_ClearMemoryStats`

fl0pedro
authored and
tensorflower-gardener
committed
PR #41481: [XLA:GPU] Insert ReshapeDecomposer before post-GemmRewriter LayoutNormalization

magaonka-amd
authored and
tensorflower-gardener
committed
PR #41167: Fix shape-dependent erf precision by adding saturation select

kredd2506
authored and
tensorflower-gardener
committed
PR #41504: [xla] Don't forget memory copies for custom calls with collective operands

ezhulenev
authored and
tensorflower-gardener
committed
[XLA:GPU] Improve layout assignment for reshapes.

akuegel
authored and
tensorflower-gardener
committed
PR #41528: Bump bant to v0.2.6 in DWYU job

nurmukhametov
authored and
tensorflower-gardener
committed
[NFC] Remove intermediate `std::vector` allocation.

allanrenucci
authored and
tensorflower-gardener
committed
[XLA:BufferAssignment] Refactor buffer assignment algorithm selection.
tensorflower-gardener
committed
Internal changes in build system.

PatriosTheGreat
authored and
tensorflower-gardener
committed
[XLA:GPU] Make StableSortExpander drop is_stable flag after expanding.

mooskagh
authored and
tensorflower-gardener
committed
[XLA:GPU] Remove usages of split_k from Triton configurations.

mooskagh
authored and
tensorflower-gardener
committed
PR #41515: [xla:gpu] Resolve multimem from SymmetricMemory

ezhulenev
authored and
tensorflower-gardener
committed
Convert unreduced and manual shardings in the format picker, and parameterise a whole lot more tests.

Kingston Mandisodza
authored and
tensorflower-gardener
committed
Internal build change.

IllogicalMoose
authored and
tensorflower-gardener
committed
[XLA:GPU] Fix the test with old devices.

PatriosTheGreat
authored and
tensorflower-gardener
committed
[XLA:GPU] Update XLABuilder::CollectivePermute docs

dimitar-asenov
authored and
tensorflower-gardener
committed
Adds support for named shardings in PatternMatchMergeOrSplitSharding.

Kingston Mandisodza
authored and
tensorflower-gardener
committed
Move rules_ml_toolchain loading from WORKSPACE and workspace0.bzl to workspace3.bzl

yuriivcs
authored and
tensorflower-gardener
committed
Fix lint errors in pjrt/c/CHANGELOG.md.

pschuh
authored and
tensorflower-gardener
committed
Introduce GpuCompiler::AsmHook and utilize Compiler hooks in hlo_opt

ermilovmaxim
authored and
tensorflower-gardener
committed
[XLA:TPU] Cache shape size in CanAllocateInAlternateMemoryNoCopy.
tensorflower-gardener
committed
Clear layout in xla::Shape if not present

hhb
authored and
tensorflower-gardener
committed
Move device events to AbstractTrackedDeviceBuffer to prepare for removing implementation specific versions of TrackedDeviceBuffer.

pschuh
authored and
tensorflower-gardener
committed
Integrate LLVM at llvm/llvm-project@bb092120f1f0
tensorflower-gardener
committed
Make the call to async_value() never return an asnyc value if it is used across the plugin boundary.

pschuh
authored and
tensorflower-gardener
committed
Add deadline and cancellation support to BatchTask.

ch1nlu
authored and
tensorflower-gardener
committed
Refactor vhlo::TensorV1Attr to store DenseElementsAttr
tensorflower-gardener
committed

Commits on Apr 25, 2026

Commits on Apr 27, 2026

Add additional rewritable parameters in the BatchFunction op
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
Update llvm BUILD patch to add linkopts to avoid link error on Windows.

akuegel
authored and
tensorflower-gardener
committed
PR #41594: [XLA:GPU] Add ncclGetLsaMultimemDevicePointer to nccl.symbols

sfvaroglu
authored and
tensorflower-gardener
committed
[XLA:GPU] Add nozapfhahn tag to tests that time out in coverage builds.

akuegel
authored and
tensorflower-gardener
committed
PR #41557: [XLA:GPU][oneAPI][BugFix] Fix oneAPI presubmit build

akhilgoe
authored and
tensorflower-gardener
committed
Convert to v2 for `GroupShardingOnDims` fallback in `ReshardNoCache`.

Kingston Mandisodza
authored and
tensorflower-gardener
committed
[XLA:MSA] Refactor some code in separate functions.

subhankarshah
authored and
tensorflower-gardener
committed
Introduce focused optimizations to reduce XLA HLO evaluator latency:

jbspooner
authored and
tensorflower-gardener
committed
[XLA:GPU] Add API to check if device address was allocated with VMM.

PatriosTheGreat
authored and
tensorflower-gardener
committed
Reverts 2863c6b6d737e9d1f79a4083a37888bf1e5ffc14

akuegel
authored and
tensorflower-gardener
committed
[XLA:GPU] Remove unused combination generation helpers from block level emitter

Moerafaat
authored and
tensorflower-gardener
committed
Optimize computation lookup in BufferAssignment.

liepieshov
authored and
tensorflower-gardener
committed
Extract block_k from TritonGemmConfig for cost model.
tensorflower-gardener
committed
Avoid crashing on unsupported bitcasts in Triton tiling propagation.

vwbaker
authored and
tensorflower-gardener
committed
[XLA:GPU] Add API to get peer address of symmetrical memory on host.

PatriosTheGreat
authored and
tensorflower-gardener
committed
Parameterize `HloExpandTest` with `HloShardingV3`

KanishAnand
authored and
tensorflower-gardener
committed
Convert shardings from V3 to V2 in `ReshardPartialReplicateWithAllToAll`

Kingston Mandisodza
authored and
tensorflower-gardener
committed
Remove ShouldTritonHandleGEMM, left over from previous logic for cublas padding which was removed in cl/904886332.

vwbaker
authored and
tensorflower-gardener
committed
Propagate manual axes to called functions during shard map import.

ekayaaslan
authored and
tensorflower-gardener
committed
Split EstimateCubScratchSpace into a pass for sort and one for scan

EusebioDM
authored and
tensorflower-gardener
committed
Use `constexpr absl::string_view` instead of `const std::string` as per go/totw/170

KanishAnand
authored and
tensorflower-gardener
committed
Add comment to StreamExecutor::IsVmmMemory.

PatriosTheGreat
authored and
tensorflower-gardener
committed
[XLA] Support reading/writing HloProto with the SplitProtoReader.

pifon2a
authored and
tensorflower-gardener
committed
Add ncclGetLsaDevicePointer to NCCL symbols.

PatriosTheGreat
authored and
tensorflower-gardener
committed
Push shardy flattener down past ShardMapImport pass.

ekayaaslan
authored and
tensorflower-gardener
committed
[XLA] Change HloReachabilityMap::Key to use only the instruction ID.

mkuperst
authored and
tensorflower-gardener
committed
Add flag to ScanExpander to control expansion of associative scans

patrick-toulme
authored and
tensorflower-gardener
committed
Sort backend configs in executable HLO text.

belitskiy
authored and
tensorflower-gardener
committed
Internal visibility only change.
tensorflower-gardener
committed
PR #39604: [NVIDIA GPU] Added annotation to allow scheduling of custom communication kernels

Tixxx
authored and
tensorflower-gardener
committed
Fix predecessor finding for conditionals reusing computations.
tensorflower-gardener
committed
[GPU] Add Request[Symmetric|Multicast|Peer]AllocationSlice API to collective_memory_requests.h

ezhulenev
authored and
tensorflower-gardener
committed
Change AbstractTrackedDeviceBuffer to manage usage events through a

pschuh
authored and
tensorflower-gardener
committed
Reverts 88fb196aa35d4ff64a317ebfdf8665c0d728838e

belitskiy
authored and
tensorflower-gardener
committed
This CL introduce a new tie breaker for two async-done:

deqiangc
authored and
tensorflower-gardener
committed
Support HloOpcode::kScan in buffer_assignment and layout_assignment. Note these are for non decomposed scans.

patrick-toulme
authored and
tensorflower-gardener
committed
xla_runner: manually register the profiler for dynamically registered PJRT plugins

abhigunj
authored and
tensorflower-gardener
committed
[XLA] Avoid calling size_bytes_dense() in the inner loop of Literal::Equal.

mkuperst
authored and
tensorflower-gardener
committed
Add an empty dataset to `testTakeWithRestartedWorker`.

yangustc07
authored and
tensorflower-gardener
committed
Internal build change.

IllogicalMoose
authored and
tensorflower-gardener
committed
[Cleanup] Refactor duplicateShardingsAtIndices to return StatusOr

Varcho
authored and
tensorflower-gardener
committed
[XLA] When canonicalizing ids also update corresponding id in HloComputation's schedule

zvikinoza
authored and
tensorflower-gardener
committed
It is impossible for the caller to be async-update or async-done
tensorflower-gardener
committed
Avoid using `std::bind`

junwhanahn
authored and
tensorflower-gardener
committed

Commits on Apr 28, 2026

[HLO Diff] Improve HLO diff determinism and constant matching.
tensorflower-gardener
committed
[XLA:MSA] Allocate alt-memory colored buffers first.

subhankarshah
authored and
tensorflower-gardener
committed
Enable checking of CompileOptions in the validation of each XLA phase's inputs.
tensorflower-gardener
committed
[ReplicaGroupV3][JAX+stablehlo][5/n] Shardy updates for RGV3. This includes

Varcho
authored and
tensorflower-gardener
committed
[XLA] When canonicalizing ids also update corresponding id in HloComputation's schedule
tensorflower-gardener
committed
Add a customizable point in MSA to allow passing a function to determine which instructions are eligible for window prefetch.
tensorflower-gardener
committed
Automated Code Change
tensorflower-gardener
committed
Automated Code Change
tensorflower-gardener
committed
Mark xla::CompilationEnvironments::HasEnv as const.
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
MlirKernelEmitter should only use single MLIRContext

ermilovmaxim
authored and
tensorflower-gardener
committed
Automated Code Change
tensorflower-gardener
committed
Automated Code Change
tensorflower-gardener
committed
PR #41603: [xla:se] Prepare allocator adaptors for sharing allocators

ezhulenev
authored and
tensorflower-gardener
committed
[XLA:GPU] Remove onehot rewriter pass

thcmbs
authored and
tensorflower-gardener
committed
PR #41407: [ROCm] Use AMD-specific allocate-shared-memory pass in Triton pipeline

nurmukhametov
authored and
tensorflower-gardener
committed
Update YNNPACK in XLA

dsharletg
authored and
tensorflower-gardener
committed
Add SwapSharding to XlaBuilder to remove redundant copies

liepieshov
authored and
tensorflower-gardener
committed
Allow to extract ragged-all-to-all

PatriosTheGreat
authored and
tensorflower-gardener
committed
Simplify CreateDotFusion API

vwbaker
authored and
tensorflower-gardener
committed
PR #41602: [xla:se] Delete unused collective memory allocators

ezhulenev
authored and
tensorflower-gardener
committed
Convert shardings to named shardings in dot handler tests.

Kingston Mandisodza
authored and
tensorflower-gardener
committed
Push shardy flattener down past ImportSdyCustomCallsPass pass.

ekayaaslan
authored and
tensorflower-gardener
committed
Get rid of const_cast in gemm_fusion.cc

vwbaker
authored and
tensorflower-gardener
committed
Handle partial replication in `PatternMatchPartiallyReplicateDim` for `NamedSharding`.

Kingston Mandisodza
authored and
tensorflower-gardener
committed
Push shardy flattener down past OpenWhileFreeVarsSharding pass.

ekayaaslan
authored and
tensorflower-gardener
committed
Discard manual computation calls explicitly during shardy unflattener.

ekayaaslan
authored and
tensorflower-gardener
committed
This change avoids a crash due to missing encoding on convert layout ops.

Moerafaat
authored and
tensorflower-gardener
committed
Push shardy flattener down past LiftInlinedMeshesPass pass.

ekayaaslan
authored and
tensorflower-gardener
committed
Update tests to avoid specifying an iota device list.

Kingston Mandisodza
authored and
tensorflower-gardener
committed
Cleanup setting shardings when they are already set previously

KanishAnand
authored and
tensorflower-gardener
committed
PR #41659: [xla] Port XLA Backend to use BFC allocator

ezhulenev
authored and
tensorflower-gardener
committed
[XLA:MLIR] Avoid mapping unused `scan` results in HLO translation.

chsigg
authored and
tensorflower-gardener
committed
Push shardy flattener down past DedupMeshes pass.

ekayaaslan
authored and
tensorflower-gardener
committed
Push shardy unflattener up past ShardMapExport pass.

ekayaaslan
authored and
tensorflower-gardener
committed
Switch to latest version of rules_ml_toolchains
alekstheod
committed
Avoid a crash when dynamic-reshape is missing an operand.
tensorflower-gardener
committed
Update partial replication check in SPMD complex reshard handling.

Kingston Mandisodza
authored and
tensorflower-gardener
committed
Push shardy flattener down past CanonicalizerPass pass.

ekayaaslan
authored and
tensorflower-gardener
committed
Push shardy unflattener up past ExportOps pass.

ekayaaslan
authored and
tensorflower-gardener
committed
Add iota instruction support to YnnEmitter

dsharletg
authored and
tensorflower-gardener
committed
Allow configuration of the `num_warmup_batch_threads` BatchFunction attribute.
tensorflower-gardener
committed
Rollback of PR #41659

belitskiy
authored and
tensorflower-gardener
committed
PR #36441: Support all-reduce hoisting with scalar multiplication pattern

sfvaroglu
authored and
tensorflower-gardener
committed
Integrate LLVM at llvm/llvm-project@7189c4bb83d9
tensorflower-gardener
committed
Correct a unit test on stablehlo import pipeline to make it actually non flat.

ekayaaslan
authored and
tensorflower-gardener
committed
Fix build failures for multiple fuzz targets in core/kernels/fuzzing

patnotz
authored and
tensorflower-gardener
committed
Push shardy unflattener up past ManualReductionCollectives pass.

ekayaaslan
authored and
tensorflower-gardener
committed
[XLA:GPU] Emit dot in two stages

vickyliu-go4it
authored and
tensorflower-gardener
committed
[XLA:TPU] BitcastConvert updated for packed `pred`s

iumitakgun
authored and
tensorflower-gardener
committed
Fix setup of cross-compile toolchains and update bazelrc.

meteorcloudy
authored and
tensorflower-gardener
committed
Added subprocess host traces to trace viewer.

bmass02
authored and
tensorflower-gardener
committed
[XLA] When canonicalizing ids also update corresponding id in HloComputation's schedule

zvikinoza
authored and
tensorflower-gardener
committed
Optimize `GlobalDecreasingSizeBestFitHeap::MakeFreeChunks` by computing `free_chunks_list_` from sorted `used_chunks_` in one pass instead of two passes.

nafi3000
authored and
tensorflower-gardener
committed
Pass tags for ThenExecuteCallback to debug slowness.

pschuh
authored and
tensorflower-gardener
committed
Fix an out-of-bounds read in the distributed runtime's RunCallable method by adding validation to ensure the number of feeds in the request matches or exceeds the number of feeds defined in the cac…

patnotz
authored and
tensorflower-gardener
committed
Update docs to use `AddDelegate` rather than `ModifyGraphWithDelegate`. `AddDelegate` is preferred, especially when using LiteRT in Play services,

fergushenderson
authored and
tensorflower-gardener
committed

Commits on Apr 29, 2026