Use convert_to_tensor inside init to avoid eager-only numpy() under XLA#2

Draft

CodersAcademy006 wants to merge 1822 commits intomasterfrom

clean/convert-constants-init

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

Commits on Nov 20, 2025

fix typo in tuple_simplifier.
tensorflower-gardener
committed
PR #34103: [GPU] Fix layout assignment of bitcast-converts.

sergachev
authored and
tensorflower-gardener
committed
[XLA:CPU][XTile] Expand single element vector ops before lowering to intrinsics.

WillFroom
authored and
tensorflower-gardener
committed
Update layout only when ```allow_spmd_sharding_propagation_to_output/parameter``` is true.
tensorflower-gardener
committed
Automated Code Change
tensorflower-gardener
committed
PR #32053: [ROCM] Added command buffers support for convolutions

pemeliya
authored and
tensorflower-gardener
committed
[XLA:SPMD] Fix on entry input/output layout changing check.

Tongfei-Guo
authored and
tensorflower-gardener
committed
[XLA] Another small refactor in call splitter.

mkuperst
authored and
tensorflower-gardener
committed
[XLA:GPU] Refactor: Update composite_rewriter to parse dot dimension numbers from a structured MLIR attribute.

loislo
authored and
tensorflower-gardener
committed
[CI] Stop optional services in Windows CI to reduce file permission error flakes.

belitskiy
authored and
tensorflower-gardener
committed
Populate input and output names of placeholder Signature

terryheo
authored and
tensorflower-gardener
committed
Minor check for code integrity.

felixwqp
authored and
tensorflower-gardener
committed
[xla:cpu] Remove unused runtime_lightweight_check

ezhulenev
authored and
tensorflower-gardener
committed
Add CollectivePermute support to CollectiveInterpolator. The interpolation model for collective-permute uses transfer size and communication pattern type as input key.

felixwqp
authored and
tensorflower-gardener
committed
Port ConditionalThunk/WhileThunk to provide shape for BufferUse

ermilovmaxim
authored and
tensorflower-gardener
committed
Integrate LLVM at llvm/llvm-project@355e0f94af5a
tensorflower-gardener
committed
Remove unused deprecated absl::testing usages in XLA GPU backend / service

apivovarov
authored and
tensorflower-gardener
committed
[XLA] Only split the body of a call once.

mkuperst
authored and
tensorflower-gardener
committed
Support sink VarHandleOp in tf.While.

SiqiaoWu1993
authored and
tensorflower-gardener
committed

Commits on Nov 21, 2025

Commits on Nov 22, 2025

Don't translate StableHLO to MHLO when serving saved models.

mrguenther
authored and
tensorflower-gardener
committed
Internal change only

SiqiaoWu1993
authored and
tensorflower-gardener
committed
Allow printing more than 1000 literal values in literal_comparison

Gregory Pataky
authored and
tensorflower-gardener
committed
Add a PJRT C API extension for TPU topology.

hhb
authored and
tensorflower-gardener
committed
Refactor CreateOutputLeafTpuBuffer to call DefineBuffer instead. Because of

pschuh
authored and
tensorflower-gardener
committed
Update workspace0.bzl
johnnynunez
authored
[NanoRt] Change the (default) memory kind to "device"

hyeontaek
authored and
tensorflower-gardener
committed
Update GraphDef version to 2419.
tensorflower-gardener
committed
compat: Update forward compatibility horizon to 2025-11-22
tensorflower-gardener
committed
[XLA:GPU] Remove SymbolicExprContext.

pifon2a
authored and
tensorflower-gardener
committed
Remove unused deprecated absl::testing usages in pjrt/ifrt

apivovarov
authored and
tensorflower-gardener
committed
[PJRT C API] Fix ASAN bug in cross-host transfer extension tests.

emilyfertig
authored and
tensorflower-gardener
committed
Integrate LLVM at llvm/llvm-project@d65be16ab6ad
tensorflower-gardener
committed

Commits on Nov 23, 2025

Commits on Nov 24, 2025

Integrate LLVM at llvm/llvm-project@f2cb5d7a05ba
tensorflower-gardener
committed
Update GraphDef version to 2421.
tensorflower-gardener
committed
compat: Update forward compatibility horizon to 2025-11-24
tensorflower-gardener
committed
Add hostname override to customize remote profiling filenames

subhamsoni-google
authored and
tensorflower-gardener
committed
[XLA:GPU] Roll forward the deletion of the legacy Triton GEMM emitter.

bchetioui
authored and
tensorflower-gardener
committed
[XLA:CPU/GPU][XTile] Make clamp return max in the case the limits are inverted.

WillFroom
authored and
tensorflower-gardener
committed
[XLA:GPU] Make setting `--xla_gpu_unsupported_generic_triton_emitter_features` into a no-op.

bchetioui
authored and
tensorflower-gardener
committed
PR #34296: [ROCm] Fix buildbreak due to missing __builtin_amdgcn_lerp

alekstheod
authored and
tensorflower-gardener
committed
[XTile] Add compatible_with_portable rules to enable CPU linking.

WillFroom
authored and
tensorflower-gardener
committed
[XTile] Enable passing fusions without gpu backend config.

WillFroom
authored and
tensorflower-gardener
committed
PR #34228: Refactor the heuristic for max unroll factor.

dimvar
authored and
tensorflower-gardener
committed
[XLA:CPU] Enable creating XLA:CPU client from a given topology

basioli-k
authored and
tensorflower-gardener
committed
Fix a layout bug in HloEvaluator.

akuegel
authored and
tensorflower-gardener
committed
[XLA][codegen] Migrate FpToFpOp to arith trunc/ext ops.

basioli-k
authored and
tensorflower-gardener
committed
Fix OSS build failure.

allanrenucci
authored and
tensorflower-gardener
committed
[XLA:GPU] Pass the llvm module explicitly to the BuildKernelPrototype.

pifon2a
authored and
tensorflower-gardener
committed
PR #34309: [XLA:GPU] Bump cuDNN version for block scaled dot support to 9.10

sergey-kozub
authored and
tensorflower-gardener
committed
[XLA:GPU] Move CollectiveOpsTestE2E base into a separate build target.

olegshyshkov
authored and
tensorflower-gardener
committed
Merge pull request #104892 from MedoX71T:master
tensorflower-gardener
committed
PR #34163: [GPU] Make cuDNN GEMM backend aware of dot algorithms.

sergachev
authored and
tensorflower-gardener
committed
Document the effort level options avaiable in XLA.
tensorflower-gardener
committed
Merge pull request #103912 from ILCSFNO:patch-4
tensorflower-gardener
committed
Merge pull request #103916 from ILCSFNO:patch-5
tensorflower-gardener
committed
Merge pull request #104548 from AshiteshSingh:patch-1
tensorflower-gardener
committed
[XLA:GPU] Emit SliceToDynamic and PadToStatic in a separate llvm module.

pifon2a
authored and
tensorflower-gardener
committed
PR #34316: [XLA:GPU] Add debug code to profile command buffer's memory types

shawnwang18
authored and
tensorflower-gardener
committed
PR #33909: [ROCm] Make multi gpu tests exclusive if executed locally

alekstheod
authored and
tensorflower-gardener
committed
PR #34049: [ROCm] Add support for rocm tar/wheels in hermetic builds

alekstheod
authored and
tensorflower-gardener
committed
[XLA:LHS] Fix minor issues:

seherellis
authored and
tensorflower-gardener
committed
[XLA:GPU] Add IrEmitterUnnested::EmitHloEntryComputation method.

pifon2a
authored and
tensorflower-gardener
committed
PR #34333: Add DSV3-1N4G HLO

mingxu1067
authored and
tensorflower-gardener
committed
PR #34335: Add HLO benchmark for llama3-8b with activation offloading

sfvaroglu
authored and
tensorflower-gardener
committed
[XLA] Increase test size to avoid timeouts.

dimitar-asenov
authored and
tensorflower-gardener
committed
Add missing symbols to CUPTI stub (follow-up for https://github.com/openxla/xla/pull/34298)

ybaturina
authored and
tensorflower-gardener
committed
Internal change only

SiqiaoWu1993
authored and
tensorflower-gardener
committed
[XLA:CPU][XTile] Add first experimental integration of tiled emitter.

WillFroom
authored and
tensorflower-gardener
committed
Reverts 4bf20e19f440a47a276cf1988b48683778f673a6

dimitar-asenov
authored and
tensorflower-gardener
committed
unify unused triton argument removal

ermilovmaxim
authored and
tensorflower-gardener
committed
Allow HloDCE to remove dead parameters from the entry computation.

bhatuzdaname
authored and
tensorflower-gardener
committed
[XLA:GPU] Move linking logic for emitters to ir_emitter_unnested.

pifon2a
authored and
tensorflower-gardener
committed

Commits on Nov 25, 2025

Reverts 335be54cf16896a093589c755dd9ee7d012216b6
tensorflower-gardener
committed
Add HBM utilization percent metric to XPlane schema.
tensorflower-gardener
committed
PR #34227: [GPU] Remove no longer necessary workaround for cuDNN convolutions.

sergachev
authored and
tensorflower-gardener
committed
Use unbounded parallelism for IFRT IR program compilation

junwhanahn
authored and
tensorflower-gardener
committed
[xla:gpu] Extract collective clique requests and acquire into separate libraries

ezhulenev
authored and
tensorflower-gardener
committed
Integrate LLVM at llvm/llvm-project@dea330b38d9c
tensorflower-gardener
committed
Remove unused deprecated tsl::testing usages

apivovarov
authored and
tensorflower-gardener
committed
[xla:ffi] Group internal CPU and GPU APIs for readability

ezhulenev
authored and
tensorflower-gardener
committed
[xla:gpu] Remove unused op kind from CollectiveConfig

ezhulenev
authored and
tensorflower-gardener
committed
[XLA:GPU] Move constant emission logic from IrEmitterContext to IrEmitterUnnested.

pifon2a
authored and
tensorflower-gardener
committed
[XLA:GPU] add an overview for thunks

metaflow
authored and
tensorflower-gardener
committed
Automated Code Change
tensorflower-gardener
committed
Automated Code Change
tensorflower-gardener
committed
Automated Code Change
tensorflower-gardener
committed
[xla:gpu] Remove unused op_id from CollectiveConfig

ezhulenev
authored and
tensorflower-gardener
committed
[XLA:GPU][XTile] Fold / squeeze xtile.mask when lowering to triton.

WillFroom
authored and
tensorflower-gardener
committed
Automated Code Change
tensorflower-gardener
committed
Automated Code Change
tensorflower-gardener
committed
[XLA:GPU] use reserved "field name"

metaflow
authored and
tensorflower-gardener
committed
compat: Update forward compatibility horizon to 2025-11-25
tensorflower-gardener
committed
Update GraphDef version to 2422.
tensorflower-gardener
committed
[xla:gpu] Remove redundant operand_count from CollectiveConfig

ezhulenev
authored and
tensorflower-gardener
committed
[XLA:CPU][XTile] Use shlo optimization passes.

WillFroom
authored and
tensorflower-gardener
committed
PR #33897: [ROCm] Make rbe amdgpu pools configurable

alekstheod
authored and
tensorflower-gardener
committed
Re-generate XLA's warnings.bazelrc.

thomasjoerg
authored and
tensorflower-gardener
committed
PR #34303: Bump github/codeql-action from 4.31.2 to 4.31.5

dependabot[bot]
authored and
tensorflower-gardener
committed
Refactor: Use OpType::create instead of rewriter.create<OpType>

chsigg
authored and
tensorflower-gardener
committed
Reverts fa3c810dc0c94f61760d4480955036659ad9bf51

thomasjoerg
authored and
tensorflower-gardener
committed
PR #34156: [ROCm] Add missing dependencies to header file

mfrancepillois
authored and
tensorflower-gardener
committed
[XLA:GPU] Pass HloInstruction to MustWrapInstruction.

olegshyshkov
authored and
tensorflower-gardener
committed
Reverts changelist 723154962

thomasjoerg
authored and
tensorflower-gardener
committed
Remove kTritonScaledDotFusionKind.

chsigg
authored and
tensorflower-gardener
committed
Remove legacy Triton pointer ops from int4 passes.

chsigg
authored and
tensorflower-gardener
committed
PR #34307: Bump ml-dtypes from 0.5.3 to 0.5.4

dependabot[bot]
authored and
tensorflower-gardener
committed
[XLA:GPU/TMA] Add heuristic for Triton TMA autotuning. This prunes down the configuration search space while maintaining the impactful configurations for TMA.

Moerafaat
authored and
tensorflower-gardener
committed
[XLA] Introduce --xla_dump_emitter_re flag to control which emitter dumps to dump

mooskagh
authored and
tensorflower-gardener
committed
Cleanup: Remove special handling for `kTritonGemmFusionKind`.

chsigg
authored and
tensorflower-gardener
committed
[XLA:GPU] Increase test timeout for `combined_ops_test_a`.

thomasjoerg
authored and
tensorflower-gardener
committed
[XLA:GPU] Introduce a flag --xla_gpu_gemm_autotuner_override_file

mooskagh
authored and
tensorflower-gardener
committed
Move stablehlo input processing logic to hlo_module_loader.cc.

akuegel
authored and
tensorflower-gardener
committed
[XLA:GPU] Return MLIR pipeline status error instead of silently ignoring.

olegshyshkov
authored and
tensorflower-gardener
committed
[XLA] Fix UndefinedBehaviorSanitizer `null-pointer-use` issue in `xla_host_send_recv_device_context_test`.

thomasjoerg
authored and
tensorflower-gardener
committed
[XLA:GPU] Fix usages of deprecated tsl::errors::NotFound in xla

loislo
authored and
tensorflower-gardener
committed
Introduce new AOT compilation support in `GpuCompiler`

EusebioDM
authored and
tensorflower-gardener
committed
[XLA:GPU] Bring hlo_op_profiler_run in sync with hlo_op_profiler_test.

akuegel
authored and
tensorflower-gardener
committed
Move `GpuExecutable` dumping logic from the compiler to the executable

EusebioDM
authored and
tensorflower-gardener
committed
[XLA:GPU/TMA] Enable TMA by default. This brings more performance to workloads on Hopper+ devices.

Moerafaat
authored and
tensorflower-gardener
committed
Integrate LLVM at llvm/llvm-project@26362c68579d
tensorflower-gardener
committed
avoid arithmetics with nullptr(UB)

ermilovmaxim
authored and
tensorflower-gardener
committed
switch from deprecated TF_CHECK_OK

ermilovmaxim
authored and
tensorflower-gardener
committed
Remove unused deprecated tsl::testing usages in tsl

apivovarov
authored and
tensorflower-gardener
committed
[XLA:GPU] Add missing dependency on absl/strings:cord.

loislo
authored and
tensorflower-gardener
committed
PR #34306: Bump numpy from 1.24.3 to 2.3.5

dependabot[bot]
authored and
tensorflower-gardener
committed
PR #34304: Bump actions/checkout from 5.0.0 to 6.0.0

dependabot[bot]
authored and
tensorflower-gardener
committed
[xla:ffi] Add correct error handling to internal FFI APIs

ezhulenev
authored and
tensorflower-gardener
committed
Convert a slice taking a single element from the minor dim of a reshape into a strided slice to allow more performant tiling on some platforms.

jbspooner
authored and
tensorflower-gardener
committed
[XLA] UnflattenCallGraph: Filter computations before hashing

zvikinoza
authored and
tensorflower-gardener
committed

Commits on Nov 26, 2025

Add FP32-->FP16 folding support to tfl.cast.

arfaian
authored and
tensorflower-gardener
committed
[xla:ffi] Split internal XLA FFI API implementation into separate target

ezhulenev
authored and
tensorflower-gardener
committed
[XLA][codegen] Migrate triton specific operations from collective emitter to xtile

basioli-k
authored and
tensorflower-gardener
committed
Implement `CreateErrorBuffer` in pjrt c api

hhb
authored and
tensorflower-gardener
committed
Add `StatType` enumerations for SparseCore

charlesalaras
authored and
tensorflower-gardener
committed
switch from deprecated TF_CHECK_OK

ermilovmaxim
authored and
tensorflower-gardener
committed
Merge pull request #104910 from johnnynunez:master
tensorflower-gardener
committed
[XLA:GPU] Rename IrEmitterUnnested to ThunkEmitter like in XLA:CPU.

pifon2a
authored and
tensorflower-gardener
committed
Automated Code Change
tensorflower-gardener
committed
Merge pull request #104948 from 372046933:fix_macos_26_compile
tensorflower-gardener
committed
[XLA:GPU] add xla_enable_scoped_logging_timers debug option

metaflow
authored and
tensorflower-gardener
committed
[XLA:GPU] Disable `sol_latency_estimator_test` for fastbuild, since it is broken.

thomasjoerg
authored and
tensorflower-gardener
committed
Update GraphDef version to 2423.
tensorflower-gardener
committed
compat: Update forward compatibility horizon to 2025-11-26
tensorflower-gardener
committed
[XLA:GPU]: Fix function arguments for metadata construction

sohaibiftikhar
authored and
tensorflower-gardener
committed
[PjRt] Avoid race conditions in LocalDeviceState destructor.

thomasjoerg
authored and
tensorflower-gardener
committed
[XLA:GPU] Fix usages of deprecated tsl::errors::InvalidArgument in xla

loislo
authored and
tensorflower-gardener
committed
[XLA:GPU] updates to thunks diagram

metaflow
authored and
tensorflower-gardener
committed
[XLA:GPU]: Emit Sort in a separate llvm module.

sohaibiftikhar
authored and
tensorflower-gardener
committed
fix heap overflow for dynamic dimension buffers

ermilovmaxim
authored and
tensorflower-gardener
committed
[XLA:GPU] Fix collective_ops_e2e_test timeout and enable it in OSS presubmits.
tensorflower-gardener
committed
PR #34372: [ROCm] Fix missing sys deps lib in hermetic builds

alekstheod
authored and
tensorflower-gardener
committed
[Autotuner] Use CustomKernel fission backend in legacy autotuner cache.
tensorflower-gardener
committed
[XLA:CPU/GPU][XTile] Split out lowering functionality from emitter_helpers.

WillFroom
authored and
tensorflower-gardener
committed
[XLA:GPU] Fix lowering of triton atomic passes.

sohaibiftikhar
authored and
tensorflower-gardener
committed
Replace the tsl_grpc_cc_dependencies macro by explicit dependencies

beckerhe
authored and
tensorflower-gardener
committed
Use GetInPlaceInputOutputPairs from AliasInfo instead of HloDataflowAnalysis.

akuegel
authored and
tensorflower-gardener
committed
PR #34362: [GPU] Link to optimization level doc in GPU flag guidance

terryysun
authored and
tensorflower-gardener
committed
PR #34112: [ROCm] Include multigpu tests

alekstheod
authored and
tensorflower-gardener
committed
[XLA:CPU/GPU][XTile] Split tiled emitting and lowering into two separate targets.

WillFroom
authored and
tensorflower-gardener
committed
[XLA:GPU] Add xla.get_dynamic_dim_size op and its lowering.

olegshyshkov
authored and
tensorflower-gardener
committed
Integrate LLVM at llvm/llvm-project@4f39a4ff0ada

boomanaiden154
authored and
tensorflower-gardener
committed
[xla:ffi] Add execution stage to all error messages when checking FFI signature

ezhulenev
authored and
tensorflower-gardener
committed
Fix protobuf dependencies in OSS

beckerhe
authored and
tensorflower-gardener
committed
[XLA:GPU] Fix usages of deprecated tsl::errors::AlreadyExists in xla

loislo
authored and
tensorflower-gardener
committed
[XLA:GPU/TMA] Disable TMA on B200 due to a timeout failure.

Moerafaat
authored and
tensorflower-gardener
committed
[XLA] Fix zstd.patch

WillFroom
authored and
tensorflower-gardener
committed
Reverts 6185a6b9ab9f84b95a9a94e55d6d16afa433b91a

beckerhe
authored and
tensorflower-gardener
committed
[xla:ffi] Relax the check for the number of attributes

ezhulenev
authored and
tensorflower-gardener
committed
[XLA:GPU] Add documentation to Priority Fusion pass.

derdrdirk
authored and
tensorflower-gardener
committed
hlo_runner_pjrt should keep hlo module config's seed() in optimized hlo module
tensorflower-gardener
committed
Return `std::nullopt` from `xla::ifrt::LoadedExecutable::devices()` for portable executables

junwhanahn
authored and
tensorflower-gardener
committed
[XLA:CPU] Use distinct element values in tiled kernel tests

basioli-k
authored and
tensorflower-gardener
committed
[XLA:GPU] Prevent all-reduce codegen when replica groups are empty

sohaibiftikhar
authored and
tensorflower-gardener
committed
[XLA:CPU][AOT] Make sure stablehlo test doesn't fail if executable loading fails due to different compilation machines

basioli-k
authored and
tensorflower-gardener
committed
[XLA] Update build_bazel_apple_support to 1.24.5

hawkinsp
authored and
tensorflower-gardener
committed
[XLA:GPU] Move ragged all to all e2e tests into a separate target.

olegshyshkov
authored and
tensorflower-gardener
committed
Refactor HloDCE to use a setter for removing dead entry parameters.

bhatuzdaname
authored and
tensorflower-gardener
committed
Refactor client_library_test_base.cc

apivovarov
authored and
tensorflower-gardener
committed
Implement PjRtCApiExecutable::GetOutputShapes.

hhb
authored and
tensorflower-gardener
committed
[XLA] Don't eagerly delete side effecting custom calls

vsytch
authored and
tensorflower-gardener
committed
switch from deprecated TF_CHECK_OK

ermilovmaxim
authored and
tensorflower-gardener
committed
Fix types and comments in replica groups V2.

ZixuanJiang
authored and
tensorflower-gardener
committed
[XLA:GPU/TMA] Pruning TMA configurations in the new autotuner infrastructure. This is done in a similar way to the old one.

Moerafaat
authored and
tensorflower-gardener
committed
Add fp16 data type to TFLite for use within the runtime.

arfaian
authored and
tensorflower-gardener
committed
[XLA:GPU] Remove unused num_devices_ member from CollectiveOpsWithFlagsBase.

olegshyshkov
authored and
tensorflower-gardener
committed
[xla:ffi] Move CPU FFI implementation to backends/cpu

ezhulenev
authored and
tensorflower-gardener
committed
[XLA] Add gutil dependency

Marcin Radomski
authored and
tensorflower-gardener
committed

Commits on Nov 27, 2025