Bhupendra Dubey
ff7eb222c2
Refactor XLA Profiler State Check to Use Low-Overhead C API
...
This CL refactors the XLA profiler's state-checking mechanism to resolve GIL deadlocks and improve performance.
Previously, the C++ profiler context would import a Python module to update the profiler's state. This operation, performed while holding the GIL, could cause deadlocks if the import failed (e.g., in a JAX-only environment).
This change replaces the fragile cross-language import with a shared C++ std::atomic<bool>. Python code now queries this state via a new, low-overhead C function (is_traceme_enabled_raw) instead of ctypes.
This approach eliminates the deadlocks, decouples the C++ profiler from Python modules, and maintains high performance for the state check. The internal C++ API was also updated to use a safer reference instead of a raw pointer.
PiperOrigin-RevId: 847261952
2025-12-20 19:56:52 -08:00
A. Unique TensorFlower
580eeae4c3
Automated Code Change
...
PiperOrigin-RevId: 847190483
2025-12-20 16:29:24 -08:00
A. Unique TensorFlower
9f41228dad
Automated Code Change
...
PiperOrigin-RevId: 847190468
2025-12-20 16:21:13 -08:00
A. Unique TensorFlower
618abe0548
Automated Code Change
...
PiperOrigin-RevId: 847190441
2025-12-20 15:51:51 -08:00
A. Unique TensorFlower
9799cb0a78
Automated Code Change
...
PiperOrigin-RevId: 847190272
2025-12-20 15:42:25 -08:00
A. Unique TensorFlower
b1b2723815
Automated Code Change
...
PiperOrigin-RevId: 847190131
2025-12-20 15:30:21 -08:00
A. Unique TensorFlower
b1082f97e5
Automated Code Change
...
PiperOrigin-RevId: 847189651
2025-12-20 15:21:34 -08:00
A. Unique TensorFlower
8633fb9dcf
Automated Code Change
...
PiperOrigin-RevId: 847189531
2025-12-20 15:12:29 -08:00
A. Unique TensorFlower
9e0976de36
Automated Code Change
...
PiperOrigin-RevId: 847189521
2025-12-20 14:59:19 -08:00
A. Unique TensorFlower
818e61079e
Automated Code Change
...
PiperOrigin-RevId: 847189518
2025-12-20 14:50:46 -08:00
A. Unique TensorFlower
3c431ffe3d
Automated Code Change
...
PiperOrigin-RevId: 847189510
2025-12-20 14:37:43 -08:00
A. Unique TensorFlower
2d5603d02e
Automated Code Change
...
PiperOrigin-RevId: 847189107
2025-12-20 14:29:52 -08:00
A. Unique TensorFlower
c0a2b0e8b7
Automated Code Change
...
PiperOrigin-RevId: 847189104
2025-12-20 14:21:27 -08:00
A. Unique TensorFlower
bc145c0cd7
Automated Code Change
...
PiperOrigin-RevId: 847188284
2025-12-20 14:09:26 -08:00
A. Unique TensorFlower
cd1558548f
Automated Code Change
...
PiperOrigin-RevId: 847188276
2025-12-20 14:01:16 -08:00
A. Unique TensorFlower
6e91c3160b
Automated Code Change
...
PiperOrigin-RevId: 847187496
2025-12-20 13:49:36 -08:00
A. Unique TensorFlower
6f24565d89
Automated Code Change
...
PiperOrigin-RevId: 847186391
2025-12-20 13:34:41 -08:00
Zixuan Jiang
68a2b28900
Run third_party/openxla/shardy/google/integrate_latest.sh for the recent unreduced axes.
...
PiperOrigin-RevId: 847181792
2025-12-20 13:12:11 -08:00
A. Unique TensorFlower
44a702d0be
Reverts 6383e3632c
...
PiperOrigin-RevId: 847161450
2025-12-20 11:32:00 -08:00
A. Unique TensorFlower
4c4373fe64
Automated Code Change
...
PiperOrigin-RevId: 847088166
2025-12-20 05:17:35 -08:00
A. Unique TensorFlower
e5a57dc540
Automated Code Change
...
PiperOrigin-RevId: 847082744
2025-12-20 04:44:54 -08:00
A. Unique TensorFlower
cd19aba91f
Automated Code Change
...
PiperOrigin-RevId: 847070976
2025-12-20 03:52:07 -08:00
A. Unique TensorFlower
f654c43ac2
Automated Code Change
...
PiperOrigin-RevId: 847065080
2025-12-20 03:14:37 -08:00
A. Unique TensorFlower
39ada7f15c
compat: Update forward compatibility horizon to 2025-12-20
...
PiperOrigin-RevId: 847036681
2025-12-20 01:22:25 -08:00
A. Unique TensorFlower
9f75c9a8e3
Update GraphDef version to 2447.
...
PiperOrigin-RevId: 847036679
2025-12-20 01:13:46 -08:00
A. Unique TensorFlower
2c61085107
Automated Code Change
...
PiperOrigin-RevId: 846987657
2025-12-19 22:02:34 -08:00
A. Unique TensorFlower
269ed39897
Automated Code Change
...
PiperOrigin-RevId: 846974498
2025-12-19 21:04:45 -08:00
Yash Katariya
4baf6a3d21
Add replicated -> unreduced test coverage
...
PiperOrigin-RevId: 846930629
2025-12-19 18:16:27 -08:00
Jian Cai
8f784b11c3
[XLA][Numerics][HLO Value Tracking] Support HLO original value in CopyFusion pass
...
This updates the HLO orignal value of a fusion accordingly if its shape is updated in the pass.
PiperOrigin-RevId: 846927675
2025-12-19 18:03:56 -08:00
Parker Schuh
9d833374f9
Move mutability tracking from Tracked buffers to RawBuffers as this is only used for importing foreign memory and clutters the API.
...
PiperOrigin-RevId: 846916941
2025-12-19 17:29:42 -08:00
Zixuan Jiang
f4a923fa82
Add a walk to convert unreduced sdy.constant to replicated sdy.constant and sdy.replicated_to_unreduced.
...
PiperOrigin-RevId: 846916791
2025-12-19 17:16:52 -08:00
Quentin Khan
911ce60c29
Use the XNNPack packing fingerprints to invalidate the weight cache.
...
PiperOrigin-RevId: 846914182
2025-12-19 17:07:07 -08:00
A. Unique TensorFlower
49fda87ad7
Update usage of llvm::PointerType::getUnqual to use non-deprecated version
...
PiperOrigin-RevId: 846902320
2025-12-19 16:23:35 -08:00
Volodymyr Kysenko
cb2c66e715
[XLA:CPU] Enable window dilation support in XLA CPU YNN convolution.
...
PiperOrigin-RevId: 846895590
2025-12-19 16:03:19 -08:00
Niklas Vangerow
3feec3590c
Migrate reshape_motion_test to PjRt.
...
PiperOrigin-RevId: 846886339
2025-12-19 15:34:30 -08:00
A. Unique TensorFlower
af8c7d0e2e
Apply llvm-use-new-mlir-op-builder fixes
...
This migrates `builder.create<Op>()` => `Op::create()`
PiperOrigin-RevId: 846865415
2025-12-19 14:49:32 -08:00
A. Unique TensorFlower
84ad581652
Apply llvm-use-new-mlir-op-builder fixes
...
This migrates `builder.create<Op>()` => `Op::create()`
PiperOrigin-RevId: 846862419
2025-12-19 14:35:51 -08:00
Bill Varcho
b35e4ed192
[ReplicaGroupV3][Refactor][4/n] Update replica group classes to hide implementation details (via private/protected) to external callers.
...
PiperOrigin-RevId: 846860541
2025-12-19 14:25:25 -08:00
A. Unique TensorFlower
d6a407c9f5
Integrate LLVM at llvm/llvm-project@7d381f2a56
...
Updates LLVM usage to match
[7d381f2a5634](https://github.com/llvm/llvm-project/commit/7d381f2a5634 )
PiperOrigin-RevId: 846858892
2025-12-19 14:15:38 -08:00
A. Unique TensorFlower
5e685fb6e1
Apply llvm-use-new-mlir-op-builder fixes
...
This migrates `builder.create<Op>()` => `Op::create()`
PiperOrigin-RevId: 846854812
2025-12-19 14:02:40 -08:00
Adrian Kuegel
d11a803d83
[XLA:GPU] Allow to fuse iotas into sort.
...
Also adjust alias info logic to allow sharing the sort fusion operands with the
right sort fusion outputs.
PiperOrigin-RevId: 846841000
2025-12-19 13:30:49 -08:00
Vlad Sytchenko
e961b23697
[XLA] Continue trying to unroll pipelined loops after failure
...
PiperOrigin-RevId: 846839467
2025-12-19 13:17:36 -08:00
Fengwu Yao
453b5bbfde
Internal changes only.
...
PiperOrigin-RevId: 846835877
2025-12-19 13:03:34 -08:00
Bill Varcho
07acbd560e
[ReplicaGroupV3][Refactor][3/n] Use CollectiveDeviceListBase for polymorphic device list support.
...
PiperOrigin-RevId: 846825118
2025-12-19 12:30:11 -08:00
Parker Schuh
068c5bfe95
Simplify TrackedCpuDeviceBuffer.
...
Update CpuRawBuffer to always have the size available and remove AfterAll implementation in TrackedCpuDeviceBuffer.
PiperOrigin-RevId: 846812637
2025-12-19 11:52:26 -08:00
Jian Cai
aa95cad5da
[XLA][Numerics][HLO Value Tracking] Support HLO original values in ConditionalSimplifier pass
...
This updates the original value of a conditional op when unused tuple elements are removed.
PiperOrigin-RevId: 846802773
2025-12-19 11:41:19 -08:00
Kevin Gleason
c698eecb5a
[StableHLO] Update CHLO broadcast ops to use StableHLO bounded-dynamism-aware broadcasting for type inference.
...
PiperOrigin-RevId: 846795035
2025-12-19 11:25:42 -08:00
Yulia Baturina
a3fad5d03e
Enable using custom hermetic NCCL version.
...
The NCCL version can be chosen via `HERMETIC_NCCL_VERSION` env var.
See docs [here](https://github.com/google-ml-infra/rules_ml_toolchain/blob/main/gpu/README.md#environment-variables-controlling-the-hermetic-cudacudnnnvshmem-versions ).
PiperOrigin-RevId: 846792626
2025-12-19 11:17:25 -08:00
Maxim Ermilov
ea362136d9
Allow serialization for sync collectives
...
PiperOrigin-RevId: 846790089
2025-12-19 11:05:30 -08:00
Yulia Baturina
843206ca81
Enable using custom hermetic NCCL version.
...
The NCCL version can be chosen via `HERMETIC_NCCL_VERSION` env var.
See docs [here](https://github.com/google-ml-infra/rules_ml_toolchain/blob/main/gpu/README.md#environment-variables-controlling-the-hermetic-cudacudnnnvshmem-versions ).
PiperOrigin-RevId: 846788706
2025-12-19 10:47:57 -08:00