[pull] master from tensorflow:master - #8859

Merged

pull[bot] merged 24 commits into

Cache-Cloud:masterfrom

tensorflow:master

Sep 30, 2026

Commits on Sep 29, 2026

[PJRT:GPU] Fail distributed initialization when hosts have different topology fingerprints.

junwhanahn
authored and
tensorflower-gardener
committed
3152ae3
Fix PjRt StreamExecutor execution to check BufferSequencingEvent error after waiting on stream.
tensorflower-gardener
committed
9eb09c7
Add function_body dependency to Grappler functions library to fix Windows build link error.

dmiltr3
authored and
tensorflower-gardener
committed
25a7578
[XLA] Fix TryRemoveTrivialCompare folding i > c to true when x == c
tensorflower-gardener
committed
2f442bb
[PJRT:GPU] Log GPU topology proto before checking topology fingerprint.

junwhanahn
authored and
tensorflower-gardener
committed
07de375
Fix nvcc compilation by using explicit traits in CUDA TopK

apivovarov
authored and
tensorflower-gardener
committed
b7c15f9
Change n2-128 runners to n4-80 and change the old L4 1GPU runner to the new L4 1GPU runner for benchmark.

quoctruong
authored and
tensorflower-gardener
committed
7393fa6
[Mosaic] Fix MemRefBitcastOp verification and tiling propagation

tlongeri
authored and
tensorflower-gardener
committed
9624b74
[XLA:GPU] Evaluate tiling candidates in parallel in FusionBlockLevelRewriter

Moerafaat
authored and
tensorflower-gardener
committed
099f9d4
[XLA:GPU] Support 1D broadcast parameters after pointwise chains in cuDNN conv fusion.

derdrdirk
authored and
tensorflower-gardener
committed
9eea599
Resolve ModuleNotFoundError in XLA CI and benchmark workflows when invoking the build script directly.

nitins17
authored and
tensorflower-gardener
committed
256b984
Enhance CUPTI activity overhead event naming for newer CUDA versions.
tensorflower-gardener
committed
c1c14ee
Factor vmem_for_operand into module level helper.

18praveenb
authored and
tensorflower-gardener
committed
65ae169
Bug fix: Isolate cuda-only dependencies to if_cuda in gpu_device tests.
tensorflower-gardener
committed
b7b6ee8
Fix more deprecated MLIR API call sites

mrguenther
authored and
tensorflower-gardener
committed
3b5cfa5
[xla:cpu] TargetMachineOptions: normalize target triple

cota
authored and
tensorflower-gardener
committed
f7d3e1d
Rollforward of: Make an overrideable Delinearize function instead of allowing DelinearizeAsync

pschuh
authored and
tensorflower-gardener
committed
2a50edb
mark tsl::Future as ABSL_MUST_USE_RESULT

ermilovmaxim
authored and
tensorflower-gardener
committed
543710f
[xla:cpu] move sanitizer passes to the very end of the pipeline

cota
authored and
tensorflower-gardener
committed
2d81e2c
[XLA:TPU] Pin convolution output layout inside copy-disabled while loops.
tensorflower-gardener
committed
ef5b0ad
[XLA:GPU] Evaluate tiling candidates in parallel in BlockLevelEmitterBackend

Moerafaat
authored and
tensorflower-gardener
committed
8d2c456

Commits on Sep 30, 2026

Add `cint2_fp32_int4_e8m0_drq` fusion pass and reference `FullyConnected` kernel.

majiddadashi
authored and
tensorflower-gardener
committed
6f21121
Prefer `std::numeric_limits<T>::lowest()` over `::min()`

pschuh
authored and
tensorflower-gardener
committed
17803f7
mark tsl::Future as ABSL_MUST_USE_RESULT
tensorflower-gardener
committed
39489a6

Navigation Menu

[pull] master from tensorflow:master - #8859

[pull] master from tensorflow:master#8859
pull[bot] merged 24 commits into
Cache-Cloud:masterfrom
tensorflow:master

Commits on Sep 29, 2026

[PJRT:GPU] Fail distributed initialization when hosts have different topology fingerprints.

Fix PjRt StreamExecutor execution to check BufferSequencingEvent error after waiting on stream.

Add function_body dependency to Grappler functions library to fix Windows build link error.

[XLA] Fix TryRemoveTrivialCompare folding i > c to true when x == c

[PJRT:GPU] Log GPU topology proto before checking topology fingerprint.

Fix nvcc compilation by using explicit traits in CUDA TopK

Change n2-128 runners to n4-80 and change the old L4 1GPU runner to the new L4 1GPU runner for benchmark.

[Mosaic] Fix MemRefBitcastOp verification and tiling propagation

[XLA:GPU] Evaluate tiling candidates in parallel in FusionBlockLevelRewriter

[XLA:GPU] Support 1D broadcast parameters after pointwise chains in cuDNN conv fusion.

Resolve ModuleNotFoundError in XLA CI and benchmark workflows when invoking the build script directly.

Enhance CUPTI activity overhead event naming for newer CUDA versions.

Factor vmem_for_operand into module level helper.

Bug fix: Isolate cuda-only dependencies to if_cuda in gpu_device tests.

Fix more deprecated MLIR API call sites

[xla:cpu] TargetMachineOptions: normalize target triple

Rollforward of: Make an overrideable Delinearize function instead of allowing DelinearizeAsync

mark tsl::Future as ABSL_MUST_USE_RESULT

[xla:cpu] move sanitizer passes to the very end of the pipeline

[XLA:TPU] Pin convolution output layout inside copy-disabled while loops.

[XLA:GPU] Evaluate tiling candidates in parallel in BlockLevelEmitterBackend

Commits on Sep 30, 2026

Add `cint2_fp32_int4_e8m0_drq` fusion pass and reference `FullyConnected` kernel.

Prefer `std::numeric_limits<T>::lowest()` over `::min()`

mark tsl::Future as ABSL_MUST_USE_RESULT