xla

xla copied to clipboard

Introduce hermetic CUDA in Google ML projects. 1) Hermetic CUDA rules allow building wheels with GPU support on a machine without GPUs, as well as running Bazel GPU tests on...

copybara-service[bot]

Always use std::array. `Eigen::array` is being removed upstream in favor of `std::array`.

copybara-service[bot]

Minor change to add logic for finding all lines with same id.

copybara-service[bot]

[XLA:GPU] Run autotuner with cublas config only if `--xla_gpu_cublas_fallback=true`. Currently we always compile in cublas by default and only later drop it from possible list of configs if the flag...

copybara-service[bot]

[XLA:GPU] Optimize command buffer update by only operating on commands that have pointer change

3

shawnwang18

Head CI test

1

shawnwang18

[GPU] Enable cuDNN GEMM fusion level 3 by default.

9

Affects only Hopper+ and cuDNN 9+: https://github.com/openxla/xla/blob/fb41b76a8b08216b80abb49ceb5c07373d9c45c5/xla/service/gpu/gemm_fusion_autotuner.cc#L556. Description of fusion level 1: https://github.com/openxla/xla/blob/fb41b76a8b08216b80abb49ceb5c07373d9c45c5/xla/xla.proto#L742.

sergachev

[XLA:GPU] Pass the CUDA / ROCm toolkit version explicitly for autotuning and GEMM rewriting. This allows to remove more `#if GOOGLE_CUDA` preprocessor directives from HLO passes.

copybara-service[bot]

There exist bug in ComputeThreadIdToOutputIndexing func. Currently, this func can not calculate indexing map correctly for sided output. Fix it and add corresponding test.

lingzhi98

Add build macro to generate hlo compilations test build rules.

copybara-service[bot]