xla

xla copied to clipboard

[fusion] Add pattern constraint for dynamic slice fusion

9

This patch constraints the dynamic-slice-fusion to only be triggered when the offset is either a constant or loop iteration offset that can guarentee that no D2H copy is required.

shraiysh

[ROCM] Enable basic functionality to run MLIR fussion tests

1

Specific tests will be fixed as we go

draganmladjenovic

mayuyuace

[Draft][XLA:GPU] Add sycl runtime

2

This PR aims to add sycl runtime support, we can run basic JAX GPU UTs with it. It includes: 1. sycl runtime crosstool build 2. sycl stream executor 3. spirv-llvm-translator...

Solaryee

[platform set error][GPU-cuda]NOT_FOUND: could not find registered compiler for platform CUDA -- was support for that platform linked in?

8

./configure.py --backend=CUDA bazel build --test_output=all --spawn_strategy=sandboxed //xla/... when set platform cuda，TF_ASSERT_OK_AND_ASSIGN(se::Platform * platform, PlatformUtil::GetPlatform("cuda")); get an error：NOT_FOUND: could not find registered compiler for platform CUDA -- was support for that...

FatJhon

In the decoding stage of some MOE model inferences, XLA squeezes dimensions of size 1 when sequence length is 1. For example, it transforms a shape of [1, 4096] into...

wenscarl

Automated Code Change

copybara-service[bot]

Automated Code Change

copybara-service[bot]

Implement MHLO->HLO conversion for entry parameter layout tiles. Does not support the following cases: * Parameters with nested layouts and tiles (e.g. tuple of tuples). * Multi result layouts and...

copybara-service[bot]

Integrate LLVM at llvm/llvm-project@070ce816dadb Updates LLVM usage to match [070ce816dadb](https://github.com/llvm/llvm-project/commit/070ce816dadb)

copybara-service[bot]