10 Pull requests merged by 6 people

• Add tuned parameters for Qwen/Qwen2.5-32B

  #8966 merged Apr 12, 2025

• Adapt Splash Attention from TorchPrime

  #8911 merged Apr 11, 2025

• Make scan-based GRU support batch_first parameter.

  #8964 merged Apr 11, 2025

• [WRONG PR] @assume_pure

  #8923 merged Apr 10, 2025

• [call_jax] support returning PyTree from the JAX function

  #8957 merged Apr 10, 2025

• Remove deprecated typing module in ansible flow

  #8955 merged Apr 10, 2025

• Add test to ensure scan-based and the standard GRU are interchangeable.

  #8949 merged Apr 9, 2025

• Update to pytorch 2.6

  #8944 merged Apr 9, 2025

• Pin update to 20250406

  #8945 merged Apr 8, 2025

• Use pages_per_seq * page_size instead of directly passing max_model_len

  #8950 merged Apr 8, 2025

6 Pull requests opened by 4 people

• Add runtime check when using non-kernel for ragged paged attn

  #8958 opened Apr 10, 2025

• assume_pure & call_jax composible.

  #8961 opened Apr 10, 2025

• @assume_pure

  #8962 opened Apr 11, 2025

• Add an option for JittableModule to dedup parameters.

  #8965 opened Apr 11, 2025

• Add a helper class to handle mesh and sharding

  #8967 opened Apr 12, 2025

• Disable one splash attention test

  #8970 opened Apr 14, 2025

6 Issues closed by 5 people

• [RFC] scan operator and scan_layers

  #8620 closed Apr 10, 2025

• [cleanup] Remove install_post_deps_pytorch_xla

  #8934 closed Apr 10, 2025

• Support Python 3.13 in nightly wheels

  #8959 closed Apr 10, 2025

• User built torch-xla wheel fails on import

  #8940 closed Apr 9, 2025

• torch.distributed.all_reduce not converted to stableHLO

  #8854 closed Apr 9, 2025

• GPU test failure: test_dynamo.py

  #8952 closed Apr 9, 2025

10 Issues opened by 5 people

• [call_jax] Bridge the torch_xla and JAX mesh

  #8972 opened Apr 14, 2025

• Splash attention test fail randomly only in github CI

  #8971 opened Apr 14, 2025

• Incorrect stableHLO output of all reduce

  #8969 opened Apr 14, 2025

• Alternative to torch.select_mask

  #8968 opened Apr 13, 2025

• `call_jax` doesn't take jax config into hashing.

  #8963 opened Apr 11, 2025

• TPU 6e-1 + Pytorch 2.6[TPU] does not work

  #8960 opened Apr 10, 2025

• GPU test failed again: AtenXlaTensorTest.TestDivInPlaceWithRoundingMode

  #8956 opened Apr 10, 2025

• MegaScale discovery is ran twice again

  #8954 opened Apr 9, 2025

• torch.linalg.lstsq issues on GPU/TPU

  #8953 opened Apr 8, 2025

• GPU master test failure: test_python_ops.py

  #8951 opened Apr 8, 2025

12 Unresolved conversations

Sometimes conversations happen on old items that aren’t yet closed. Here is a list of all the Issues and Pull Requests with unresolved conversations.

• Large number of graph break with flash_attention on dynamo openxla backend

  #8913 commented on Apr 7, 2025 • 0 new comments

• Persistent cache doesn't work for GPU/TPU

  #8930 commented on Apr 8, 2025 • 0 new comments

• Run backward on CPU before importing torch_xla cause future backward on XLA crash

  #4174 commented on Apr 9, 2025 • 0 new comments

• TPU memory use increased significantly in torch/xla - 2.6.0.dev20241107

  #8423 commented on Apr 9, 2025 • 0 new comments

• Torch-XLA not compatible with static python

  #8948 commented on Apr 10, 2025 • 0 new comments

• [Deprecation Tracking] API deprecation timeline summary

  #8915 commented on Apr 11, 2025 • 0 new comments

• [torchax] jit compile the model constructor

  #8635 commented on Apr 11, 2025 • 0 new comments

• [scan] Avoid re-tracing the combine function on every call

  #8632 commented on Apr 14, 2025 • 0 new comments

• Transition to Hermetic CUDA

  #8665 commented on Apr 8, 2025 • 0 new comments

• [Draft] Add Experimental limited sparse embedding bag

  #8905 commented on Apr 11, 2025 • 0 new comments

• test on debian12

  #8928 commented on Apr 10, 2025 • 0 new comments

• add CreateGlobalShardedData prototype

  #8932 commented on Apr 14, 2025 • 0 new comments