21 Pull requests merged by 11 people

• Introduce HLO graph bindings

  #8564 merged Jan 14, 2025

• pin update

  #8559 merged Jan 13, 2025

• Set gradient_as_bucket_view=True in test_train_mp_imagenet

  #8558 merged Jan 13, 2025

• fix batch_norm amp autocast

  #8556 merged Jan 13, 2025

• [scan] Test we don't recompile under debugging env flags

  #8555 merged Jan 13, 2025

• Metadata agnostic user computation hash

  #8557 merged Jan 11, 2025

• [Computation Hash] Introduce deterministic hash for user computations

  #8554 merged Jan 11, 2025

• Introduce HLO graph bindings

  #8551 merged Jan 11, 2025

• Always run TPU tests on new PRs

  #8552 merged Jan 10, 2025

• Metadata agnostic user computation hash

  #8550 merged Jan 10, 2025

• Lower cummax op

  #8491 merged Jan 10, 2025

• [torch_xla2] Fix reenabled op info tests

  #8548 merged Jan 10, 2025

• [Computation Hash] Introduce deterministic hash for user computations

  #8539 merged Jan 10, 2025

• Revert "fix batch_norm amp autocast"

  #8547 merged Jan 9, 2025

• Fix typos in pytorch on xla docs

  #8543 merged Jan 9, 2025

• fix batch_norm amp autocast

  #8498 merged Jan 8, 2025

• doc change 1

  #8544 merged Jan 8, 2025

• [torch_xla2] Update test_ops.py

  #8403 merged Jan 7, 2025

• fix ab in flash attention

  #8540 merged Jan 7, 2025

• Fully revamp aten embedding bag for JAX backend

  #8535 merged Jan 7, 2025

• pin update

  #8536 merged Jan 7, 2025

4 Pull requests opened by 3 people

• Improve and refine MLP tests for extensibility and A/B testing

  #8561 opened Jan 13, 2025

• [scan] Reduce memory usage

  #8562 opened Jan 14, 2025

• Cherrypick #8521 into r2.6

  #8563 opened Jan 14, 2025

• Lower cummin op

  #8565 opened Jan 14, 2025

6 Issues closed by 5 people

• [Computation Hash] User Computation hash is agnostic to debug metadata

  #8538 closed Jan 10, 2025

• Lower cummax

  #8230 closed Jan 10, 2025

• [Computation Hash] User Computation hash disregards protobuf requirements

  #8537 closed Jan 10, 2025

• Export training model to StableHlo

  #8366 closed Jan 9, 2025

• AMP BF16 issue with batch norm layer

  #8496 closed Jan 8, 2025

• Input tensor is not an XLA tensor on AWS Trainium instance

  #8510 closed Jan 8, 2025

5 Issues opened by 5 people

• [v2.5] Import torch_xla crashed with "TypeError: must be called with a dataclass type or instance" with torch-xla 2.5 and torch 2.5.1+cpu (CPU flavor)

  #8560 opened Jan 13, 2025

• Multigpu training hangs using single and multiple nodes

  #8549 opened Jan 10, 2025

• [Torch_xla2] `torch_xla2.default_env()` guard doesn't enforce XLATensor2

  #8546 opened Jan 9, 2025

• [Q][GPU][BF16] torch.mul is lowered to HLO as an f32 multiply

  #8545 opened Jan 8, 2025

• [RFC] Add HuggingFace tests with pinned dependencies to CI

  #8542 opened Jan 7, 2025

9 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.

• Lowering Aten op to composite op instead of small ops

  #8502 commented on Jan 10, 2025 • 8 new comments

• Implement torch.linspace and torch.logspace

  #8533 commented on Jan 8, 2025 • 6 new comments

• 2 questions for the composite op feature

  #8486 commented on Jan 9, 2025 • 0 new comments

• Slow XLA training performance.

  #8541 commented on Jan 9, 2025 • 0 new comments

• Build slowdown issue

  #3569 commented on Jan 9, 2025 • 0 new comments

• PyTorch ResNet18 GPU Training Failed on Colab

  #3403 commented on Jan 9, 2025 • 0 new comments

• PyTorch toy model fails to execute on DTensor constructed from "xla" type device mesh

  #8534 commented on Jan 10, 2025 • 0 new comments

• PyTorch DTensor device mesh interface with device type "xla" fails at `get_rank()`

  #8528 commented on Jan 12, 2025 • 0 new comments

• 2.6 backport PR request list

  #8455 commented on Jan 14, 2025 • 0 new comments