7 Pull requests merged by 6 people

• Update XLA pin to 11/22

  #8411 merged Nov 25, 2024

• scan and scan_layers

  #7901 merged Nov 23, 2024

• Add an example that train the torchtitan version of llama.

  #8400 merged Nov 23, 2024

• Force specialize_float = True for torch_xla

  #8404 merged Nov 21, 2024

• Only write to HBM at the last iteration.

  #8393 merged Nov 20, 2024

• Unify how cpp test is being run

  #8391 merged Nov 19, 2024

• Update ARC runner version to 2.321.0

  #8397 merged Nov 19, 2024

6 Pull requests opened by 5 people

• [torch_xla2] Create jtorchvision_roi_align.py

  #8394 opened Nov 19, 2024

• Migrate from capture_pre_autograd_graph to export_for_training

  #8398 opened Nov 19, 2024

• [torch_xla2] Update test_ops.py

  #8403 opened Nov 20, 2024

• Switch to libtpu package from the libtpu-wheels registry

  #8409 opened Nov 22, 2024

• [Documentation] Added a section pointing readers to the AI-Hypercomputer/tpu-recipies repo for reference model implementations.

  #8412 opened Nov 25, 2024

• Unit test pass

  #8416 opened Nov 26, 2024

11 Issues opened by 8 people

• [LoweringContext] Modularize for enhanced user control and optimization

  #8415 opened Nov 25, 2024

• [LoweringContext] Support explicit device data parameters for scalar inputs

  #8414 opened Nov 25, 2024

• Review documentation in the docs/source/learn directory

  #8413 opened Nov 25, 2024

• ``torch.split`` followed by ``torch.cat`` fails to restore the tensor (v2.1-2.5)

  #8410 opened Nov 23, 2024

• Dataloading hangs on Trillium when using recent whls

  #8407 opened Nov 21, 2024

• Getting "undefined symbol: _ZN5torch4lazy13MetricFnValueB5cxx11E" with torch-xla nightly wheel for 2.6

  #8406 opened Nov 21, 2024

• Einsum is not added to the supported list for autocast

  #8405 opened Nov 21, 2024

• Kaggle Notebook: model return loss None on TPU

  #8402 opened Nov 20, 2024

• Add multiple TPU generations to the TPU CI tests

  #8399 opened Nov 19, 2024

• DDP doesn't synchronize weights

  #8396 opened Nov 19, 2024

• torch.layer_norm causes RESOURCE_EXHAUSTED

  #8395 opened Nov 19, 2024

11 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.

• A process in the process pool was terminated abruptly while the future was running or pending.

  #8233 commented on Nov 20, 2024 • 0 new comments

• How are PJRT asynchronous executions throttled by torch_xla?

  #8380 commented on Nov 21, 2024 • 0 new comments

• [fp8] Support `fp8e4m3` in torch_xla

  #8005 commented on Nov 21, 2024 • 0 new comments

• [XLA:GPU compile Error] nvcc fatal : Unsupported gpu architecture 'compute_35'

  #7823 commented on Nov 24, 2024 • 0 new comments

• int8 StableHLO export

  #8373 commented on Nov 25, 2024 • 0 new comments

• TPU Trillium Base Docker Image cannot initialize

  #8371 commented on Nov 25, 2024 • 0 new comments

• How to write in-place custom ops compatible with torch.compile using pallas

  #8385 commented on Nov 25, 2024 • 0 new comments

• Need help validating TPU/XLA devices support for ComfyUI.

  #8388 commented on Nov 25, 2024 • 0 new comments

• Support of AOT compilation (refine #6992)

  #7581 commented on Nov 22, 2024 • 0 new comments

• Align rrelu_with_noise schema to reflect noise mutation

  #8309 commented on Nov 22, 2024 • 0 new comments

• custom_kernel: fix shape mismatch by sharding segment_ids in flash attn.

  #8333 commented on Nov 25, 2024 • 0 new comments