All pull requests

Pull requests

• Arm backend: Fix provenance for GroupNorm.

  ciflow/trunkCLA Signedmodule: armpartner: armrelease notes: arm
  #24240·wwwind opened Oct 28, 2026Collaborator·

  ·
  1

• Keep a folded int8 dtype through every retrace

  CLA Signedmeta-exportedrelease notes: none
  #24239·pssrawat opened Oct 28, 2026Contributor·

  ·
  2

• Call distutils log.warn, not log.warning, in setup.py

  ciflow/cudaCLA Signedrelease notes: build
  #24238·shoumikhin opened Oct 28, 2026Contributor·

  ·
  2

• Arm backend: Gate int16 rsqrt lowering on TOSA int16

  ciflow/trunkCLA Signedmodule: armpartner: arm
  #24237·bdemirb opened Oct 28, 2026Collaborator·

  ·
  3

• Support max pooling in the CPU XNNPACK provider

  CLA Signedmeta-exported
  #24236·JakeStevens opened Oct 28, 2026Contributor·

  ·
  3

• [ET][native] Verify bundled LLM samples through the runner

  CLA Signed
  #24234·pytorchbot opened Oct 28, 2026Collaborator·

  ·
  1

• [ET][native] Bundle LLM prefill and decode samples from export_hf_llm

  CLA Signed
  #24233·pytorchbot opened Oct 28, 2026Collaborator·

  ·
  1

• [ET][native] Read bundled samples in C++

  CLA Signed
  #24232·pytorchbot opened Oct 28, 2026Collaborator·

  ·
  1

• [ET][native] Make the LLM runner's next-token choice pluggable

  CLA Signed
  #24231·pytorchbot opened Oct 28, 2026Collaborator·

  ·
  1

• [ET][native] Bundle sample method calls in .ptn packages

  CLA Signed
  #24230·pytorchbot opened Oct 28, 2026Collaborator·

  ·
  2

• transforms: propagate permute/views through binary ops

  ciflow/trunkCLA Signedmodule: armpartner: armrelease notes: none
  #24229·AdrianLundell opened Oct 28, 2026Collaborator·

  ·
  2

• Fix two crashes caused by constant_prop_pass

  CLA Signed
  #24224·pytorchbot opened Oct 28, 2026Collaborator·

  ·
  1

• Fix crashes and lowering failures in the AOTI, OpenVINO and Qualcomm backends and in constant_prop_pass

  CLA Signedrelease notes: backends
  #24222·shoumikhin opened Oct 28, 2026Contributor·

  ·
  6

• Give backends the method's metadata at init (#24220)

  CLA Signedmeta-exported
  #24220·sivapatibandla opened Oct 28, 2026·

  ·
  4

• Skip xnn_reshape_runtime when the input shapes have not changed (#24219)

  CLA Signedmeta-exported
  #24219·billmguo opened Oct 27, 2026Contributor·

  ·
  3

• Support torch.empty_like when lowering to Edge, and speed up GroupBasedPartitioner

  CLA Signedrelease notes: exir
  #24218·shoumikhin opened Oct 27, 2026Contributor·

  ·
  9

• Build operator lookup names without snprintf

  CLA Signed
  #24217·rascani opened Oct 27, 2026Contributor·

  ·
  3

• Pin OpenVINO CI to the 2026.4.1 release instead of a deleted nightly

  CLA Signed
  #24212·pytorchbot opened Oct 27, 2026Collaborator·

  ·
  1

• Update the TensorRT backend docs page

  CLA Signed
  #24211·pytorchbot opened Oct 27, 2026Collaborator·

  ·
  1

• Fix device placement of constants, shared inputs and delegate outputs

  CLA Signed
  #24210·pytorchbot opened Oct 27, 2026Collaborator·

  ·
  1

• [PyBindings] Torch-free follow-ups: reused inputs, error reasons, bfloat16 shape, dtype names, tests, docs

  CLA Signed
  #24209·pytorchbot opened Oct 27, 2026Collaborator·

  ·
  1

• Serialize only a view constant's own bytes in XNNPACK and Vulkan

  CLA Signedmodule: vulkan
  #24208·pytorchbot opened Oct 27, 2026Collaborator·

  ·
  1

• Let a backend share one delegate across calls of the same lowered module

  CLA Signed
  #24207·pytorchbot opened Oct 27, 2026Collaborator·

  ·
  1

• Stop installing test-only packages with the executorch package

  CLA Signed
  #24206·pytorchbot opened Oct 27, 2026Collaborator·

  ·
  1

• Qualcomm AI Engine Direct - Add QNN/HTP joint-graph training toy example

  release notes: qualcomm
  #24205·likhitanavali opened Oct 27, 2026·

  ·
  6