All pull requests

Pull requests

• [2.15] Backport #4882: install the matching runtime companion with the executorch extra

  cla signedcomponent: api [Python]component: build systemcomponent: testsdocumentation
  #4913·shoumikhin opened Oct 21, 2026Contributor·

  ·

• [2.15] Backport #4881: do not retrace by default when saving a .pte

  cla signedcomponent: api [C++]component: api [Python]component: testsdocumentation
  #4912·shoumikhin opened Oct 21, 2026Contributor·

  ·

• [2.15] Backport #4821: store each engine once, as named data, when saving a .pte

  cla signedcomponent: api [C++]component: api [Python]component: tests
  #4911·shoumikhin opened Oct 21, 2026Contributor·

  ·

• [2.15] Backport #4839: docs: add an ExecuTorch deployment guide

  cla signedcomponent: testsdocumentation
  #4910·shoumikhin opened Oct 21, 2026Contributor·

  ·

• [2.15] Backport #4833: allow CUDA graph replay on shared activation scratch

  cla signedcomponent: api [C++]component: tests
  #4909·shoumikhin opened Oct 21, 2026Contributor·

  ·

• fix(runtime): create the warm-up inputs before unflattening them under whole-graph CUDA graphs

  cla signedcomponent: api [Python]component: corecomponent: dynamocomponent: runtimecomponent: tests
  #4907·narendasan opened Oct 21, 2026Collaborator·

  ·

• fix(runtime): keep the output structure under whole-graph CUDA graphs

  cla signedcomponent: api [Python]component: corecomponent: dynamocomponent: runtimecomponent: tests
  #4906·narendasan opened Oct 21, 2026Collaborator·

  ·

• fix(runtime): flatten the inputs by the input spec under whole-graph CUDA graphs

  cla signedcomponent: api [Python]component: corecomponent: dynamocomponent: runtimecomponent: tests
  #4905·narendasan opened Oct 21, 2026Collaborator·

  ·

• fix(runtime): record inputs updated in place as the caller's tensors under whole-graph CUDA graphs

  cla signedcomponent: api [Python]component: corecomponent: dynamocomponent: runtimecomponent: tests
  #4904·narendasan opened Oct 21, 2026Collaborator·

  ·

• fix(runtime): release the captured CUDA graph with the execution context

  cla signedcomponent: corecomponent: runtimecomponent: tests
  #4903·shoumikhin opened Oct 21, 2026Contributor·

  ·

• fix(executorch): make test_shared_scratch_backend safe on Jetson

  cla signedcomponent: tests
  #4902·shoumikhin opened Oct 21, 2026Contributor·

  ·

• fix(conversion): treat a 0 in reshape dims as an empty dimension

  cla signedcomponent: api [Python]component: conversioncomponent: converterscomponent: corecomponent: dynamocomponent: tests
  #4901·narendasan opened Oct 21, 2026Collaborator·

  ·

• fix(conversion): clamp a negative slice start on a dynamic axis

  cla signedcomponent: api [Python]component: conversioncomponent: converterscomponent: corecomponent: dynamocomponent: tests
  #4900·narendasan opened Oct 21, 2026Collaborator·

  ·

• fix(lowering): decompose unfold with a gather instead of as_strided

  cla signedcomponent: api [Python]component: corecomponent: dynamocomponent: loweringcomponent: tests
  #4899·narendasan opened Oct 21, 2026Collaborator·

  ·

• perf(conversion): write a contiguous slice_scatter as a concatenation

  cla signedcomponent: api [Python]component: conversioncomponent: converterscomponent: corecomponent: dynamocomponent: loweringcomponent: testsdocumentation
  #4898·shoumikhin opened Oct 21, 2026Contributor·

  ·

• fix(conversion): keep multi-token KV writes converting across engines and off half precision

  cla signedcomponent: api [Python]component: conversioncomponent: corecomponent: dynamocomponent: tests
  #4897·narendasan opened Oct 21, 2026Collaborator·

  ·

• fix(lowering): keep dynamic dims through complex lowering

  cla signedcomponent: api [Python]component: corecomponent: dynamocomponent: loweringcomponent: tests
  #4896·narendasan opened Oct 21, 2026Collaborator·

  ·

• fix(conversion): map as_strided reads through the input's PyTorch memory layout

  cla signedcomponent: api [Python]component: conversioncomponent: converterscomponent: corecomponent: dynamocomponent: tests
  #4895·narendasan opened Oct 21, 2026Collaborator·

  ·

• fix(lowering): keep the write-back of an in-place update to a graph input

  cla signedcomponent: api [Python]component: corecomponent: dynamocomponent: loweringcomponent: tests
  #4894·narendasan opened Oct 21, 2026Collaborator·

  ·

• fix(conversion): keep the grouped_mm validator from specializing dynamic dims

  cla signedcomponent: api [Python]component: conversioncomponent: corecomponent: dynamocomponent: tests
  #4893·narendasan opened Oct 21, 2026Collaborator·

  ·

• feat(lowering): decompose scatter_reduce with include_self=False

  cla signedcomponent: api [Python]component: corecomponent: dynamocomponent: loweringcomponent: tests
  #4892·narendasan opened Oct 21, 2026Collaborator·

  ·

• fix(lowering): keep the exported signature through the complex decomposition retrace

  cla signedcomponent: api [Python]component: corecomponent: dynamocomponent: loweringcomponent: tests
  #4891·narendasan opened Oct 21, 2026Collaborator·

  ·

• fix(conversion): keep one engine output per graph output when outputs share a tensor

  cla signedcomponent: api [Python]component: conversioncomponent: corecomponent: dynamocomponent: tests
  #4890·narendasan opened Oct 21, 2026Collaborator·

  ·

• fix(conversion): keep a reshape's inferred size static when PyTorch knows it

  cla signedcomponent: api [Python]component: conversioncomponent: corecomponent: dynamocomponent: tests
  #4889·narendasan opened Oct 21, 2026Collaborator·

  ·

• fix(partitioning): widen 0/1-specialized dims before bounding derived input shapes

  cla signedcomponent: api [Python]component: corecomponent: dynamocomponent: tests
  #4888·narendasan opened Oct 21, 2026Collaborator·

  ·