All pull requests

Pull requests

• Flatten shared async computations per chain before MemorySpaceAssignment.

  #50103·copybara-service[bot] opened Oct 6, 2026Bot·

  ·
  

• Duplicate shared async computations per async chain in FlattenCallGraph.

  #50102·copybara-service[bot] opened Oct 6, 2026Bot·

  ·
  

• [ROCm] Emit dynamic-slice copy thunks only on CUDA Ampere+

  #50101·magaonka-amd opened Oct 6, 2026Contributor·

  ·

• Improve erf_inv accuracy near |x| = 1.

  #50099·copybara-service[bot] opened Oct 6, 2026Bot·

  ·
  

• Skip infeasible candidate offsets when searching for sliced allocations in FindInRoot.

  #50097·copybara-service[bot] opened Oct 5, 2026Bot·

  ·

• [XLA:TPU] Internal refactoring of a TPU backend pass.

  #50096·copybara-service[bot] opened Oct 5, 2026Bot·

  ·
  

• [XLA:TPU] Rewrite f64 Log with bit level range reduction

  #50095·copybara-service[bot] opened Oct 5, 2026Bot·

  ·
  

• - Replace the previous primary/fallback approach by testing a list of algorithm strategies and choosing the first one that fits the estimated memory limits (or the most memory-efficient one if all exceed limits).

  #50094·copybara-service[bot] opened Oct 5, 2026Bot·

  ·

• [XLA] Hoist literal allocation and add __restrict in LiteralBase::Convert

  #50093·copybara-service[bot] opened Oct 5, 2026Bot·

  ·
  

• Improve float32 log1p accuracy on TPU v5p, v6e and 7x.

  #50092·copybara-service[bot] opened Oct 5, 2026Bot·

  ·
  

• [XLA:CPU] Use xla::cpu::Worker::Parallelize for DUCC FFT parallel loops.

  #50091·copybara-service[bot] opened Oct 5, 2026Bot·

  ·
  1

  

• [XLA:GPU] Only fuse 1D broadcasts along the feature dimension into cuDNN conv fusions.

  #50090·copybara-service[bot] opened Oct 5, 2026Bot·

  ·

• [XLA:GPU] ragged_all_to_all dev_kernel: fan remote GIN puts across contexts

  #50089·dlcompilers-infra-bot opened Oct 5, 2026Contributor·

  ·
  

• [XLA:GPU] ragged_all_to_all dev_kernel: fire remote GIN puts before LSA copy

  #50088·dlcompilers-infra-bot opened Oct 5, 2026Contributor·

  ·
  

• Integrate LLVM at llvm/llvm-project@62e30e3326d9

  #50087·copybara-service[bot] opened Oct 5, 2026Bot·

  ·

• [XLA] Add more sin/cos/tan tests with large inputs

  #50086·copybara-service[bot] opened Oct 5, 2026Bot·

  ·
  

• [XLA:GPU] Report oversized shared memory only on SM 10.7 to fix cuDNN deviceless plan deserialization

  #50085·dlcompilers-infra-bot opened Oct 5, 2026Contributor·

  ·
  

• [XLA] Dispatch same-type element copy/compare in xla/literal.cc by byte size and hoist index computation

  #50084·copybara-service[bot] opened Oct 5, 2026Bot·

  ·
  

• [XLA:PJRT] Add interleave/deinterleave transpose kernels for a stride-1 dimension of size 3

  #50083·copybara-service[bot] opened Oct 5, 2026Bot·

  ·
  

• PR #49980: [XLA:GPU][oneAPI] Switch sycl_device_description to cc_library for faster compile time with oneAPI 2026.1

  #50082·copybara-service[bot] opened Oct 5, 2026Bot·

  ·

• Prevent calling logical_on_device_shape() for pad_real_to_static use case.

  #50081·copybara-service[bot] opened Oct 5, 2026Bot·

  ·

• [XLA] HloDataflowAnalysis: forward loop-carried async state for async cross_buffer_slice fusions.

  #50079·copybara-service[bot] opened Oct 5, 2026Bot·

  ·
  

• [XLA] Move type-independent helpers from HloEvaluatorTypedVisitor to HloEvaluator

  #50078·copybara-service[bot] opened Oct 5, 2026Bot·

  ·
  

• Avoid using Status for control flow in HeapSimulator.

  #50077·copybara-service[bot] opened Oct 5, 2026Bot·

  ·

• Previously, if any buffer color exceeded its memory limit during fast buffer assignment, fallback cleared all allocations and ran buffer assignment for every color from scratch.

  #50074·copybara-service[bot] opened Oct 5, 2026Bot·

  ·