All pull requests

Pull requests

• [executorch][cuda] Run INT5 decode linears on autotuned QuantizedGemmFamily Triton kernels and delete the last C shim (int5_plain_mm)

  CLA Signedmeta-exported
  #23516·Gasoonjia opened Oct 6, 2026Contributor·

  ·
  2

• [executorch][cuda] Run INT8 decode linears on autotuned QuantizedGemmFamily Triton kernels and delete the int8_plain_mm C shim

  CLA Signedmeta-exported
  #23515·Gasoonjia opened Oct 6, 2026Contributor·

  ·
  2

• [executorch][cuda] Run INT6 decode linears on autotuned QuantizedGemmFamily Triton kernels and delete the int6_plain_mm C shim

  CLA Signedmeta-exported
  #23514·Gasoonjia opened Oct 6, 2026Contributor·

  ·
  2

• [executorch][cuda] Run INT4 decode linears on autotuned QuantizedGemmFamily Triton kernels and delete the int4_plain_mm C shim

  CLA Signedmeta-exported
  #23513·Gasoonjia opened Oct 6, 2026Contributor·

  ·
  2

• [executorch][cuda] Add QuantizedGemmFamily: per-M quantized GEMM ops with format-owned legality and shared dispatch

  CLA Signedmeta-exported
  #23512·Gasoonjia opened Oct 6, 2026Contributor·

  ·
  2

• [executorch][cuda] Autotune data-dependent kernels on representative inputs declared by the kernel author

  CLA Signedmeta-exported
  #23511·Gasoonjia opened Oct 6, 2026Contributor·

  ·
  2

• [executorch][cuda] Time AOTI autotune candidates with CUDA graphs so a busy CPU does not decide the pick

  CLA Signedmeta-exported
  #23510·Gasoonjia opened Oct 6, 2026Contributor·

  ·
  2

• Add opaque inputs to serving layer

  CLA Signed
  #23509·metascroy opened Oct 6, 2026Contributor·

  ·
  4

• MLX: lower index_copy into a mutated user input functionally

  #23508·IlyasMoutawwakil opened Oct 6, 2026·

  ·
  4

• Support accumulating index_put in the portable in-place kernel

  CLA Signedrelease notes: ops & kernels
  #23506·JakeStevens opened Oct 6, 2026Contributor·

  ·
  1

• Keep YOLO26Model in contiguous layout after predict()

  CLA Signedcommunity: contributionrelease notes: none
  #23505·HiroForYou opened Oct 6, 2026·

  ·
  7

• Initial commit - migrate to Pyrefly

  ciflow/dockerciflow/trunkCLA Signedmodule: arm
  #23504·maggiemoss opened Oct 6, 2026Contributor·

  ·
  6

• Add 5% perf regression gate to cuda.yml benchmarks

  CLA Signedmeta-exported
  #23503·Gasoonjia opened Oct 6, 2026Contributor·

  ·
  3

• Pin the OpenVINO test harness to a release instead of a deleted nightly

  ciflow/nightlyCLA Signedmodule: cirelease notes: none
  #23502·shoumikhin opened Oct 6, 2026Contributor·

  ·
  3

• Optimized HiFi4 kernel for 4- and 6-bit packed weights

  CLA Signedmeta-exported
  #23501·mcremon-meta opened Oct 6, 2026Contributor·

  ·
  3

• Fix exit safety, package linking and header hygiene in the shared CUDA allocator

  ciflow/binariesciflow/cudaciflow/rocmciflow/trunkCLA Signedrelease notes: backends
  #23499·shoumikhin opened Oct 6, 2026Contributor·

  ·
  26

• Add CPU runtime with XNNPACK provider (#23498)

  CLA Signedmeta-exported
  #23498·JakeStevens opened Oct 6, 2026Contributor·

  ·
  3

• Add static FP32 AOT export and Recipe for CpuBackend (#23497)

  CLA Signedmeta-exported
  #23497·JakeStevens opened Oct 6, 2026Contributor·

  ·
  3

• Key lifted constants by content across partitions

  CLA Signedmeta-exported
  #23496·JakeStevens opened Oct 6, 2026Contributor·

  ·
  3

• Reserve lifted tensor constant names

  CLA Signedmeta-exported
  #23495·JakeStevens opened Oct 6, 2026Contributor·

  ·
  3

• [ET-VK][gemm] Nest linear and matmul sources by scheme

  CLA Signedmeta-exported
  #23493·SS-JIA opened Oct 6, 2026Contributor·

  ·
  2

• [ET-VK][convolution] Nest convolution sources by route

  CLA Signedmeta-exported
  #23492·SS-JIA opened Oct 6, 2026Contributor·

  ·
  2

• [ET-VK][binary] Nest binary operator sources

  CLA Signedmeta-exported
  #23491·SS-JIA opened Oct 6, 2026Contributor·

  ·
  2

• [ET-VK][codegen] Support nested shader directories

  CLA Signedmeta-exported
  #23490·SS-JIA opened Oct 6, 2026Contributor·

  ·
  2

• Fix reinplace liveness for returned tensors

  CLA Signed
  #23488·JakeStevens opened Oct 6, 2026Contributor·

  ·
  2