All pull requests

Pull requests

• Report the Vulkan, WebGPU, QNN, OpenVINO and NXP backend suites to HUD

  CLA Signedmeta-exportedmodule: vulkan
  #25068·psiddh opened Nov 14, 2026Contributor·

  ·
  4

• [executorch][cuda] Add triton::convolution and the conv routing framework

  CLA Signed
  #25066·Gasoonjia opened Nov 14, 2026Contributor·

  ·
  2

• [executorch][cuda] Add planned replacement to the Triton replacement pass

  CLA Signed
  #25065·Gasoonjia opened Nov 14, 2026Contributor·

  ·
  2

• [executorch][cuda] Deprecate triton_kernel_mode=OFF

  CLA Signed
  #25064·Gasoonjia opened Nov 14, 2026Contributor·

  ·
  2

• [executorch][cuda] Add triton::convolution and the conv routing framework

  CLA Signed
  #25063·Gasoonjia opened Nov 14, 2026Contributor·

  ·
  2

• [executorch][cuda] Add planned replacement to the Triton replacement pass

  CLA Signed
  #25062·Gasoonjia opened Nov 14, 2026Contributor·

  ·
  2

• [executorch][cuda] Deprecate triton_kernel_mode=OFF

  CLA Signed
  #25061·Gasoonjia opened Nov 14, 2026Contributor·

  ·
  2

• Read the CUDA cache counters once per autotune capture

  CLA Signed
  #25060·Gasoonjia opened Nov 14, 2026Contributor·

  ·
  7

• Write each CUDA weight once across methods

  CLA Signed
  #25059·Gasoonjia opened Nov 14, 2026Contributor·

  ·
  8

• Lower last-dim softmax through ggml_soft_max

  CLA Signedmeta-exported
  #25057·JakeStevens opened Nov 13, 2026Contributor·

  ·
  3

• Lower elu through ggml_elu

  CLA Signedmeta-exported
  #25056·JakeStevens opened Nov 13, 2026Contributor·

  ·
  3

• [Native] Add a batching executor for native-backend packages

  CLA Signed
  #25055·metascroy opened Nov 13, 2026Contributor·

  ·
  1

• cpu: lower avg_pool2d with XNNPACK

  CLA Signedmeta-exported
  #25054·JakeStevens opened Nov 13, 2026Contributor·

  ·
  3

• Reject channels-last ops in DecomposeFusedQuant before rewriting

  CLA Signedmeta-exported
  #25053·JakeStevens opened Nov 13, 2026Contributor·

  ·
  3

• [executorch][vulkan] Fetch embedding_q4gsw indices of rank 2 and 3 from textures

  CLA Signedmodule: vulkan
  #25050·pytorchbot opened Nov 13, 2026Collaborator·

  ·
  1

• [executorch][native] Keep float scales of int4 ggml embeddings exact

  CLA Signed
  #25049·pytorchbot opened Nov 13, 2026Collaborator·

  ·
  1

• [executorch][native] Record embedding op samples

  CLA Signed
  #25048·pytorchbot opened Nov 13, 2026Collaborator·

  ·
  1

• [executorch][native] Quantize nn.Embedding weights in the op sample schemes

  CLA Signed
  #25047·pytorchbot opened Nov 13, 2026Collaborator·

  ·
  2

• XNNPACK: Skip the SDPA safe-softmax guard when no row can be fully masked

  ciflow/trunkCLA Signedmodule: armrelease notes: armrelease notes: xnnpack
  #25043·JakeStevens opened Nov 13, 2026Contributor·

  ·
  1

• [XNNPACK] One polyphase pass for dense and depthwise transposed convs

  CLA Signed
  #25041·JakeStevens opened Nov 13, 2026Contributor·

  ·
  2

• Fit the reference model's ring buffer to its 32-token prefill in the Muse Glimmer export test

  CLA Signedrelease notes: nonetopic: not user facing
  #25038·shoumikhin opened Nov 13, 2026Contributor·

  ·
  2

• [XNNPACK] Lower dense transposed convs as one polyphase conv + depth to space

  CLA Signed
  #25037·JakeStevens opened Nov 13, 2026Contributor·

  ·
  2

• Support 2-bit per-channel weights (qcint2)

  CLA Signedmeta-exported
  #25036·JakeStevens opened Nov 13, 2026Contributor·

  ·
  4

• Fix invalid padding in shared pooling tests

  CLA Signedhelp wantedpartner: armrelease notes: none
  #25034·zingo opened Nov 13, 2026Collaborator·

  ·
  1

• Add quantized KV attention primitives

  ciflow/trunkCLA Signedhelp wantedmodule: armrelease notes: arm
  #25033·xingguo01 opened Nov 13, 2026Collaborator·

  ·
  1