All pull requests

Pull requests

• Integrate SME2 QC4W decode GEMV

  #11635·damdoo001-arm opened Oct 11, 2026Contributor·

  ·

• Enable SME2 QP8 F32 QC8W BMM dispatch

  #11634·damdoo001-arm opened Oct 11, 2026Contributor·

  ·

• [YNNPACK:SIMD] Improve float32 log accuracy with relative-error minimax [2/2] rational coefficients

  #11627·copybara-service[bot] opened Oct 9, 2026Bot·

  ·

• Adds new ARM SME2 ACLE based microkernels for QP8 F32 QC4W/QC8W GEMM and packing.

  #11623·copybara-service[bot] opened Oct 9, 2026Bot·

  ·

• Automated Code Change

  #11622·copybara-service[bot] opened Oct 9, 2026Bot·

  ·

• Use same strategy for loads in AVX512 and AVX2 for int8 k1 kernels

  #11620·copybara-service[bot] opened Oct 9, 2026Bot·

  ·

• Add f32-qc8w GEMM SVE2 kernels: LD1SB Z.S + SCVTF weight conversion via the SVE/NEON bridge.

  #11612·copybara-service[bot] opened Oct 9, 2026Bot·

  ·

• Optimize AArch64 NEON bf8 -> f16 (f8_e5m2x16 -> f16x16) conversion using ZIP1/ZIP2 with zero in YNNPACK.

  #11608·copybara-service[bot] opened Oct 8, 2026Bot·

  ·

• Add f32-qc4w-gemm AArch64 NEON fmagic (TBL2 + FSUB) microkernels (1x8, 4x8, 6x8, 8x8).

  #11599·copybara-service[bot] opened Oct 8, 2026Bot·

  ·

• skip sub-byte datatypes in get_scalar_value_as_float

  #11598·aizu-m opened Oct 8, 2026Contributor·

  ·

• [YNNPACK] Fix dot with a B operand that is broadcast in n and/or k

  #11596·kasper0406 opened Oct 8, 2026·

  ·
  1

• Added scaled_dot, a dot product that skips zero-scaled rows.

  #11593·copybara-service[bot] opened Oct 7, 2026Bot·

  ·

• Copybara import of the project:

  #11586·copybara-service[bot] opened Oct 6, 2026Bot·

  ·

• When constructing a dot node in ynnpack, kernel selection previously estimated the optimal kernel and packing layout for b using only the shape of b, leaving shape.m unset even when the row extent of a was a known constant. Without shape.m, kernel selection assumes a large row dimension and can choose a packing layout that constrains execution for small or fixed row counts.

  #11583·copybara-service[bot] opened Oct 6, 2026Bot·

  ·

• Fix slinky_thread_pool::enqueue idle-worker accounting so the destructor waits for all scheduled workers.

  #11582·copybara-service[bot] opened Oct 5, 2026Bot·

  ·

• reject out of range channel_dim in ynnpack shim channelwise define

  #11577·aizu-m opened Oct 5, 2026Contributor·

  ·

• Wrapping object files in a static archive impacts debug info which negatively impacts profile guided opts.

  #11574·copybara-service[bot] opened Oct 5, 2026Bot·

  ·

• Copybara import of the project:

  #11573·copybara-service[bot] opened Oct 5, 2026Bot·

  ·

• new pf16 8vsx8vs kernel family is integrated

  #11571·melkap01-Arm opened Oct 5, 2026Contributor·

  ·

• gate unpooling indirection reuse on channels and pixel strides

  #11567·aizu-m opened Oct 4, 2026Contributor·

  ·

• Validate every blockwise scale the qb4w kernels read

  #11566·anagnorisis2peripeteia opened Oct 4, 2026Contributor·

  ·
  4

• Require static-slice num_dims to match the input rank

  #11565·anagnorisis2peripeteia opened Oct 4, 2026Contributor·

  ·
  4

• [WIP] Add per-node control of BMM dequantization fusion and benchmark harness

  #11563·copybara-service[bot] opened Oct 4, 2026Bot·

  ·

• Re-evaluate the in-place skip on every setup of a copy operator

  #11562·anagnorisis2peripeteia opened Oct 4, 2026Contributor·

  ·
  5

• Release the weights cache descriptors when create fails

  #11561·anagnorisis2peripeteia opened Oct 4, 2026Contributor·

  ·
  3