Pull requests

• riscv: fix ShuffleChannel scratch buffer safety

  riscv
  #7160·luke-opus opened Oct 21, 2026Contributor·

  ·
  1

• pnnx: escape python keywords in generated module attribute names

  pnnxtool
  #7159·drakeo338 opened Oct 21, 2026·

  ·
  1

  2

• pnnx: support positive-step tensor slicing with StridedSlice

  coredoclayerpnnxtesttool
  #7158·Edwardssss opened Oct 21, 2026Contributor·

  ·
  1

• riscv: vectorize fp32 InnerProduct for unpacked 2D inputs

  riscv
  #7155·luke-opus opened Oct 20, 2026Contributor·

  ·
  1

• riscv: reuse adjacent columns in fp32 5x5s2 depthwise

  riscv
  #7154·luke-opus opened Oct 20, 2026Contributor·

  ·
  2

• riscv: reuse four output rows in fp16 5x5s1 depthwise

  riscv
  #7153·luke-opus opened Oct 20, 2026Contributor·

  ·
  1

• Validate runtime blob shapes in GRU, RNN, LSTM, DeformableConv2D, Gemm and MultiHeadAttention, and check int8 scale records

  armlayerloongarchmipsriscvtestx86
  #7152·professor-moody opened Oct 20, 2026Contributor·

  ·
  6

• vulkan: add a barrier before record_clone writes the dst buffer

  core
  #7151·kongxinjun opened Oct 20, 2026Contributor·

  ·
  2

• fix integer overflow and missing dimension checks in yuv420sp pixel apis

  core
  #7149·ihb2032 opened Oct 20, 2026Contributor·2 tasks done·

  ·
  1

  1

• Validate multi-blob layer arity, ModelBin shape products, and layer weight/parameter invariants

  armcorelayerloongarchmipsriscvtestvulkanx86
  #7148·professor-moody opened Oct 19, 2026Contributor·

  ·
  10

• pnnx: add Caffe frontend

  pnnxtool
  #7146·nihui opened Oct 19, 2026Member·

  ·
  1

• vulkan: single pass square sum for Normalize on subgroups

  testvulkan
  #7143·futz12 opened Oct 19, 2026Contributor·

  ·
  3

• vulkan: reduce the flash attention row statistics on subgroups

  coretestvulkan
  #7142·futz12 opened Oct 19, 2026Contributor·

  ·
  4

• softmax vulkan subgroup reduce optimization

  testvulkan
  #7141·futz12 opened Oct 19, 2026Contributor·

  ·
  1

• x86: implement softplus with SIMD

  x86
  #7138·losewayy opened Oct 18, 2026Contributor·

  ·
  3

• pnnx: add ncnn optimize functionality reusing pass_ncnn pipeline

  pnnxtool
  #7132·futz12 opened Oct 17, 2026Contributor·

  ·
  4

• Fix GridSample nearest tie rounding on CPU

  layerloongarchmipstestx86
  #7131·FOWEPJF255 opened Oct 17, 2026·

  ·
  1

  9

• support concat batch inference

  armdoclayerloongarchmipspnnxriscvtesttoolvulkanx86
  #7129·futz12 opened Oct 16, 2026Contributor·

  ·
  3

• pnnx: preserve LayerNorm scale when bias is disabled

  pnnxtool
  #7128·ktz03 opened Oct 16, 2026Contributor·

  ·
  1

  1

• pnnx: reject unsupported RNN nonlinearities

  pnnxtool
  #7125·ktz03 opened Oct 16, 2026Contributor·

  ·
  1

  1

• pnnx: reject unsupported MHA appended key/value options

  pnnxtool
  #7123·ktz03 opened Oct 16, 2026Contributor·

  ·
  1

  1

• pnnx: preserve max_norm in Embedding module fusion

  pnnxtool
  #7118·ktz03 opened Oct 15, 2026Contributor·

  ·
  1

  10

• pnnx: preserve scalar division precision in exported python

  pnnxtool
  #7116·Edwardssss opened Oct 15, 2026Contributor·

  ·
  1

  3

• pnnx: preserve average pooling ceil tail divisors

  pnnxtool
  #7114·ktz03 opened Oct 15, 2026Contributor·

  ·
  1

  4

• riscv: add bf16 gemm with optional zvfbfwma

  cmakecoreriscvtest
  #7112·nihui opened Oct 14, 2026Member·

  ·
  6