All pull requests

Pull requests

• vulkan: fix sdpa_cross out-of-bounds load when M or N is not a multiple of 4

  testvulkan
  #7121·MikalaiKryvusha opened Oct 15, 2026·

  ·
  1

  2

• vulkan: use roundEven for UnaryOp round

  testvulkan
  #7119·kongxinjun opened Oct 15, 2026·

  ·
  1

• pnnx: preserve max_norm in Embedding module fusion

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

  ·
  1

  6

• 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·

  ·
  3

• riscv: add bf16 detection, cast and clip

  cmakecoreriscvtest
  #7111·nihui opened Oct 14, 2026Member·

  ·
  3

• pnnx: reject unsupported torch.norm orders

  pnnxtool
  #7109·ktz03 opened Oct 14, 2026Contributor·

  ·
  1

  1

• pnnx: apply scalar rsub alpha to the input operand

  pnnxtool
  #7107·ktz03 opened Oct 14, 2026Contributor·

  ·
  1

  1

• pnnx: preserve torch.div rounding modes

  coredocpnnxtesttool
  #7105·ktz03 opened Oct 14, 2026Contributor·

  ·
  1

  9

• Add optional spine-runtime OpenMP backend for SpacemiT K3

  cmakecoredoctest
  #7103·alex-spacemit opened Oct 14, 2026·

  ·
  8

• pnnx: reject unsupported round decimals during expression fusion

  pnnxtool
  #7102·ktz03 opened Oct 14, 2026Contributor·

  ·
  1

  1

• pnnx: reject unsupported zero-alpha addmm lowering

  pnnxtool
  #7100·ktz03 opened Oct 14, 2026Contributor·

  ·
  1

  4

• pnnx: eliminate static onnx tensor sequences

  pnnxtool
  #7098·Edwardssss opened Oct 14, 2026Contributor·

  ·
  20

• pnnx: preserve full reductions for empty Torch dimensions

  pnnxtool
  #7095·ktz03 opened Oct 13, 2026Contributor·

  ·
  1

  1

• pnnx: add unpool tnn.Unsqueeze

  pnnxtool
  #7092·Edwardssss opened Oct 13, 2026Contributor·

  ·
  1

• riscv: reuse four output rows in fp32 5x5s1 depthwise

  riscv
  #7091·luke-opus opened Oct 12, 2026·

  ·
  5

• pnnx: preserve uneven torch.tensor_split boundaries

  pnnxtool
  #7088·ktz03 opened Oct 11, 2026Contributor·

  ·
  1

• pnnx: preserve uneven torch.chunk boundaries

  pnnxtool
  #7087·ktz03 opened Oct 11, 2026Contributor·

  ·
  1

• vulkan: low-register winograd43 pack4 transforms for Intel GPUs on Mesa

  vulkan
  #7085·swsammy opened Oct 10, 2026·

  ·
  1

• riscv(permute): RVV kernel for Permute

  riscv
  #7084·HougeLangley opened Oct 10, 2026Contributor·

  ·
  5

• riscv(rotaryembed): optional RotaryEmbed+Permute fusion via a layout flag

  armlayerloongarchmipstoolx86
  #7083·HougeLangley opened Oct 10, 2026Contributor·

  ·
  28

• riscv(rotaryembed): RVV kernel for RotaryEmbed

  riscv
  #7082·HougeLangley opened Oct 10, 2026Contributor·

  ·
  5

• riscv(rmsnorm): RVV kernel for RMSNorm

  riscv
  #7081·HougeLangley opened Oct 10, 2026Contributor·

  ·
  15

• riscv(sdpa): IME2 flash prefill and an M=1 RVV decode path

  riscv
  #7080·HougeLangley opened Oct 10, 2026Contributor·

  ·
  80