All pull requests

Pull requests

• [LLM:Bugfix] Propagate text tokenizer vocabulary load failures

  #5034·kerry0520 opened Oct 23, 2026Contributor·5 of 12 tasks·

  ·

• [PyMNN:Bugfix] Honor shape contracts in array construction and composition

  module:pymnntype:bug
  #5030·hulkbig opened Oct 22, 2026Contributor·

  ·
  

• [PyMNN:Bugfix] Preserve dimensions in statistical reductions

  module:pymnntype:bug
  #5029·hulkbig opened Oct 22, 2026Contributor·

  ·
  

• [CPU:Bugfix] Preserve C4 quantization alignment across RVV chunks

  module:quantizationtype:bug
  #5024·luke-opus opened Oct 21, 2026Collaborator·4 of 12 tasks·

  ·
  

• [CPU:Bugfix] Match RVV Float2Int8 rounding to scalar

  module:quantizationtype:bug
  #5023·luke-opus opened Oct 21, 2026Collaborator·4 of 12 tasks·

  ·
  

• [CPU:Bugfix] Use the unary int8 table only for int8 inputs

  backend:metalmodule:quantizationtype:bug
  #5022·kongxinjun opened Oct 21, 2026·

  ·
  1

  

• [CPU:Perf] Skip padding work in dilated convolutions

  type:performance
  #5019·patrickkabwe opened Oct 20, 2026Contributor·5 of 12 tasks·

  ·
  1

  

• [CPU:Feature] Integrate KleidiAI FP16 INT4 kernels

  module:quantizationtype:feature
  #5009·HenryDen opened Oct 20, 2026Contributor·12 tasks·

  ·
  

• [Hexagon:Perf] Optimize V79 INT4 block-wise prefill

  module:quantizationtype:bugtype:performance
  #5001·yydhYYDH opened Oct 19, 2026Collaborator·6 of 12 tasks·

  ·
  

• [CPU:Perf] Use 1x1 convolution when only the center tap is live

  type:performance
  #5000·patrickkabwe opened Oct 19, 2026Contributor·6 of 12 tasks·

  ·
  1

  

• [iOS:Feature] MNN Chat: shared MNNChatCore package, Agent mode and chat experience overhaul

  app:transformermodule:llmplatform:iostype:feature
  #4991·Tangesion opened Oct 19, 2026Collaborator·

  ·
  1

  

• [QNN:Bugfix] Support optional ONNX Slice inputs

  awaiting contributorbackend:qnntype:bug
  #4990·merller opened Oct 19, 2026·5 of 12 tasks·

  ·
  1

  3

  

• [CPU:Bugfix] 修复 packed FP16 RMSNorm 与单行路径的数值不一致

  awaiting contributortype:bug
  #4988·vklonghml opened Oct 19, 2026Contributor·

  ·
  4

  

• [Vulkan:Feature] Add experimental ARM NX MatMul and INT8 Linear

  backend:vulkanmodule:quantizationtype:feature
  #4976·wangzhaode opened Oct 16, 2026Collaborator·

  ·
  1

  

• [Core:Bugfix] Track floating-point shape inputs in Module resize cache

  awaiting contributor
  #4974·vklonghml opened Oct 15, 2026Contributor·5 of 12 tasks·

  ·
  3

  

• [Python] Fix audio bindings and document the public audio API

  #4954·ZhuoenX opened Oct 11, 2026Contributor·

  ·
  1

  

• [CUDA:Bugfix] Fall back to PTX for GPUs newer than the toolkit

  awaiting contributor
  #4951·Qxinyu opened Oct 10, 2026Collaborator·3 of 12 tasks·

  ·
  3

  

• [Converter:Bugfix] Honor ONNX ArgMax and ArgMin last-index selection

  awaiting contributormodule:convertertype:bug
  #4945·hulkbig opened Oct 9, 2026Contributor·5 of 12 tasks·

  ·
  1

  

• [CPU:Bugfix] Validate mmap weight cache manifests

  awaiting contributortype:bug
  #4926·kerry0520 opened Oct 2, 2026Contributor·5 of 12 tasks·

  ·
  14

  

• [CPU:Perf] Optimize RVV memory access for packing and depthwise convolution

  awaiting contributortype:performance
  #4913·typer-J opened Sep 24, 2026Contributor·

  ·
  1

  

• [CPU] Optimize large 4-byte raster bitcopy broadcasts with RVV

  awaiting contributortype:performance
  #4907·typer-J opened Sep 23, 2026Contributor·

  ·
  1

  

• [CPU:Bugfix] Fix DFlash cross-attention KV truncation

  type:bug
  #4898·yydhYYDH opened Sep 20, 2026Collaborator·5 of 12 tasks·

  ·
  

• [CPU:Perf] Vectorize the conv int8 bias-float init with a corrected reciprocal

  awaiting contributormodule:quantizationtype:performance
  #4893·typer-J opened Sep 19, 2026Contributor·

  ·
  2

  

• [CPU:Perf] Vectorize the conv int8 bias-float init with a reciprocal

  awaiting contributormodule:quantizationtype:performance
  #4892·typer-J opened Sep 19, 2026Contributor·

  ·
  1

  

• [CPU:Perf] Add an RVV kernel for conv int8 weight kernel sum

  awaiting contributormodule:quantizationtype:performance
  #4865·typer-J opened Sep 12, 2026Contributor·

  ·
  5