Issues

• PNNX average pooling counts implicit ceil tail in the divisor

  Tencent/ncnn#7113 · Filter by author ktz03 opened on Oct 15, 2026

• pnnx rewrites unsupported torch.norm orders to a loadable Reduction

  Tencent/ncnn#7108 · Filter by author ktz03 opened on Oct 14, 2026

• pnnx scales the wrong operand for scalar torch.rsub alpha

  Tencent/ncnn#7106 · Filter by author ktz03 opened on Oct 14, 2026

• pnnx changes torch.div floor and trunc rounding semantics

  Tencent/ncnn#7104 · Filter by author ktz03 opened on Oct 14, 2026

• pnnx drops nonzero decimals when converting torch.round

  Tencent/ncnn#7101 · Filter by author ktz03 opened on Oct 14, 2026

• pnnx: zero-alpha torch.addmm conversion can silently produce wrong output

  Tencent/ncnn#7099 · Filter by author ktz03 opened on Oct 14, 2026

• SIGSEGV in Vulkan compute encoder on Adreno 840 (8 Elite Gen 5) - deterministic, and it used to work

  Tencent/ncnn#7097 · Filter by author fakingpescobar-ctrl opened on Oct 13, 2026

• squeezenet 示例在 arm cpu 上输出 nan（fp16 winograd63 输入变换溢出）

  Tencent/ncnn#7096 · Filter by author kongxinjun opened on Oct 13, 2026

• pnnx loses full reductions with empty TorchScript dimensions

  Tencent/ncnn#7094 · Filter by author ktz03 opened on Oct 13, 2026

• We need deformable attention support

  Tencent/ncnn#7089 · Filter by author futz12 opened on Oct 11, 2026

• pnnx chunk and tensor_split use incorrect uneven partition sizes

  Tencent/ncnn#7071 · Filter by author ktz03 opened on Oct 9, 2026

• pnnx: torch.roll with repeated dimensions produces wrong output

  Tencent/ncnn#7066 · Filter by author ktz03 opened on Oct 8, 2026

• pnnx loses the Softplus threshold during ncnn conversion

  Tencent/ncnn#7064 · Filter by author ktz03 opened on Oct 8, 2026

• pnnx silently changes circular padding to reflect or constant padding

  Tencent/ncnn#7061 · Filter by author ktz03 opened on Oct 8, 2026

• pnnx: dynamic torch.stack emits a reshape with multiple inferred dimensions

  Tencent/ncnn#7059 · Filter by author ktz03 opened on Oct 8, 2026

• Vulkan GPU detections wrong on Mali-G715 (driver 54.x, fp16 only) and Adreno 650 (fp16 and fp32), ncnn-android-nanodet, ncnn 20260526

  Tencent/ncnn#7054 · Filter by author kekpan opened on Oct 8, 2026

• Reduction LogSumExp overflows or underflows for finite FP32 inputs

  Tencent/ncnn#7053 · Filter by author ktz03 opened on Oct 8, 2026

• [SECURITY] 公开 API yuv420sp2rgb w*h 整数溢出：vuptr 负偏移越界读（绘制系列代码层同型越界写）

  Tencent/ncnn#7035 · Filter by author JerryGW opened on Oct 1, 2026
  1

• [SECURITY] MatMul 层跨输入 K 维一致性缺失：A.w > B.w 时 gemm B 矩阵打包越界读

  Tencent/ncnn#7034 · Filter by author JerryGW opened on Oct 1, 2026
  1

• load_param_bin 加载含未量化层（int8_scale_term=0）的 ncnn2int8 部分量化模型时推理输出错误——同一模型的文本 load_param 加载结果正确[BUG]

  Tencent/ncnn#7031 · Filter by author zengjie617789 opened on Sep 30, 2026
  1

• Heap-buffer-overflow in Concat::forward (AVX and generic x86 kernel) with multi-scale YOLO export

  Tencent/ncnn#7025 · Filter by author sergshadrap opened on Sep 25, 2026
  5

• AMD CPU SDPA result error

  Tencent/ncnn#6987 · Filter by author DickyQi opened on Sep 10, 2026
  1

• feature plan

  enhancement
  Tencent/ncnn#6981 · Filter by author nihui opened on Sep 9, 2026

• Vulkan compute output corrupted on Imagination/PowerVR with opt.lightmode=1 blob recycling — allocator bookkeeping proven correct by internal tracing during a corrupted inference; barriers and submit boundaries ruled out

  Tencent/ncnn#6935 · Filter by author di2g opened on Aug 29, 2026
  2

• ARM fp16 arithmetic produces all-NaN output for a ViT-tiny graph / model (ncnn 20260526)

  Tencent/ncnn#6932 · Filter by author cesarpgouveia opened on Aug 25, 2026
  1