Issues

• pnnx exports ReLU RNNs as native tanh RNNs after warning

  Tencent/ncnn#7124 · Filter by author ktz03 opened on Oct 16, 2026

• pnnx drops MultiheadAttention appended key/value options in native conversion

  Tencent/ncnn#7122 · Filter by author ktz03 opened on Oct 16, 2026

• Vulkan SDPA: sdpa_cross.comp reads past the buffer when seqlen % 4 != 0, output wrong on Mali-G615

  Tencent/ncnn#7120 · Filter by author MikalaiKryvusha opened on Oct 15, 2026
  1

• pnnx drops max_norm when fusing nn.Embedding

  Tencent/ncnn#7117 · Filter by author ktz03 opened on Oct 15, 2026

• pnnx: exported python loses precision for float16 scalar-left division

  Tencent/ncnn#7115 · Filter by author Edwardssss opened on Oct 15, 2026

• 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