Issues

• [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

• 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

• aarch64：padding 之后内核分发使用了过期的 elempack → conv3x3s2_pack4_neon 越界读（非确定性 NaN / SIGSEGV），触发输入为合法的 dims=2 Mat

  Tencent/ncnn#7019 · Filter by author potato34 opened on Sep 22, 2026
  1

• 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
  1

• 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

• python 版后处理例子基本上都跑不通了。。

  Tencent/ncnn#6873 · Filter by author yst-one opened on Jul 28, 2026
  3

• __kmp_wait_4_ptr+220 Android运行初始化时有极低概率发生阻塞，引起耗电增加

  Tencent/ncnn#6872 · Filter by author sandingman opened on Jul 28, 2026
  1

• Vulkan: MatMul has no Vulkan implementation — attention models fall back to CPU mid-graph (yolo12n 1.81x vs 5-8x for CNN peers)

  Tencent/ncnn#6858 · Filter by author giuliocorradi opened on Jul 27, 2026
  2

• Vulkan: record_download() bypasses the HOST_CACHED staging buffer, so downloads run at 0.33 GB/s (12.7x fix)

  Tencent/ncnn#6857 · Filter by author giuliocorradi opened on Jul 27, 2026
  1

• __kmp_abort_process+52 Android运行有极低概率崩溃

  Tencent/ncnn#6856 · Filter by author sandingman opened on Jul 27, 2026
  1

• Access violation 0xC0000005 during Vulkan GPU initialization on Windows with RTX 5060 + Intel Graphics

  Tencent/ncnn#6843 · Filter by author yianzhou opened on Jul 22, 2026
  3

• Interp bilinear/bicubic use the wrong sampling scale for non-integer scale_factor (same as #3555, but for bilinear/bicubic)

  Tencent/ncnn#6836 · Filter by author Acture opened on Jul 20, 2026
  1

• 如何将Qwen2.5-vl-3b的模型(从HF下载的),转换为ncnn的格式?

  Tencent/ncnn#6835 · Filter by author superbayes opened on Jul 19, 2026
  1

• RotaryEmbed drops the second half of cos/sin — wrong output for 2D / vision RoPE

  Tencent/ncnn#6833 · Filter by author Acture opened on Jul 17, 2026

• YOLOv8 detection results on Mali-G720 Vulkan GPU (CPU results are correct)

  Tencent/ncnn#6818 · Filter by author taiji1985 opened on Jul 12, 2026
  1

• Runtime segfault on YOLO26n NCNN model exported with pnnx 20260704

  bug
  Tencent/ncnn#6804 · Filter by author glenn-jocher opened on Jul 4, 2026
  3

• 可以发布android 端的whl 吗？pip 官方有了编译工具，希望作者可以发布一个

  enhancement
  Tencent/ncnn#6783 · Filter by author yst-one opened on Jun 17, 2026

• Android更新到20250503版本后偶现__kmp_abort_process崩溃

  Tencent/ncnn#6775 · Filter by author epwtrq opened on Jun 8, 2026
  7

• SGEMM (im2col + GEMM) convolution path may occasionally output NaN due to uninitialized workspace memory reuse

  Tencent/ncnn#6766 · Filter by author suncan0207 opened on Jun 2, 2026
  2

• Inference results are not deterministic due to uninitialized padding memory in Mat::create()

  Tencent/ncnn#6765 · Filter by author suncan0207 opened on Jun 1, 2026
  3

• Floating point exception (core dumped)

  Tencent/ncnn#6750 · Filter by author CallmeZhangChenchen opened on May 27, 2026
  2