Issues

• OSS-Fuzz integration for ncnn?

  Tencent/ncnn#7186 · Filter by author JayOfTheKeyboard opened on Oct 27, 2026

• Vulkan: VkTransfer::record_upload dereferences an empty staging buffer after allocation failure

  Tencent/ncnn#7185 · Filter by author yeliannaa opened on Oct 27, 2026

• pnnx Python convert ignores fp16=False with input_shapes

  Tencent/ncnn#7183 · Filter by author ktz03 opened on Oct 27, 2026

• pnnx: manual tanh GELU fusion loses the approximation mode

  Tencent/ncnn#7175 · Filter by author ktz03 opened on Oct 26, 2026

• test_squeezenet.cpp crashes: SEGV due to a vector being zero sized

  Tencent/ncnn#7171 · Filter by author yurivict opened on Oct 24, 2026
  1

• pnnx: complex scalar expressions lose Float32 component precision

  Tencent/ncnn#7168 · Filter by author ktz03 opened on Oct 24, 2026

• PNNX: torch.var(correction=2) generates an invalid unbiased keyword

  Tencent/ncnn#7165 · Filter by author ktz03 opened on Oct 23, 2026
  1

• Windows Python process exits with 0xC0000005 after Vulkan GPU enumeration or inference

  Tencent/ncnn#7157 · Filter by author grikdotnet opened on Oct 20, 2026

• Vulkan GELU ignores fast_gelu=0 and computes tanh approximation instead of exact GELU

  Tencent/ncnn#7156 · Filter by author grikdotnet opened on Oct 20, 2026

• pnnx: generated Python is invalid for a submodule named from

  Tencent/ncnn#7150 · Filter by author ktz03 opened on Oct 20, 2026

• pnnx drops learned LayerNorm weight when bias=False

  Tencent/ncnn#7127 · Filter by author ktz03 opened on Oct 16, 2026

• GridSample nearest CPU sampling rounds half-pixel coordinates upward instead of to even

  Tencent/ncnn#7126 · Filter by author ktz03 opened on Oct 16, 2026

• 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

• 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
  1

• 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