forward出现pnnx.param的内容

[1311523821]

error log | 日志或报错信息 | ログ

pnnxparam = ./pnnx/depth_anything_v2_vits.pnnx.param
pnnxbin = ./pnnx/depth_anything_v2_vits.pnnx.bin
pnnxpy = ./pnnx/depth_anything_v2_vits_pnnx.py
pnnxonnx = ./pnnx/depth_anything_v2_vits.pnnx.onnx
ncnnparam = ./pnnx/depth_anything_v2_vits.ncnn.param
ncnnbin = ./pnnx/depth_anything_v2_vits.ncnn.bin
ncnnpy = ./pnnx/depth_anything_v2_vits_ncnn.py
fp16 = 0
optlevel = 0
device = cpu
inputshape = [1,3,532,532]f32
inputshape2 =
customop =
moduleop =
get inputshape from traced inputs
inputshape = [1,3,532,532]f32
############# pass_level0
inline module = depth_anything_v2.dinov2_layers.attention.MemEffAttention
inline module = depth_anything_v2.dinov2_layers.block.NestedTensorBlock
inline module = depth_anything_v2.dinov2_layers.layer_scale.LayerScale
inline module = depth_anything_v2.dinov2_layers.mlp.Mlp
inline module = depth_anything_v2.dinov2_layers.patch_embed.PatchEmbed
inline module = depth_anything_v2.dpt.DPTHead
inline module = depth_anything_v2.util.blocks.FeatureFusionBlock
inline module = depth_anything_v2.util.blocks.ResidualConvUnit
inline module = torch.nn.modules.linear.Identity
inline module = depth_anything_v2.dinov2_layers.attention.MemEffAttention
inline module = depth_anything_v2.dinov2_layers.block.NestedTensorBlock
inline module = depth_anything_v2.dinov2_layers.layer_scale.LayerScale
inline module = depth_anything_v2.dinov2_layers.mlp.Mlp
inline module = depth_anything_v2.dinov2_layers.patch_embed.PatchEmbed
inline module = depth_anything_v2.dpt.DPTHead
inline module = depth_anything_v2.util.blocks.FeatureFusionBlock
inline module = depth_anything_v2.util.blocks.ResidualConvUnit
inline module = torch.nn.modules.linear.Identity

---

############# pass_level1
############# pass_level2
Traceback (most recent call last):
File "e:\AI\Code\WorkCode\Depth-Anything-V2\export_ncnn.py", line 124, in
export_depth_anything_v2(
File "e:\AI\Code\WorkCode\Depth-Anything-V2\export_ncnn.py", line 52, in export_depth_anything_v2
pnnx.export(model, export_path_base,
File "D:\ProgramData\anaconda3\envs\bridgedepth\lib\site-packages\pnnx\utils\export.py", line 22, in export
return convert(ptpath, inputs, inputs2, input_shapes, input_types, input_shapes2, input_types2, device, customop, moduleop, optlevel, pnnxparam, pnnxbin, pnnxpy, pnnxonnx, ncnnparam, ncnnbin, ncnnpy, fp16)
File "D:\ProgramData\anaconda3\envs\bridgedepth\lib\site-packages\pnnx\utils\convert.py", line 160, in convert
spec.loader.exec_module(foo)
File "", line 879, in exec_module
File "", line 1017, in get_code
File "", line 947, in source_to_code
File "", line 241, in _call_with_frames_removed
File "E:\AI\Code\WorkCode\Depth-Anything-V2./pnnx/depth_anything_v2_vits_pnnx.py", line 458
v_97 = prim::Constant(value=14)

depth_anything_v2_vits_pnnx.py
def forward(self, v_x_1):
v_97 = prim::Constant(value=14)
v_147 = prim::Constant(value=-1)
v_154 = prim::Constant(value=1)
v_157 = prim::Constant(value=2147483647)
v_190 = prim::Constant(value=37)
v_pos_embed_1 = self.pretrained_data
v_cls_token_1 = self.pretrained_data
v_94 = v_x_1.size(dim=2)
v_95 = prim::NumToTensor(v_94)
v_patch_h_1 = aten::floor_divide(v_95, v_97)
v_101 = aten::Int(v_patch_h_1)
v_104 = aten::Int(v_patch_h_1)
v_107 = aten::Int(v_patch_h_1)
v_110 = aten::Int(v_patch_h_1)
v_113 = v_x_1.size(dim=3)
v_114 = prim::NumToTensor(v_113)
v_1468 = prim::Constant(value=14)

context | 编译/运行环境 | バックグラウンド

torch 2.7.0+cu118
pnnx 20251016
ncnn 1.0.20250916
Python 3.10.18

[dosubot]

你遇到的 prim::Constant 出现在 forward 函数（如 v_97 = prim::Constant(value=14)），本质上是 pnnx 在导出 PyTorch 计算图时，将 JIT IR 里的常量节点直接转成了 Python 代码。正常情况下，这些常量应该会被 constant folding 优化掉，或者被转成普通的 Python 常量表达式，但如果模型结构复杂、包含动态 shape、特殊索引、未支持的 PyTorch 操作，或者 pnnx 优化 pass 没有完全覆盖，就会遗留 prim::Constant 这样的节点，导致生成的 Python 代码无法直接运行或报错（比如你遇到的行 458 错误）相关源码分析。

常见原因包括：输入 shape/dtype 不匹配、模型有零维参数、chunk 操作维度不对、模型用到了 pnnx 暂不支持的 PyTorch 高级用法或自定义模块等类似问题与建议。

排查建议：

1. 检查模型输入 shape 和 dtype，确保没有异常值。
2. 尝试用 moduleop 参数让 pnnx 保留部分复杂模块为整体算子，避免内联失败。
3. 检查模型是否包含未支持的 PyTorch 操作或自定义 C++/CUDA op。
4. 如果是 view/reshape/concat 等 shape 相关报错，手动修改 pnnx 导出的 Python 脚本（如 depth_anything_v2_vits_pnnx.py），调整相关参数后重新导出。
5. 如果 torch.jit.trace 导出不稳定，可以尝试 torch.jit.script 或先导出 ONNX 再用 pnnx 转换。

这种问题在导出 YOLOv8、Depth-Anything 等复杂模型时较常见，pnnx 的优势是可以生成可编辑的 Python 脚本，允许你手动修正 shape、常量、拼接等细节，再重新导出，最终保证 ncnn 能正常推理详细原理和流程。

如需进一步定位，可贴出出错的 forward 片段和相关 shape 信息。

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}