ncnn啥时候考虑全面兼容pytorch或onnx算子？

[Baiyuetribe]

日志

9成新模型，都无法使用pnnx正常导出

模型

f5-tts

复现步骤

% pnnx s.pt inputshape="[1,1226,1000],[2,16,1226,64],[2,16,1226,64],[1,1226,612],[1,1226,612],[1]i32"
pnnxparam = s.pnnx.param
pnnxbin = s.pnnx.bin
pnnxpy = s_pnnx.py
pnnxonnx = s.pnnx.onnx
ncnnparam = s.ncnn.param
ncnnbin = s.ncnn.bin
ncnnpy = s_ncnn.py
fp16 = 1
optlevel = 2
device = cpu
inputshape = [1,1226,1000]f32,[2,16,1226,64]f32,[2,16,1226,64]f32,[1,1226,612]f32,[1,1226,612]f32,[1]i32
inputshape2 = 
customop = 
moduleop = 
############# pass_level0
inline module = f5_tts.model.backbones.dit.DiT
inline module = f5_tts.model.backbones.dit.InputEmbedding
inline module = f5_tts.model.modules.AdaLayerNorm
inline module = f5_tts.model.modules.AdaLayerNorm_Final
inline module = f5_tts.model.modules.Attention
inline module = f5_tts.model.modules.ConvPositionEmbedding
inline module = f5_tts.model.modules.DiTBlock
inline module = f5_tts.model.modules.FeedForward
inline module = f5_tts.model.backbones.dit.DiT
inline module = f5_tts.model.backbones.dit.InputEmbedding
inline module = f5_tts.model.modules.AdaLayerNorm
inline module = f5_tts.model.modules.AdaLayerNorm_Final
inline module = f5_tts.model.modules.Attention
inline module = f5_tts.model.modules.ConvPositionEmbedding
inline module = f5_tts.model.modules.DiTBlock
inline module = f5_tts.model.modules.FeedForward

----------------

libc++abi: terminating due to uncaught exception of type std::runtime_error: The following operation failed in the TorchScript interpreter.
Traceback of TorchScript, serialized code (most recent call last):
  File "code/__torch__/f5_tts/model/backbones/dit.py", line 115, in forward
    proj = self.proj
    input = torch.cat([noise, cat_mel_text], -1)
    _25 = (proj).forward(input, )
           ~~~~~~~~~~~~~ <--- HERE
    _26 = torch.add((conv_pos_embed).forward(_25, ), _25)
    return _26
  File "code/__torch__/torch/nn/modules/linear/___torch_mangle_25.py", line 12, in forward
    bias = self.bias
    weight = self.weight
    return torch.linear(input, weight, bias)
           ~~~~~~~~~~~~ <--- HERE
  def forward1(self: __torch__.torch.nn.modules.linear.___torch_mangle_25.Linear,
    input: Tensor) -> Tensor:

Traceback of TorchScript, original code (most recent call last):
/Users/baiyue/Library/Python/3.10/lib/python/site-packages/torch/nn/modules/linear.py(117): forward
/Users/baiyue/Library/Python/3.10/lib/python/site-packages/torch/nn/modules/module.py(1543): _slow_forward
/Users/baiyue/Library/Python/3.10/lib/python/site-packages/torch/nn/modules/module.py(1562): _call_impl
/Users/baiyue/Library/Python/3.10/lib/python/site-packages/torch/nn/modules/module.py(1553): _wrapped_call_impl
/Users/baiyue/work/F5-TTS-ONNX/F5_TTS/src/f5_tts/model/backbones/dit.py(86): forward
/Users/baiyue/Library/Python/3.10/lib/python/site-packages/torch/nn/modules/module.py(1543): _slow_forward
/Users/baiyue/Library/Python/3.10/lib/python/site-packages/torch/nn/modules/module.py(1562): _call_impl
/Users/baiyue/Library/Python/3.10/lib/python/site-packages/torch/nn/modules/module.py(1553): _wrapped_call_impl
/Users/baiyue/work/F5-TTS-ONNX/F5_TTS/src/f5_tts/model/backbones/dit.py(185): forward
/Users/baiyue/Library/Python/3.10/lib/python/site-packages/torch/nn/modules/module.py(1543): _slow_forward
/Users/baiyue/Library/Python/3.10/lib/python/site-packages/torch/nn/modules/module.py(1562): _call_impl
/Users/baiyue/Library/Python/3.10/lib/python/site-packages/torch/nn/modules/module.py(1553): _wrapped_call_impl
/Users/baiyue/work/F5-TTS-ONNX/Export_ONNX/F5_TTS/Export_F5.py(318): forward
/Users/baiyue/Library/Python/3.10/lib/python/site-packages/torch/nn/modules/module.py(1543): _slow_forward
/Users/baiyue/Library/Python/3.10/lib/python/site-packages/torch/nn/modules/module.py(1562): _call_impl
/Users/baiyue/Library/Python/3.10/lib/python/site-packages/torch/nn/modules/module.py(1553): _wrapped_call_impl
/Users/baiyue/Library/Python/3.10/lib/python/site-packages/torch/jit/_trace.py(1275): trace_module
/Users/baiyue/Library/Python/3.10/lib/python/site-packages/torch/jit/_trace.py(695): _trace_impl
/Users/baiyue/Library/Python/3.10/lib/python/site-packages/torch/jit/_trace.py(1000): trace
/Users/baiyue/work/F5-TTS-ONNX/Export_ONNX/F5_TTS/Export_F5.py(556): <module>
RuntimeError: mat1 and mat2 shapes cannot be multiplied (1226x1612 and 712x1024)

zsh: abort      pnnx s.pt