INVALID_ARGUMENT : unsupported conv activation mode "LeakyRelu"

[baishouwujianfei]

Describe the issue

hello, I encountered an error while using onnxruntime-gpu to start the model service through CUDA: onnxruntime.capi.onnxruntime_pybind11_state.RuntimeException: [ONNXRuntimeError] : 6 : RUNTIME_EXCEPTION : Exception during initialization: D:\a\_work\1\s\onnxruntime\contrib_ops\cuda\fused_conv.cc:19 onnxruntime::contrib::cuda::FusedConv<float>::FusedConv [ONNXRuntimeError] : 2 : INVALID_ARGUMENT : unsupported conv activation mode "LeakyRelu". The model runs normally when switched to the CPU with other parameters unchanged. How can I resolve this issue?

To reproduce

none

Urgency

general

Platform

Windows

OS Version

win 11

ONNX Runtime Installation

Built from Source

ONNX Runtime Version or Commit ID

onnxruntime-gpu 1.19.2

ONNX Runtime API

Python

Architecture

X64

Execution Provider

CUDA

Execution Provider Library Version

CUDA12.2

[skottmckay]

Seems like a bug in the Conv fusion optimizer. If the CUDA EP doesn't support LeakyRelu the optimizer shouldn't fuse the Conv and LeakyRelu if they're assigned to the CUDA EP.

[baishouwujianfei]

Seems like a bug in the Conv fusion optimizer. If the CUDA EP doesn't support LeakyRelu the optimizer shouldn't fuse the Conv and LeakyRelu if they're assigned to the CUDA EP.

"Thank you for your reply. How can I avoid this bug, or should I wait for a fix?"

[skottmckay]

It's actually setup to ignore the nodes and not fuse if it's assigned to CUDA (technically it could fuse with Relu but right now it's ignoring all activations).

onnxruntime/onnxruntime/core/optimizer/conv_activation_fusion.cc

Lines 118 to 119 in a24723d

	if (node_ep == kCudaExecutionProvider) {
	return std::nullopt;

How was the model created? Wondering if it was saved with CPU EP optimizations applied, and now when it's loaded with CUDA enabled it's invalid. i.e. the optimizer ran previously when the Conv + Activation were assigned to the CPU EP so they were fused.

[baishouwujianfei]

It's actually setup to ignore the nodes and not fuse if it's assigned to CUDA (technically it could fuse with Relu but right now it's ignoring all activations).

onnxruntime/onnxruntime/core/optimizer/conv_activation_fusion.cc

Lines 118 to 119 in a24723d

if (node_ep == kCudaExecutionProvider) {
return std::nullopt;
How was the model created? Wondering if it was saved with CPU EP optimizations applied, and now when it's loaded with CUDA enabled it's invalid. i.e. the optimizer ran previously when the Conv + Activation were assigned to the CPU EP so they were fused.

Sorry, this is someone else's model that I downloaded from Hugging Face, so I'm not sure how the model was created.
link：https://huggingface.co/seasonstudio/openvoice_tone_clone_onnx/blob/main/tone_clone_model.onnx