transformers 4.49.0 didn't correctly assist in converting pytorch model into onnx when using optimun_cli

[Kotomi-Du]

Environment info

• transformers version: 4.49.0
• Platform: Windows
• Python version: 3.12.0
• PyTorch version (GPU): 2.6.0+cu118

Library:

• text generation: @patrickvonplaten

Information

Model I am using Mistral https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2 ;
Python enviroment:

numpy==2.2.3
onnx==1.17.0
onnxruntime==1.21.0
optimum==1.24.0
tokenizers==0.21.1
torch==2.6.0+cu118  # cuda is required if converting to fp16 onnx
torchaudio==2.6.0+cu118
torchvision==0.21.0+cu118
transformers==4.49.0  # transformers==4.49.0 has a bug when transferring LLM to onnx， 4.48.3 has no issue

To reproduce

Steps to reproduce the behavior:

1. optimum-cli export onnx --model mistralai/Mistral-7B-Instruct-v0.2 --no-dynamic-axes --batch_size 1 --task text-generation-with-past --dtype=fp16 --device cuda -- onnx_models\mistralai_Mistral-7B-Instruct-v0.2
2. You will error message below

3. After adding onnx::Gather_67 manually with code snippet at https://github.com/huggingface/optimum/blob/main/optimum/exporters/onnx/convert.py#L350, another issue came

 gather_tensor = torch.tensor(0, dtype=torch.int64)
 onnx_inputs['onnx::Gather_67'] = gather_tensor.cpu().numpy()

Expected behavior

If switching transformers into version 4.48.3, there is no issues.