[Optimum export ONNX] Bug when export model TrOCR with dtype BF16

[vu0607]

System Info

optimum: `1.17.1`
onnx: `1.15.0`
onnxruntime-gpu: `1.17.1` 
torch: `1.12.0+cu113`
CUDA: `11.4`

Who can help?

I meet bug when export model TrOCR to ONNX dtype bf16 (support with optimum 1.17.0): INVALID_GRAPH : Load model from output/encoder_model.onnx failed:This is an invalid model. Type Error: Type 'tensor(bfloat16)' of input parameter (pixel_values) of operator (Conv) in node (/embeddings/patch_embeddings/projection/Conv) is invalid.
Mr @michaelbenayoun Hope you can consider and support me.
Beside I want ask about plan support export ONNX dtype float8 in the future. I see quanto support quantize model to float8

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction (minimal, reproducible, runnable)

optimum-cli export onnx \
      --task image-to-text-with-past \
      --model "microsoft/trocr-small-handwritten" "output" \
      --device cuda \
      --dtype bf16 \
      --batch_size 8 \
      --width 384 \
      --height 384 \
      --num_channels 3

Expected behavior

successfully export BF16 and plan about dtype FP8