does M2 work with ONNX?
andersonbcdefg opened this issue · comments
i know there are special long convolutions, not sure if ONNX supports them / what would happen if I tried to export to ONNX
I'm not very familiar with ONNX - what you would need for the long convolution to do it efficiently is an FFT operation.
Out of curiosity, can you describe the use cases that need ONNX?
Sorry Dan I just saw this! Typical use cases for ONNX (although it does/can support GPUs) is edge, on-device CPU inference, which is especially desirable with embeddings, since most models for embeddings are small enough to run locally or even in the browser. A great example of this is Transformers.js, which I believe is actually the main reason people started putting ONNX checkpoints on the HuggingFace Hub.
ONNX is nice because you can run models exported by many frameworks (PyTorch, Tensorflow, etc.) on a single runtime, and there are some graph optimizations and quantization stuff that it makes really easy. Out of all the frameworks for compiling or exporting models to run fast locally (Llama.cpp, MLC, etc.), ONNX is the most mature and easiest to use, esp. for embeddings which haven't been as much of a focus of some of the newer libraries. It's been around long before the rise of LLMs. :) (Well... "long before" == 2017)
i know there are special long convolutions, not sure if ONNX supports them / what would happen if I tried to export to ONNX
I have tried to convert M2-BERT to ONNX but unfortunately, an issue with an unsupported operator came up:
torch.onnx.errors.UnsupportedOperatorError: Exporting the operator 'aten::fft_rfft' to ONNX opset version 19 is not supported. Please feel free to request support or submit a pull request on PyTorch GitHub: https://github.com/pytorch/pytorch/issues.
I think it is related to this issue
I have used onnxruntime 1.16.3, ONNX opset version 19 and torch 2.1.0.
Maybe I will have some time to deep dive into this in the future, so if anyone else wants to try, I could share my code.