Change Transformers export default sequence length to max_position_embeddings

[mgoin]

Usage

sparseml.transformers.export_onnx --model_path ./TinyLlama-1.1B-Chat-v0.3 --task text-generation --trust_remote_code 
2023-11-13 18:56:23 sparseml.transformers.export INFO     Attempting onnx export for model at /home/mgoin/models/TinyLlama-1.1B-Chat-v0.3 for task text-generation
2023-11-13 18:56:23 sparseml.transformers.export INFO     Using default sequence length of 2048
Special tokens have been added in the vocabulary, make sure the associated word embeddings are fine-tuned or trained.

Reason

Unfortunately I think it is very important to export LLMs with the largest sequence length possible.

For instance I was exporting llama models with a sequence length of 512 for convenience
sparseml.transformers.export_onnx --model_path ./llama2.c-stories15M --task text-generation --sequence_length 512
However due to the PyTorch export and constant folding it seems to produce a cached, fixed rotary embedding of the sequence length used. See it as 512 in this screenshot. The ONNX can still run and compile with a sequence length larger than 512, but the output is unstable and quickly starts to repeat

[bfineran]

LGTM - might want to have a fallback default in case max_position_embeddings is never set, but it may just be the case that HF has a default already

[mgoin]

max_position_embeddings

The config class itself has a default value for this, so I'm inclined to leave it be. We could raise ValueError but that will be reflected downstream