`BetterTransformer` optimizations can't be applied to `Falcon`

[pcuenca]

System Info

Python 3.10, optimum @ main, transformers @ main

Who can help?

@fxmarty

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)

Reproduction:

from transformers import AutoTokenizer, AutoModelForCausalLM
from optimum.bettertransformer import BetterTransformer
import torch

model_id = "tiiuae/falcon-rw-1b"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)
model = BetterTransformer.transform(model)

inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=10)

Falcon attention was refactored in huggingface/transformers@05ea7b7#diff-81c616a9db6f569c579ccf03c30c2f69aa7b65fa40959ac7e882fb8d541891d7. This removed the property maybe_rotary and adopted llama conventions for rotary embeddings.

We could modify the use of maybe_rotary here by using something like:

        submodules = ["query_key_value", "dense", "attention_dropout"]
        if not config.alibi:
            submodules.append("rotary_emb")

And then we'd need to adapt the code here, applying rotary embeddings when alibi is not in use.

Expected behavior

Transformation would work.

[fxmarty]

Hi, Falcon with SDPA is supported by default in Transformers now huggingface/transformers#26572, and we deprecate the usage of BetterTransformer for this architecture.

See https://huggingface.co/docs/transformers/perf_infer_gpu_one#flashattention-and-memory-efficient-attention-through-pytorchs-scaleddotproductattention