Add GPT-J support + benchmark results

[tomaarsen]

Closes #11

Hello!

Pull Request overview

• Add support for all GPT-J models.
• Add benchmark results for GPT-J-6B
• Add benchmarking script for GPT-J-6B

Details

As simple as

from attention_sinks import AutoModel

model = AutoModel.from_pretrained("EleutherAI/gpt-j-6b", device_map="auto")

Benchmarks

python benchmark/perplexity.py --model_name_or_path EleutherAI/gpt-j-6b --experiment attention_sinks --output_dir benchmark/outputs_gptj_6b
python benchmark/perplexity.py --model_name_or_path EleutherAI/gpt-j-6b --experiment transformers --output_dir benchmark/outputs_gptj_6b
python benchmark/perplexity.py --model_name_or_path EleutherAI/gpt-j-6b --experiment windowed --output_dir benchmark/outputs_gptj_6b

python benchmark/plot_perplexity.py --features perplexity vram --title "Log perplexity & VRAM usage of GPT-J 6B as a function of input lengths" --output_dir benchmark/outputs_gptj_6b --log_perplexity_limit 4

• Tom Aarsen