Long-context-transformers

Exploring finetuning public checkpoints on filtered 8K sequences on Pile

Exmple of running 8K sequences on Pile

Single GPU and single node

CUDA_VISIBLE_DEVICES=0 HF_MODULES_CACHE=./cache/ HF_DATASETS_CACHE=./cache/ TRANSFORMERS_CACHE=./cache/ python finetune.py --per_device_train_batch_size 1 --per_device_eval_batch_size 1 --output_dir pythia-1.4b --gradient_accumulation_steps 8 --fp16 --evaluation_strategy "epoch" --max_steps 100000 --model_name_or_path EleutherAI/pythia-1.4b

Note that this self-contained script holds everything you need to run this finetuning, as long as you set up dependencies, such as flash attention correctly. For a 1.3 B model, it should work on a single A100 80G.

Multiple GPUs and Single node with DeepSpeed

HF_MODULES_CACHE=./cache/ HF_DATASETS_CACHE=./cache/ TRANSFORMERS_CACHE=./cache/ deepspeed --num_gpus=8 finetune.py --per_device_train_batch_size 1 --per_device_eval_batch_size 1 --output_dir pythia-6.9b --gradient_accumulation_steps 8 --fp16 --evaluation_strategy "epoch" --max_steps 100000 --deepspeed ds_config.json --model_name_or_path EleutherAI/pythia-6.9b

If you hit "RuntimeError: Tensors must be contiguous" , follow this simple fix and modify your deepSpeed library

Multiple GPUs and Multiple Nodes with DeepSpeed with Slurm

sbatch slurm.sh

Note that you can launch up to pythia-20B with 16 80GB A100s, aka two nodes. Since the above slurm script relies on openmpi, you should be able to generalize it to more than 2 nodes without problems.

Warnings

• For OPT, it auto pads 2 in the end, so the max position should be subtracted by 2 (e.g instead of 8192, you will have to put 8190)
• For Bloom, to support Alibi, we had to compute pesudoinverse, its backward is unfriendly to gradient checkpointing, if you see backward precision issue, try to disable gradient checkpointing.

Dependencies

Not much besides typical pytorch and transformers, the most likely issue will come from flash-attention, where you should follow exactly what the official repo, in better case, if you have the choice to use the docker provided, it will save you from many headaches.

To do:

• enable multiple GPUs and model parallel
• supporting alibi (Bloom) and normal trainable embedding (OPT)
• Instruction tuning on long context benchmarks such as Scroll

Citation

You can find the citation option under the wedge in the repo. Beyond that please make sure to cite the amazing work by the incredible Tri Dao. Without his flash-attention this repo will not be possible.

Dao, Tri, et al. "Flashattention: Fast and memory-efficient exact attention with io-awareness." arXiv preprint arXiv:2205.14135 (2022).