This repository showcase a working pipeline that can be used for the NeurIPS Large Language Model Efficiency Challenge: 1 LLM + 1GPU + 1Day challenge. The aim of this challenge is to come up with strategies to finetune a large language model either on 1 Nvidia 4090 or 1 A100 (40 GB) inorder to improve the performance on a subset of HELM benchmark and on a set of secret holdout tasks.
Coming up with strategies is one thing but one needs to first have a working pipeline to fit atleast a 7B LLM in one GPU. This itself is a challenge. This repo is created to outline the setup and gotchas and is built on the shoulder of giants.
The most important thing is to get your hands on a A100 (40 GB) or a 4090. Note that in this challenege these two are separate tracks.
I will be using conda but you can use your choice of environment management tool.
conda create -n neurips-llm python==3.10.0
git clone --recurse-submodule https://github.com/ayulockin/neurips-llm-efficiency-challenge
cd neurips-llm-efficiency-challenge
Note the usage of --recurse-submodule here. The repo uses my fork of lit-gpt as a submodule. My fork is instrumented with Weights and Biases experiment tracking and model versioning capabilities. This is also in sync with the upstream repo (lit-gpt).
lit-gpt requires PyTorch 2.1 (nightly) to work. Let's install this:
pip3 install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu118
This will download CUDA driver version 11.8 if it is compatible with your installed CUDA runtime version. We will revisit CUDA again.
Let's install other requirements by lit-gpt
cd lit-gpt
pip install -r requirements.txt tokenizers sentencepiece huggingface_hub wandb>=0.12.10
cd ..
In order to finetune a large model (7B) efficiently, flash attention is a must imo. In this section we will install flash attention 2.0.
This library requires CUDA runtime >= 11.4 and PyTorch >= 1.12.
Update CUDA to 11.4 or above
The CUDA runtime and driver are two different APIs. You can check the runtime version using `nvcc --version` and the driver version using `nvidia-smi`.
If your runtime is less than 11.4, you need to update it to 11.4 or above. This runtime is also dependent on the OS (eg: Debian 10 supports till 11.4).
If you have to update the cuda runtime, follow the steps:
- Remove cuda from your system.
sudo apt-get --purge remove "cublas*" "cuda*"
- Google Nvidia Toolkit 11.x download. You will find the appropriate url with steps listed there. In my case, I was on Debian 10 and thus could install 11.4. The official instructions page for CUDA 11.4 can be found here.
wget https://developer.download.nvidia.com/compute/cuda/11.4.1/local_installers cuda-repo-debian10-11-4-local_11.4.1-470.57.02-1_amd64.deb
sudo dpkg -i cuda-repo-debian10-11-4-local_11.4.1-470.57.02-1_amd64.deb
sudo apt-key add /var/cuda-repo-debian10-11-4-local/7fa2af80.pub
sudo add-apt-repository contrib
sudo apt-get update
sudo apt-get -y install cuda
To download Flash Attention here are the required steps:
pip install packaging
pip uninstall -y ninja && pip install ninja
MAX_JOBS=8 pip install flash-attn --no-build-isolation
An A100 will typically come with 83.6 GB of usable RAM.
ninjawill do parallel compilation jobs that could exhaust the amount of RAM. Set the max number of jobs usingMAX_JOBS. AMAX_JOBS=4will take 30+ minutes to compile flash attention, whileMAX_JOBS=8might take 20ish minutes (with 35ish GB of RAM usage). On an A100,MAX_JOBSof 16 might work (haven't tested).
We start by downloading the model and preparing the model so that it can be consumed by lit-gpt's finetuning pipeline.
python lit-gpt/scripts/download.py --repo_id meta-llama/Llama-2-7b-hf --token <HuggingFace Token>
python lit-gpt/scripts/convert_hf_checkpoint.py --checkpoint_dir checkpoints/meta-llama/Llama-2-7b-hf
Note: In order to use Llama 2 model you need to fill a small form in the HuggingFace model card page for this model. The permission is usually granted in 1-2 days. Tip: Please provide the same email id that you used to create your HuggingFace account.
Note: You can generate your HuggingFace access token here.
Download the dataset and prepare it using a convenient script provided by lit-gpt. Below I am downloading the databricks-dolly-15k dataset.
python lit-gpt/scripts/prepare_dolly.py --checkpoint_dir checkpoints/meta-llama/Llama-2-7b-hf
Note: The tokenizer used by the model checkpoint is used to tokenize the dataset. The dataset will be split into train and test set in the
.ptformat.
You are not only limited to the databricks-dolly-15k dataset. You can also download and prepare RedPajama-Data-1T dataset.
python lit-gpt/scripts/prepare_redpajama.py --checkpoint_dir checkpoints/meta-llama/Llama-2-7b-hf
Follow these steps to create your own data preparation script.
Tip: You will ideally want to combine datasets from varying benchmarks and sample them properly.
At this point, before going ahead, let's validate if our setup is working.
python lit-gpt/generate/base.py --checkpoint_dir checkpoints/meta-llama/Llama-2-7b-hf --prompt "Tell me an interesting fun fact about earth:"
lit-gpt provides a few resource constrainted finetuning strategies like lora, qlora, etc., out of the box.
- LoRA finetuning
python lit-gpt/finetune/lora.py --data_dir data/dolly/ --checkpoint_dir checkpoints/meta-llama/Llama-2-7b-hf --precision bf16-true --out_dir out/lora/llama-2-7b
- QLoRA finetuning
To finetune with QLoRA, you will have to install the bitsandbytes library.
pip install bitsandbytes
Finetune with QLoRA by passing the --quantize flag to the lora.py script
python lit-gpt/finetune/lora.py --data_dir data/dolly/meta-llama/ --checkpoint_dir checkpoints/meta-llama/Llama-2-7b-hf --precision bf16-true --out_dir out/lora/llama-2-7b --quantize "bnb.nf4"
Once you have a working setup for finetuing, coming up with finetuing strategies is going to be one of the most important task but this will be guided by an even bigger task - thinking through the evaluation strategy.
As per the organizers of this competition:
The evaluation process in our competition will be conducted in two stages. In the first stage, we will run a subset of HELM benchmark along with a set of secret holdout tasks. The holdout tasks will consist of logic reasoning type of multiple-choice Q&A scenarios as well as conversational chat tasks. Submissions will be ranked based on their performance across all tasks. The ranking will be determined by the geometric mean across all evaluation tasks.
We cannot do anything about the secret hold out tasks. But we can try to improve the finetuned model on a subset of the HELM benchmark.
We can also consider using other benchmarks like EluetherAI's lm-evaluation-harness. You can find the tasks available in this benchmark here. A few tasks here can be considered your validation set since there is an overlap between lm-evaluation-harness and HELM.
Install this library and perform evaluation with the base checkpoint and the LoRA finetuned checkpoint.
- Install
lm-evaluation-harness
cd ..
git clone https://github.com/EleutherAI/lm-evaluation-harness
cd lm-evaluation-harness
pip install -e .
cd neurips-llm-efficiency-challenge
- Evaluate using the base checkpoint
python lit-gpt/eval/lm_eval_harness.py --checkpoint_dir checkpoints/meta-llama/Llama-2-7b-hf --precision "bf16-true" --eval_tasks "[truthfulqa_mc, wikitext, openbookqa, arithmetic_1dc]" --batch_size 4 --save_filepath "results-falcon-7b.json"
- Evaluate the finetuned checkpoint
python lit-gpt/eval/lm_eval_harness_lora.py --lora_path out/lora/llama-2-7b/lit_model_lora_finetuned.pth --checkpoint_dir checkpoints/meta-llama/Llama-2-7b-hf --precision "bf16-true" --eval_tasks "[truthfulqa_mc, wikitext, openbookqa, arithmetic_1dc]" --batch_size 4 --save_filepath "results-falcon-7b.json"
Here
out/lora/<model-name>/lit_model_lora_finetuned.pthis something we get after finetuning the model.
Notice that for base model evaluation we used the
lm_eval_harness.pyscript while for finetuned model evaluation we are usinglm_eval_harness_lora.py. If you are using other strategy, you might have to modify the scipt to cater to your strategy.
The organizers of this challenge, have provided an useful toy example to demonstrate the submission steps. Check out the official repo here.
I have copied the files required to setup the submisison pipeline to this repo to simplify things.
There are a lot of details mentioned in the official documentation to install Nvidia container toolkit.
Here are the steps that worked for me:
Setup the GPG key (don't ignore this step)
distribution=$(. /etc/os-release;echo $ID$VERSION_ID) \
&& curl -s -L https://nvidia.github.io/libnvidia-container/$distribution/libnvidia-container.repo | sudo tee /etc/yum.repos.d/nvidia-container-toolkit.repo
Run the following commands:
sudo apt-get install -y nvidia-container-toolkit
sudo nvidia-ctk runtime configure --runtime=docker
sudo systemctl restart docker
docker build -t submission .
sudo docker run --rm --gpus all -p 8080:80 toy_submission
If everything works out, you will have a successful running HTTP server. The steps mentioned in the Dockerfile should be self satisfactory.
A subset (unknown) will be used for the leaderboard evaluation. We can choose to use the official HELM repository but it is not configured to hit the endpoints of the HTTP server. We will instead be using this fork (neruips_client branch - note the typo) of the HELM repo.
cd ..
git clone https://github.com/drisspg/helm.git@neruips_client
Create a new environment because this repo uses different versions of multiple repositories.
conda create -n helm-eval python==3.10.0
conda activate helm-eval
cd helm
pip install -e .
Note: If 8080 is already in use, do the following: change the base url that's hardcoded from http://localhost:8080 (to avoid conflict with Jupyter Notebook, most probable) to http://localhost:9000 in line 30 of this file - helm/src/helm/proxy/clients/http_model_client.py.
Run the following lines to benchmark on the mmlu (subset of HELM) benchmark:
cd ..
cd neurips-llm-efficiency-challenge
echo 'entries: [{description: "mmlu:model=neurips/local,subject=college_computer_science", priority: 4}]' > run_specs.conf
helm-run --conf-paths run_specs.conf --suite v1 --max-eval-instances 1000
helm-summarize --suite v1
Check out the various benchmarks that are present in this benchmark here
You might encounter this issue - OSError: /opt/conda/envs/helm-eval/lib/python3.10/site-packages/nvidia/cublas/lib/libcublas.so.11: symbol cublasLtGetStatusString version libcublasLt.so.11 not defined in file libcublasLt.so.11 with link time reference. If so do the following steps:
To solve this check out this Stack OverFlow answer.
After successful evaluation, you can find the results in the benchmark_output directory. I am working on logging the results to Weights and Biases for easy comparison and tracking.
I hope the documented steps will expedite the setting up process so that more time can be spent on doing ML.
The steps were tested on a GCP Compute Engine VM with A100 (40 GB) GPU. If you have a different setup and if something doesn't work, feel free to open an issue or raise a PR.