This project is the JAX implementation of Llama 2. The objectives of this project are threefold:
- Implement the Llama 2 model using JAX to enable efficient training and inference on Google Cloud TPU;
- Develop a high-quality codebase that serves as an exemplary implementation of the Transformer model using JAX;
- Facilitate the identification of common errors and inconsistencies across various transformer models through the implementation of a high-quality codebase, thereby providing valuable insights for the NLP community.
Related projects:
- hyunwoongko/transformer: PyTorch implementation of the original Transformer
- ayaka14732/TrAVis: An in-browser Transformer attention visualiser that includes a NumPy implementation of BERT
- ayaka14732/bart-base-jax: JAX implementation of BART-base
- ztjhz/t5-jax: JAX implementation of T5
- young-geng/EasyLM: LLM framework that includes Flax implementations of LLaMA, GPT-J and RoBERTa
This project is supported by Cloud TPUs from Google's TPU Research Cloud (TRC).
- Parameter conversion
- Data loading
- Model architecture
- Cross entropy loss
- Training
- Data parallelism
- Model parallelism
- Other parallelisation schemes
- Generation
- KV cache
- Left padding
- Presence penalty
- Frequency penalty
- Beam search
- Beam sampling
- Top-k sampling
- Top-p sampling
- Documentation
The documentation of the library of this project is published on GitHub Pages.
This project requires at least Python 3.11, JAX 0.4.19, PyTorch 2.1.0, Optax 0.1.8.dev0 and Transformers 4.35.0.dev0.
PyTorch and Transformers are needed for testing purposes. Additionally, the data loader depends on PyTorch DataLoader
, while the profiling functionality requires TensorFlow.
For Ubuntu users, you can follow How to install Python 3.11 on Ubuntu 22.04 to Install Python 3.11. The tutorial applied to Ubuntu 20.04 as well.
python3.11 -m venv venv
. venv/bin/activate
pip install -U pip
pip install -U wheel
If you are running on TPU pods, you need to put the IP address of all other hosts in ~/podips.txt
(one IP address per line). Besides, you should make sure that the local host can SSH into itself and all other hosts listed in the file.
You need to follow the installation instructions on JAX's official GitHub page.
Typically, you only need to install the CPU version of PyTorch since we perform most of the computation using JAX. However, it's worth noting that the current codebase's generation process is not fully optimised yet. To expedite the inference, one effective approach would involve converting the model back to Hugging Face format and running the inference in PyTorch.
To install PyTorch, you can follow the official installation guide.
On TPU VMs, this is usually:
pip install --pre torch --index-url https://download.pytorch.org/whl/nightly/cpu
On TPU Pods:
./podrun -i -- ~/venv/bin/pip install --pre torch --index-url https://download.pytorch.org/whl/nightly/cpu
pip install git+https://github.com/huggingface/transformers.git
pip install git+https://github.com/deepmind/optax.git # https://github.com/google-deepmind/optax/issues/472
pip install -r requirements.txt
On TPU Pods:
./podrun -i -- ~/venv/bin/pip install git+https://github.com/huggingface/transformers.git
./podrun -i -- ~/venv/bin/pip install git+https://github.com/deepmind/optax.git
./podrun -iw -- ~/venv/bin/pip install -r requirements.txt
LLaMA 1:
If you couldn't obtain the LLaMA weights, you can download them with shawwn/llama-dl.
mkdir ../llama-weights-original && cd ../llama-weights-original
curl -o- https://github.com/raw/shawwn/llama-dl/56f50b96072f42fb2520b1ad5a1d6ef30351f23c/llama.sh | bash
python ../llama-2-jax/venv/lib/python3.11/site-packages/transformers/models/llama/convert_llama_weights_to_hf.py --input_dir ../llama-weights-original --model_size 7B --output_dir ../llama-weights/7B
Llama 2:
You can request to access the Llama weights from the official website. After your request is approved, you will automatically get access to the Hugging Face Llama 2 models. You can verify that the models are accessible by trying to access the Llama 2 7B version.
If you need to work with Llama 2 models, you need to login into Hugging Face CLI:
huggingface-cli login
Alternatively, in case you are not using an interactive shell, you can login in non-interactive mode:
python -c "from huggingface_hub.hf_api import HfFolder; HfFolder.save_token('<YOUR_HUGGING_FACE_TOKEN>')"
On TPU Pods:
./podrun -i -- ~/venv/bin/python -c "from huggingface_hub.hf_api import HfFolder; HfFolder.save_token('<YOUR_HUGGING_FACE_TOKEN>')"
python scripts/convert_params_runner.py llama1-7B
python scripts/convert_params_runner.py llama2-7B
python scripts/convert_params_runner.py llama2-70B
python generate.py
On TPU pods, the command is:
./podrun -icw ~/venv/bin/python generate.py
Before training, you need to login into W&B:
wandb login <YOUR_WANDB_API_KEY>
I present a simple example of the training pipeline by fine-tuning the model on the GSM dataset.
cd .. && git clone --depth=1 https://github.com/openai/grade-school-math.git
python train.py
On TPU pods, the command is:
./podrun -icw ~/venv/bin/python train.py
- B: batch_size
- L: seq_len
- S: src_seq_len
- D: dst_seq_len
- C: vocab_size
- N: n_layers
- K: d_k
- V: d_v
- H: n_heads_kv
- R: n_rep_kv
- M: d_model
- F: d_ff
Name | Parameters | C | N | K/V | H | R | M | F |
---|---|---|---|---|---|---|---|---|
LLaMA 1 7B | 6738415616 | 32000 | 32 | 128 | 32 | 1 | 4096 | 11008 |
Llama 2 7B | 6738415616 | 32000 | 32 | 128 | 32 | 1 | 4096 | 11008 |
LLaMA 1 13B | 32000 | 40 | 40 | 1 | 5120 | |||
Llama 2 13B | 13015864320 | 32000 | 40 | 128 | 40 | 1 | 5120 | 13824 |
LLaMA 1 33B | 32000 | 60 | 52 | 1 | 6656 | |||
LLaMA 1 65B | 32000 | 80 | 64 | 1 | 8192 | |||
Llama 2 70B | 68976648192 | 32000 | 80 | 128 | 8 | 8 | 8192 | 28672 |
n_params = 2CM + (2N + 1)M + 2NMRHK + 2NMHK + 3NMF
Hugging Face format:
LlamaForCausalLM(
(model): LlamaModel(
(embed_tokens): Embedding(32000, 4096, padding_idx=0)
(layers): ModuleList(
(0-31): 32 x LlamaDecoderLayer(
(self_attn): LlamaAttention(
(q_proj): Linear(in_features=4096, out_features=4096, bias=False)
(k_proj): Linear(in_features=4096, out_features=4096, bias=False)
(v_proj): Linear(in_features=4096, out_features=4096, bias=False)
(o_proj): Linear(in_features=4096, out_features=4096, bias=False)
(rotary_emb): LlamaRotaryEmbedding()
)
(mlp): LlamaMLP(
(gate_proj): Linear(in_features=4096, out_features=11008, bias=False)
(down_proj): Linear(in_features=11008, out_features=4096, bias=False)
(up_proj): Linear(in_features=4096, out_features=11008, bias=False)
(act_fn): SiLUActivation()
)
(input_layernorm): LlamaRMSNorm()
(post_attention_layernorm): LlamaRMSNorm()
)
)
(norm): LlamaRMSNorm()
)
(lm_head): Linear(in_features=4096, out_features=32000, bias=False)
)
The format used in this project:
model
embedding: (32000, 4096)
decoder: decoder_block
input_norm: (32, 4096)
attention
q_proj: (32, 4096, 1, 32, 128)
k_proj: (32, 4096, 32, 128)
v_proj: (32, 4096, 32, 128)
out_proj: (32, 1, 32, 128, 4096)
post_attn_norm: (32, 4096)
gate_proj: (32, 4096, 11008)
up_proj: (32, 4096, 11008)
down_proj: (32, 11008, 4096)
norm: (4096)
lm_head: (4096, 32000)
Hugging Face format:
LlamaForCausalLM(
(model): LlamaModel(
(embed_tokens): Embedding(32000, 8192, padding_idx=0)
(layers): ModuleList(
(0-79): 80 x LlamaDecoderLayer(
(self_attn): LlamaAttention(
(q_proj): Linear(in_features=8192, out_features=8192, bias=False)
(k_proj): Linear(in_features=8192, out_features=1024, bias=False)
(v_proj): Linear(in_features=8192, out_features=1024, bias=False)
(o_proj): Linear(in_features=8192, out_features=8192, bias=False)
(rotary_emb): LlamaRotaryEmbedding()
)
(mlp): LlamaMLP(
(gate_proj): Linear(in_features=8192, out_features=28672, bias=False)
(up_proj): Linear(in_features=8192, out_features=28672, bias=False)
(down_proj): Linear(in_features=28672, out_features=8192, bias=False)
(act_fn): SiLUActivation()
)
(input_layernorm): LlamaRMSNorm()
(post_attention_layernorm): LlamaRMSNorm()
)
)
(norm): LlamaRMSNorm()
)
(lm_head): Linear(in_features=8192, out_features=32000, bias=False)
)
The format used in this project:
model
embedding: (32000, 8192)
decoder: decoder_block
input_norm: (80, 8192)
attention
q_proj: (80, 8192, 8, 8, 128)
k_proj: (80, 8192, 8, 128)
v_proj: (80, 8192, 8, 128)
out_proj: (80, 8, 8, 128, 8192)
post_attn_norm: (80, 8192)
gate_proj: (80, 8192, 28672)
up_proj: (80, 8192, 28672)
down_proj: (80, 28672, 8192)
norm: (8192)
lm_head: (8192, 32000)
- LLaMA utilises rotary positional embeddings.
- There is no bias in the Q, K, V matrices and the linear projections in the FFNs, which is the same as the original transformer, but different from BERT and BART.
- In Llama models, each FFN has 3 linear projections, while in BART there are only 2.
- There is no dropout in the original LLaMA implementation.
- No left padding in the original implementation, see huggingface/transformers#26072.
- Llama 2 70B utilises Grouped-Query Attention (GQA).
- Many people fine-tunes Llama in a 16-bit precision (float16 or bfloat16), but the performance would be impacted and thus comparisons with other models trained in 32-bit precision would be unfair. Another thing that is worth noticing is that the parameters for rotary embedding should be always in 32-bit to avoid collision.