0️⃣1️⃣🤗 BitNet-Transformers: Huggingface Transformers Implementation of "BitNet: Scaling 1-bit Transformers for Large Language Models" in pytorch with Mistral Architecture

• Paper Link: https://arxiv.org/pdf/2310.11453.pdf

Prepare Dev env

# Clone this repo
git clone https://github.com/DewEfresh/bitnet-transformers
cd bitnet-transformers

# Install requirements
pip install -r clm_requirements.txt

# Clone transformers repo
git clone https://github.com/huggingface/transformers
pip install -e transformers

# Update Llama(2) model
rm ./transformers/src/transformers/models/llama/modeling_llama.py
ln -s $(pwd)/bitnet_mistral/modeling_llama.py ./transformers/src/transformers/models/llama/modeling_llama.py

# Update Llama(2) model
rm ./transformers/src/transformers/models/mistral/modeling_mistral.py
ln -s $(pwd)/bitnet_mistral/modeling_mistral.py ./transformers/src/transformers/models/mistral/modeling_mistral.py

We'll overwrite bitnet_llama/modeling_llama.py into transformers. Since the file is linked, any changes made to the file will be reflected in the transformers repo.

Train Wikitext-103

You can track metrics via wandb

./train_wikitext.sh

GPU Mem Usage Comparison

Train Config

• Batch size: 1
• Gradient accumulation: 1
• Seq length: 2048
• Model: LLamaForCausalLM with BitLinear layer
• Model size: 47,452,672 (47.5M)

Original LLAMA - 16bit

• Uses 250MB GPU memory for Model weights

BitLLAMA - Mixed 16bit

• Uses 200MB GPU memory for Model weights
• Use bf16(or fp16) to store model weights
• Use int8 to store -1/1 1-bit weights
• Use more memory when training than original LLAMA: It saves 1-bit weight and 16bit weight together

BitLLAMA - 8bit

• Uses 100MB GPU memory for Model weights
• Use bf16(or fp16) on-the-fly when needed
• Use 8bit to save 1-bit BitLinear weight & other weights

BitLLAMA - 1bit

• Use bf16(or fp16) on-the-fly when needed
• Use 1bit to save 1-bit weight

TBD

Todo

• Add BitLinear layer
• Add LLamaForCausalLM model with BitLinear layer
  • Update .save_pretrained method (for 1-bit weight saving)
• Add sample code for LM training
• Update BitLinear layer to use 1-bit weight
  • Use uint8 instead of bfloat16
  • Use custom cuda kernel for 1-bit weight