Bitnet-LLama-70M

Bitnet-LLama-70M is a 70M parameter model trained using the method described in The Era of 1-bit LLMs: All Large Language Models are in 1.58 Bits.

It was trained on the subset of the HuggingFaceTB/cosmopedia dataset. This is just a small experiment to try out BitNet. Bitnet-LLama-70M was trained for 2 epochs on 1xA100.

This model is just an experiment and you might not get good results while chatting with it due to smaller model size and less training.

Wandb training report is as follows:

Sample inference code

from transformers import AutoModelForCausalLM, AutoTokenizer

# Load a pretrained BitNet model
model = "abideen/Bitnet-Llama-70M"
tokenizer = AutoTokenizer.from_pretrained(model)
model = AutoModelForCausalLM.from_pretrained(model)

def convert_to_bitnet(model, copy_weights):
    for name, module in model.named_modules():
        # Replace linear layers with BitNet
        if isinstance(module, LlamaSdpaAttention) or isinstance(module, LlamaMLP):
            for child_name, child_module in module.named_children():
                if isinstance(child_module, nn.Linear):
                    bitlinear = BitLinear(child_module.in_features, child_module.out_features, child_module.bias is not None).to(device="cuda:0")
                    if copy_weights:
                        bitlinear.weight = child_module.weight
                        if child_module.bias is not None:
                            bitlinear.bias = child_module.bias
                    setattr(module, child_name, bitlinear)
        # Remove redundant input_layernorms
        elif isinstance(module, LlamaDecoderLayer):
            for child_name, child_module in module.named_children():
                if isinstance(child_module, LlamaRMSNorm) and child_name == "input_layernorm":
                    setattr(module, child_name, nn.Identity().to(device="cuda:0"))
              

convert_to_bitnet(model, copy_weights=True)
model.to(device="cuda:0")

prompt = "What is Machine Learning?"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
generate_ids = model.generate(inputs.input_ids, max_length=100)
tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]