wpferrell/gpt2-bigsmall

GPT-2 (BigSmall compressed)

548MB → 414MB (75.5%). Bit-identical. Under 500MB peak RAM with streaming.

This is GPT-2 117M compressed with BigSmall — lossless neural network weight compression. Not quantization. Every weight is bit-identical to the original.

Install

pip install bigsmall

Load and run inference (streaming)

from bigsmall import StreamingLoader
from transformers import GPT2LMHeadModel, GPT2Tokenizer

# Streams one layer at a time — under 500MB peak RAM
loader = StreamingLoader("wpferrell/gpt2-bigsmall")
model = loader.load_model(GPT2LMHeadModel)
tokenizer = GPT2Tokenizer.from_pretrained("openai-community/gpt2")

inputs = tokenizer("Hello, I'm a language model", return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=50)
print(tokenizer.decode(outputs[0]))

Or decompress to disk first

from bigsmall import from_pretrained
from transformers import GPT2LMHeadModel
model = from_pretrained("wpferrell/gpt2-bigsmall", model_class=GPT2LMHeadModel)

What's inside

File	Original	Compressed	Ratio
model.safetensors (FP32)	548 MB	414 MB	75.5%

Verified lossless: md5 of every weight tensor matches original after decompression.

Comparison

Tool	BF16 Ratio	FP32 Ratio	Inference Overhead	Hardware
ZipNN	67%	83%	None	CPU
DFloat11	~70%	BF16 only	~2x at batch=1	CUDA only
ZipServ	~70%	BF16 only	1.22x faster	GDDR GPU
BigSmall	65.6%	75.5%	None	CPU + any GPU

Lower ratio = better compression. BigSmall BF16 measured on Mistral 7B.

About BigSmall

BigSmall compresses at the joint entropy floor for neural network weights. It codes sign+exponent jointly and mantissa conditioned on exponent, achieving the information-theoretic minimum. The streaming loader decompresses one transformer layer at a time directly into VRAM.

• GitHub: wpferrell/Bigsmall
• PyPI: pip install bigsmall