| --- |
| license: apache-2.0 |
| tags: |
| - bigsmall |
| - compression |
| - lossless |
| - qwen2 |
| --- |
| |
| # Qwen 2.5 7B Instruct (BigSmall compressed) |
|
|
| **15.2 GB → 10.1 GB (66.0%). Under 2 GB peak RAM. Full quality — not quantization.** |
|
|
| This is Qwen2.5-7B-Instruct compressed with [BigSmall](https://github.com/wpferrell/Bigsmall) — lossless neural network weight compression. Every weight is bit-identical to the original. No accuracy loss whatsoever. |
|
|
| ## Install |
|
|
| ```bash |
| pip install bigsmall |
| ``` |
|
|
| ## Load and run inference (streaming — under 2GB peak RAM) |
|
|
| ```python |
| from bigsmall import StreamingLoader |
| from transformers import AutoModelForCausalLM, AutoTokenizer |
| |
| loader = StreamingLoader("wpferrell/qwen2.5-7b-instruct-bigsmall") |
| model = loader.load_model(AutoModelForCausalLM) |
| tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B-Instruct") |
| |
| messages = [{"role": "user", "content": "Explain lossless compression."}] |
| text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True) |
| inputs = tokenizer([text], return_tensors="pt") |
| outputs = model.generate(**inputs, max_new_tokens=200) |
| print(tokenizer.decode(outputs[0])) |
| ``` |
|
|
| ## Or use AutoModel with the transparent hook |
|
|
| ```python |
| import bigsmall |
| bigsmall.install_hook() |
| |
| from transformers import AutoModelForCausalLM |
| model = AutoModelForCausalLM.from_pretrained("wpferrell/qwen2.5-7b-instruct-bigsmall") |
| ``` |
|
|
| ## Compression stats |
|
|
| | Metric | Value | |
| |--------|-------| |
| | Original size | 15.2 GB | |
| | Compressed size | 10.1 GB | |
| | Ratio | 66.0% (BF16) | |
| | Format | BF16 → BigSmall (.bs shards) | |
| | Lossless verified | md5 every tensor | |
| | Peak RAM (streaming) | < 2 GB | |
|
|
| ## Comparison |
|
|
| | Tool | BF16 Ratio | FP32 Ratio | Inference Overhead | Hardware | |
| |------|------------|------------|-------------------|---------| |
| | [ZipNN](https://arxiv.org/abs/2411.05239) | 67% | 83% | None | CPU | |
| | [DFloat11](https://arxiv.org/abs/2504.11651) | ~70% | BF16 only | ~2x at batch=1 | CUDA only | |
| | [ZipServ](https://arxiv.org/abs/2603.17435) | ~70% | BF16 only | 1.22x faster | GDDR GPU | |
| | **BigSmall** | **65.6%** | **75.5%** | **None** | **CPU + any GPU** | |
|
|
| ## 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](https://github.com/wpferrell/Bigsmall) |
| - PyPI: `pip install bigsmall` |
| - Paper: [BigSmall: Lossless Neural Network Weight Compression at the Joint Entropy Floor](https://github.com/wpferrell/Bigsmall/blob/main/paper.pdf) |
|
|
