docs: v3.13.0 card update
Browse files
README.md
CHANGED
|
@@ -1 +1,89 @@
|
|
| 1 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
tags:
|
| 4 |
+
- bigsmall
|
| 5 |
+
- compressed
|
| 6 |
+
- lossless
|
| 7 |
+
---
|
| 8 |
+
|
| 9 |
+
[](https://doi.org/10.5281/zenodo.20279248)
|
| 10 |
+
|
| 11 |
+
# GPT-2 — BigSmall Compressed
|
| 12 |
+
|
| 13 |
+
**75.5% smaller than [gpt2](https://huggingface.co/gpt2). Bit-identical weights. Drop-in replacement.**
|
| 14 |
+
|
| 15 |
+
## Use it in 2 lines
|
| 16 |
+
|
| 17 |
+
```bash
|
| 18 |
+
pip install bigsmall
|
| 19 |
+
```
|
| 20 |
+
|
| 21 |
+
```python
|
| 22 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 23 |
+
|
| 24 |
+
model = AutoModelForCausalLM.from_pretrained("wpferrell/gpt2-bigsmall")
|
| 25 |
+
tokenizer = AutoTokenizer.from_pretrained("gpt2")
|
| 26 |
+
```
|
| 27 |
+
|
| 28 |
+
## Size comparison
|
| 29 |
+
|
| 30 |
+
| | Size |
|
| 31 |
+
|---|---|
|
| 32 |
+
| Original ([gpt2](https://huggingface.co/gpt2)) | 548 MB |
|
| 33 |
+
| This compressed version | 414 MB |
|
| 34 |
+
| Compression ratio | 75.5% of original |
|
| 35 |
+
|
| 36 |
+
## What "lossless" means
|
| 37 |
+
|
| 38 |
+
Every weight is mathematically identical to the original. Not quantized. Not pruned. Identical model behaviour. Compressed using entropy coding — like ZIP but designed for neural weights.
|
| 39 |
+
|
| 40 |
+
Every tensor's md5 is stored in the file header; `bigsmall verify` decompresses and verifies bit-for-bit.
|
| 41 |
+
|
| 42 |
+
## Low VRAM streaming
|
| 43 |
+
|
| 44 |
+
```python
|
| 45 |
+
from bigsmall import BigSmallStreamingModel
|
| 46 |
+
|
| 47 |
+
model = BigSmallStreamingModel.from_pretrained(
|
| 48 |
+
"wpferrell/gpt2-bigsmall",
|
| 49 |
+
device="cuda",
|
| 50 |
+
lru_max_vram_gb=2.0, # cache 2 GB of decoded layers (v3.13.0+)
|
| 51 |
+
)
|
| 52 |
+
out = model.generate(input_ids, max_new_tokens=100)
|
| 53 |
+
```
|
| 54 |
+
|
| 55 |
+
Streams layers on demand — uses ~12× less VRAM than a full load.
|
| 56 |
+
|
| 57 |
+
## New in BigSmall v3.13.0
|
| 58 |
+
|
| 59 |
+
- **Delta compression** — if you have the base model already, store a fine-tune as a tiny patch on top:
|
| 60 |
+
```bash
|
| 61 |
+
bigsmall compress my_finetune.safetensors --delta-from base.safetensors -o patch.bs
|
| 62 |
+
```
|
| 63 |
+
- **`bigsmall scan`** — analyse a model before compressing.
|
| 64 |
+
- **`bigsmall apply`** — reconstruct a fine-tune from base + patch.
|
| 65 |
+
- **`bigsmall verify --sample 0.001`** — fast probabilistic integrity check.
|
| 66 |
+
- **`bigsmall compress --ecc`** — optional Reed-Solomon error recovery.
|
| 67 |
+
- **`bigsmall compress --resume`** — resumable compression.
|
| 68 |
+
|
| 69 |
+
[Full release notes →](https://github.com/wpferrell/Bigsmall/releases/tag/v3.13.0)
|
| 70 |
+
|
| 71 |
+
## All pre-compressed models
|
| 72 |
+
|
| 73 |
+
See [wpferrell on HuggingFace](https://huggingface.co/wpferrell) for all available models.
|
| 74 |
+
|
| 75 |
+
## License
|
| 76 |
+
|
| 77 |
+
Model weights: same license as the original model ([gpt2](https://huggingface.co/gpt2)).
|
| 78 |
+
BigSmall format: [Elastic License 2.0](https://www.elastic.co/licensing/elastic-license) — free for personal, research, and internal commercial use.
|
| 79 |
+
Commercial/SaaS licensing: wpferrell@gmail.com
|
| 80 |
+
|
| 81 |
+
## Requires
|
| 82 |
+
|
| 83 |
+
`bigsmall >= 3.13.0` to access the latest features. Older versions (>= 3.0.0) can still decode this model.
|
| 84 |
+
|
| 85 |
+
## Citation
|
| 86 |
+
|
| 87 |
+
If you use BigSmall in research:
|
| 88 |
+
|
| 89 |
+
> Ferrell, W. (2026). *BigSmall: Lossless Compression for Neural Network Weights.* Zenodo. https://doi.org/10.5281/zenodo.20279248
|