Create readme
Browse files
readme
ADDED
|
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
3bit quantized version of this: https://huggingface.co/ausboss/llama-30b-supercot
|
| 2 |
+
|
| 3 |
+
GPTQ quantization using https://github.com/0cc4m/GPTQ-for-LLaMa
|
| 4 |
+
|
| 5 |
+
Made at the request of someone that wanted a 3bit version. The file is 17% smaller than 4bit non-groupsize, but the wikitext2 ppl is 12% worse. I don't have a functioning Ooba install so I can't test this myself.
|
| 6 |
+
|
| 7 |
+
Command used to quantize:
|
| 8 |
+
```python llama.py c:\llama-30b-supercot c4 --wbits 3 --true-sequential --groupsize 128 --save_safetensors 4bit-128g.safetensors```
|
| 9 |
+
|
| 10 |
+
Evaluation & Score (Lower is better):
|
| 11 |
+
* WikiText2: 5.22 (12% worse than 4bit non-groupsize)
|
| 12 |
+
* PTB: 19.63 (11% worse than 4bit non-groupsize)
|
| 13 |
+
* C4: 6.93 (7% worse than 4bit non-groupsize)
|
| 14 |
+
|
| 15 |
+
4bit non-groupsize version is here: https://huggingface.co/tsumeone/llama-30b-supercot-4bit-cuda
|
| 16 |
+
4bit 128 groupsize version is here: https://huggingface.co/tsumeone/llama-30b-supercot-4bit-128g-cuda
|