Update README.md
Browse files
README.md
CHANGED
|
@@ -6,16 +6,22 @@ tags:
|
|
| 6 |
---
|
| 7 |
|
| 8 |
|
|
|
|
|
|
|
| 9 |
Q. Why quantized in 8bit instead of 4bit?
|
| 10 |
A. For evaluation purpose. In theory, a 8bit quantized model should provide slightly better perplexity (maybe not noticeable - To Be Evaluated...) over a 4bit quatized version. If your available GPU VRAM is over 15GB you may want to try this out.
|
| 11 |
Note that quatization in 8bit does not mean loading the model in 8bit precision. Loading your model in 8bit precision (--load-in-8bit) comes with noticeable quality (perplexity) degradation.
|
|
|
|
| 12 |
|
| 13 |
Refs:
|
| 14 |
- https://github.com/ggerganov/llama.cpp/pull/951
|
| 15 |
- https://news.ycombinator.com/item?id=35148542
|
|
|
|
|
|
|
| 16 |
- https://arxiv.org/abs/2105.03536
|
| 17 |
-
- https://github.com/IST-DASLab/gptq
|
| 18 |
- https://arxiv.org/abs/2212.09720
|
|
|
|
|
|
|
| 19 |
|
| 20 |
<br>
|
| 21 |
|
|
@@ -25,7 +31,7 @@ Refs:
|
|
| 25 |
- wbits: 8
|
| 26 |
- true-sequential: yes
|
| 27 |
- act-order: yes
|
| 28 |
-
- 8-bit
|
| 29 |
- Conversion process: LLaMa 13B -> LLaMa 13B HF -> Vicuna13B-v1.1 HF -> Vicuna13B-v1.1-8bit-128g
|
| 30 |
|
| 31 |
<br>
|
|
@@ -84,7 +90,9 @@ pip install -r requirements.txt
|
|
| 84 |
- i9-7980XE OC @4.6Ghz
|
| 85 |
|
| 86 |
- 11 tokens/s on average with Triton
|
| 87 |
-
-
|
|
|
|
|
|
|
| 88 |
- Tested and working in both chat mode and text generation mode
|
| 89 |
|
| 90 |
|
|
|
|
| 6 |
---
|
| 7 |
|
| 8 |
|
| 9 |
+
This is a 8bit GPTQ (not to be confused with 8bit RTN) version of Vicuna 13B v1.1 HF.
|
| 10 |
+
|
| 11 |
Q. Why quantized in 8bit instead of 4bit?
|
| 12 |
A. For evaluation purpose. In theory, a 8bit quantized model should provide slightly better perplexity (maybe not noticeable - To Be Evaluated...) over a 4bit quatized version. If your available GPU VRAM is over 15GB you may want to try this out.
|
| 13 |
Note that quatization in 8bit does not mean loading the model in 8bit precision. Loading your model in 8bit precision (--load-in-8bit) comes with noticeable quality (perplexity) degradation.
|
| 14 |
+
This model is also only useful until Vicuna30B or higher come to light, in which case a 8bit GPTQ version for these models would not fit consumer cards and might be less than a 4bit GPTQ (To Be Evaluated).
|
| 15 |
|
| 16 |
Refs:
|
| 17 |
- https://github.com/ggerganov/llama.cpp/pull/951
|
| 18 |
- https://news.ycombinator.com/item?id=35148542
|
| 19 |
+
- https://github.com/ggerganov/llama.cpp/issues/53
|
| 20 |
+
- https://arxiv.org/abs/2210.17323
|
| 21 |
- https://arxiv.org/abs/2105.03536
|
|
|
|
| 22 |
- https://arxiv.org/abs/2212.09720
|
| 23 |
+
- https://arxiv.org/abs/2301.00774
|
| 24 |
+
- https://github.com/IST-DASLab/gptq
|
| 25 |
|
| 26 |
<br>
|
| 27 |
|
|
|
|
| 31 |
- wbits: 8
|
| 32 |
- true-sequential: yes
|
| 33 |
- act-order: yes
|
| 34 |
+
- 8-bit GPTQ
|
| 35 |
- Conversion process: LLaMa 13B -> LLaMa 13B HF -> Vicuna13B-v1.1 HF -> Vicuna13B-v1.1-8bit-128g
|
| 36 |
|
| 37 |
<br>
|
|
|
|
| 90 |
- i9-7980XE OC @4.6Ghz
|
| 91 |
|
| 92 |
- 11 tokens/s on average with Triton
|
| 93 |
+
- Equivalent tokens/s observed over the 4bit version
|
| 94 |
+
- Pending preliminary observation: better quality results than 8bit RTN (--load-in-8bits) (To Be Confirmed)
|
| 95 |
+
- Pending preliminary observation: slightly better quality results than 4bit GPTQ (To Be Confirmed)
|
| 96 |
- Tested and working in both chat mode and text generation mode
|
| 97 |
|
| 98 |
|