File size: 1,608 Bytes
09dc0f0
 
dc5d361
09dc0f0
 
 
 
 
 
 
dbc7df4
 
09dc0f0
 
 
 
 
 
dbc7df4
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
<p><strong><font size="5">Information</font></strong></p>
GPT4-X-Alpaca 30B 4-bit working with GPTQ versions used in Oobabooga's Text Generation Webui and KoboldAI.
<p>There are 3 quantized versions, one is quantized using GPTQ's <i>--true-sequential</i> and <i>--act-order</i> optimizations, the second is quantized using GPTQ's <i>--true-sequential</i> and <i>--groupsize 128</i> optimization, and the third one is quantized for GGML using q4_1</p>
This was made using Chansung's GPT4-Alpaca Lora: https://huggingface.co/chansung/gpt4-alpaca-lora-30b

<p><strong>Training Parameters</strong></p>
<ul><li>num_epochs=10</li><li>cutoff_len=512</li><li>group_by_length</li><li>lora_target_modules='[q_proj,k_proj,v_proj,o_proj]'</li><li>lora_r=16</li><li>micro_batch_size=8</li></ul>

<p><strong><font size="5">Benchmarks</font></strong></p>

<p><strong><font size="4">--true-sequential --act-order</font></strong></p>

<strong>Wikitext2</strong>: 4.481280326843262

<strong>Ptb-New</strong>: 8.539161682128906

<strong>C4-New</strong>: 6.451964855194092

<strong>Note</strong>: This version does not use <i>--groupsize 128</i>, therefore evaluations are minimally higher. However, this version allows fitting the whole model at full context using only 24GB VRAM.

<p><strong><font size="4">--true-sequential --groupsize 128</font></strong></p>

<strong>Wikitext2</strong>: 4.285132884979248

<strong>Ptb-New</strong>: 8.34856128692627

<strong>C4-New</strong>: 6.292652130126953

<strong>Note</strong>: This version uses <i>--groupsize 128</i>, resulting in better evaluations. However, it consumes more VRAM.