Guanaco-33B-4bit / README.md
MetaIX's picture
Update README.md
3c7dd84
|
raw
history blame
2.39 kB
<p><strong><font size="5">Information</font></strong></p>
Guanaco 33b working with Oobabooga's Text Generation Webui and KoboldAI.
<p>This is a quantized version of <a href="https://huggingface.co/timdettmers/guanaco-33b">Tim Dettmers' Guanaco 33b.</a></p>
<p><strong>What's included</strong></p>
<P>GPTQ: 2 quantized versions. One quantized --true-sequential and act-order optimizations, and the other was quantized using --true-sequential --groupsize 128 optimizations.</P>
<P>GGML: 3 quantized versions. One quantized using q4_1, another was quantized using q5_0, and the last one was quantized using q5_1.</P>
<p><strong>GPU/GPTQ Usage</strong></p>
<p>To use with your GPU using GPTQ pick one of the .safetensors along with all of the .jsons and .model files.</p>
<p>Oobabooga: If you require further instruction, see <a href="https://github.com/oobabooga/text-generation-webui/blob/main/docs/GPTQ-models-(4-bit-mode).md">here</a> and <a href="https://github.com/oobabooga/text-generation-webui/blob/main/docs/LLaMA-model.md">here</a></p>
<p>KoboldAI: If you require further instruction, see <a href="https://github.com/0cc4m/KoboldAI">here</a></p>
<p><strong>CPU/GGML Usage</strong></p>
<p>To use your CPU using GGML(Llamacpp) you only need the single .bin ggml file.</p>
<p>Oobabooga: If you require further instruction, see <a href="https://github.com/oobabooga/text-generation-webui/blob/main/docs/llama.cpp-models.md">here</a></p>
<p>KoboldAI: If you require further instruction, see <a href="https://github.com/LostRuins/koboldcpp">here</a></p>
<p><strong><font size="5">Benchmarks</font></strong></p>
<p><strong><font size="4">--true-sequential --act-order</font></strong></p>
<strong>Wikitext2</strong>: 4.582493305206299
<strong>Ptb-New</strong>: 8.697775840759277
<strong>C4-New</strong>: 6.67733097076416
<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>: 4.369843006134033
<strong>Ptb-New</strong>: 8.53034496307373
<strong>C4-New</strong>: 6.496636390686035
<strong>Note</strong>: This version uses <i>--groupsize 128</i>, resulting in better evaluations. However, it consumes more VRAM.