license: apache-2.0
inference: true
tags:
- vicuna
This model is a 8bit quantization of Vicuna 13B.
- 13B parameters
- Group size: 128
- wbits: 8
- true-sequential: yes
- act-order: yes
- 8-bit quantized - Read more about this here: https://github.com/ggerganov/llama.cpp/pull/951
- Conversion process: Llama 13B -> Llama 13B HF -> Vicuna13B-v1.1 HF -> Vicuna13B-v1.1-8bit-128g
Basic installation procedure
- It was a nightmare, I will only detail briefly what you'll need. WSL was quite painful to sort out.
- I will not provide installation support, sorry.
- You can certainly use llama.cpp and other loaders that support 8bit quantization, I just chose oobabooga/text-generation-webui.
- You will likely face many bugs until text-generation-webui loads, ranging between missing PATH or env variables to having to manually pip uninstall/install packages.
- The notes below will likely become outdated once both text-generation-webui and GPTQ-for-LLaMa receive the appropriate bug fixes.
- If this model produces very slow answers (1 token/s), it means you are not using Cuda for bitsandbytes or that your hardware needs an upgrade.
- If this model produces answers with weird characters, it means you are not using the correct version of qwopqwop200/GPTQ-for-LLaMa as mentioned below.
- If this model produces answers that are out of topic or if it talks to itself, it means you are not using the correct checkout 508de42 of qwopqwop200/GPTQ-for-LLaMa as mentioned below.
Cuda (Slow tokens/s):
git clone https://github.com/oobabooga/text-generation-webui
cd text-generation-webui
pip install -r requirements.txt
mkdir repositories
cd repositories
git clone https://github.com/qwopqwop200/GPTQ-for-LLaMa.git -b cuda # Make sure you obtain the qwopqwop200 version, not the oobabooga one! (because "act-order: yes")
cd GPTQ-for-LLaMa
pip install -r requirements.txt
python setup_cuda.py install
Triton (Fast tokens/s) - Works on Windows with WSL (what I've used) or Linux:
git clone https://github.com/oobabooga/text-generation-webui
cd text-generation-webui
git fetch origin pull/1229/head:triton # This is the version that supports Triton - https://github.com/oobabooga/text-generation-webui/pull/1229
git checkout triton
pip install -r requirements.txt
mkdir repositories
cd repositories
git clone https://github.com/qwopqwop200/GPTQ-for-LLaMa.git # -b cuda
cd GPTQ-for-LLaMa
git checkout 508de42 # Before qwopqwop200 broke everything... - https://github.com/qwopqwop200/GPTQ-for-LLaMa/issues/183
pip install -r requirements.txt
Testbench detail and results
Latest version of oobabooga + https://github.com/oobabooga/text-generation-webui/pull/1229
NVIDIA GTX 3090
32BG DDR4
i9-7980XE OC @4.6Ghz
11 tokens/s on average with Triton
Preliminary observations: better results than --load-in-8bits (To Be Confirmed)
Tested and working in both chat mode and text generation mode
Vicuna Model Card
Model details
Model type: Vicuna is an open-source chatbot trained by fine-tuning LLaMA on user-shared conversations collected from ShareGPT. It is an auto-regressive language model, based on the transformer architecture.
Model date: Vicuna was trained between March 2023 and April 2023.
Organizations developing the model: The Vicuna team with members from UC Berkeley, CMU, Stanford, and UC San Diego.
Paper or resources for more information: https://vicuna.lmsys.org/
License: Apache License 2.0
Where to send questions or comments about the model: https://github.com/lm-sys/FastChat/issues
Intended use
Primary intended uses: The primary use of Vicuna is research on large language models and chatbots.
Primary intended users: The primary intended users of the model are researchers and hobbyists in natural language processing, machine learning, and artificial intelligence.
Training dataset
70K conversations collected from ShareGPT.com.
Evaluation dataset
A preliminary evaluation of the model quality is conducted by creating a set of 80 diverse questions and utilizing GPT-4 to judge the model outputs. See https://vicuna.lmsys.org/ for more details.
Major updates of weights v1.1
- Refactor the tokenization and separator. In Vicuna v1.1, the separator has been changed from
"###"
to the EOS token"</s>"
. This change makes it easier to determine the generation stop criteria and enables better compatibility with other libraries. - Fix the supervised fine-tuning loss computation for better model quality.