license: apache-2.0
language: fr
pipeline_tag: text-generation
inference:
parameters:
temperature: 0.7
tags:
- LLM
- finetuned
Vigostral-7B-Chat: A French chat LLM
Preview of Vigostral-7B-Chat, a new addition to the Vigogne LLMs family, fine-tuned on Mistral-7B-v0.1.
For more information, please visit the Github repository.
License: A significant portion of the training data is distilled from GPT-3.5-Turbo and GPT-4, kindly use it cautiously to avoid any violations of OpenAI's terms of use.
Prompt Template
We used a prompt template adapted from the chat format of Llama-2.
You can apply this formatting using the chat template through the apply_chat_template()
method.
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("bofenghuang/vigostral-7b-chat")
conversation = [
{"role": "user", "content": "Bonjour ! Comment ça va aujourd'hui ?"},
{"role": "assistant", "content": "Bonjour ! Je suis une IA, donc je n'ai pas de sentiments, mais je suis prêt à vous aider. Comment puis-je vous assister aujourd'hui ?"},
{"role": "user", "content": "Quelle est la hauteur de la Tour Eiffel ?"},
{"role": "assistant", "content": "La Tour Eiffel mesure environ 330 mètres de hauteur."},
{"role": "user", "content": "Comment monter en haut ?"},
]
print(tokenizer.apply_chat_template(conversation, tokenize=False, add_generation_prompt=True))
You will get
<s>[INST] <<SYS>>
Vous êtes Vigogne, un assistant IA créé par Zaion Lab. Vous suivez extrêmement bien les instructions. Aidez autant que vous le pouvez.
<</SYS>>
Bonjour ! Comment ça va aujourd'hui ? [/INST] Bonjour ! Je suis une IA, donc je n'ai pas de sentiments, mais je suis prêt à vous aider. Comment puis-je vous assister aujourd'hui ? </s>[INST] Quelle est la hauteur de la Tour Eiffel ? [/INST] La Tour Eiffel mesure environ 330 mètres de hauteur. </s>[INST] Comment monter en haut ? [/INST]
Usage
Inference using the quantized versions
The quantized versions of this model are generously provided by TheBloke!
- AWQ for GPU inference: TheBloke/Vigostral-7B-Chat-AWQ
- GTPQ for GPU inference: TheBloke/Vigostral-7B-Chat-GPTQ
- GGUF for CPU+GPU inference: TheBloke/Vigostral-7B-Chat-GGUF
These versions facilitate testing and development with various popular frameworks, including AutoAWQ, vLLM, AutoGPTQ, GPTQ-for-LLaMa, llama.cpp, text-generation-webui, and more.
Inference using the unquantized model with 🤗 Transformers
from typing import Dict, List, Optional
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig, TextStreamer
model_name_or_path = "bofenghuang/vigostral-7b-chat"
tokenizer = AutoTokenizer.from_pretrained(model_name_or_path, padding_side="right", use_fast=False)
model = AutoModelForCausalLM.from_pretrained(model_name_or_path, torch_dtype=torch.float16, device_map="auto")
streamer = TextStreamer(tokenizer, timeout=10.0, skip_prompt=True, skip_special_tokens=True)
def chat(
query: str,
history: Optional[List[Dict]] = None,
temperature: float = 0.7,
top_p: float = 1.0,
top_k: float = 0,
repetition_penalty: float = 1.1,
max_new_tokens: int = 1024,
**kwargs,
):
if history is None:
history = []
history.append({"role": "user", "content": query})
input_ids = tokenizer.apply_chat_template(history, return_tensors="pt").to(model.device)
input_length = input_ids.shape[1]
generated_outputs = model.generate(
input_ids=input_ids,
generation_config=GenerationConfig(
temperature=temperature,
do_sample=temperature > 0.0,
top_p=top_p,
top_k=top_k,
repetition_penalty=repetition_penalty,
max_new_tokens=max_new_tokens,
pad_token_id=tokenizer.eos_token_id,
**kwargs,
),
streamer=streamer,
return_dict_in_generate=True,
)
generated_tokens = generated_outputs.sequences[0, input_length:]
generated_text = tokenizer.decode(generated_tokens, skip_special_tokens=True)
history.append({"role": "assistant", "content": generated_text})
return generated_text, history
# 1st round
response, history = chat("Un escargot parcourt 100 mètres en 5 heures. Quelle est sa vitesse ?", history=None)
# 2nd round
response, history = chat("Quand il peut dépasser le lapin ?", history=history)
# 3rd round
response, history = chat("Écris une histoire imaginative qui met en scène une compétition de course entre un escargot et un lapin.", history=history)
You can also use the Google Colab Notebook provided below.
Inference using the unquantized model with vLLM
Set up an OpenAI-compatible server with the following command:
# Install vLLM
# This may take 5-10 minutes.
# pip install vllm
# Start server for Vigostral-Chat models
python -m vllm.entrypoints.openai.api_server --model bofenghuang/vigostral-7b-chat
# List models
# curl http://localhost:8000/v1/models
You can also use the docker image provided below.
# Launch inference engine
docker run --gpus '"device=0"' \
-e HF_TOKEN=$HF_TOKEN -p 8000:8000 \
ghcr.io/bofenghuang/vigogne/vllm:latest \
--host 0.0.0.0 \
--model bofenghuang/vigostral-7b-chat
# Launch inference engine on mutli-GPUs (4 here)
docker run --gpus all \
-e HF_TOKEN=$HF_TOKEN -p 8000:8000 \
ghcr.io/bofenghuang/vigogne/vllm:latest \
--host 0.0.0.0 \
--tensor-parallel-size 4 \
--model bofenghuang/vigostral-7b-chat
# Launch inference engine using the quantized AWQ version
# Note only supports Ampere or newer GPUs
docker run --gpus '"device=0"' \
-e HF_TOKEN=$HF_TOKEN -p 8000:8000 \
ghcr.io/bofenghuang/vigogne/vllm:latest \
--host 0.0.0.0 \
--quantization awq \
--model TheBloke/Vigostral-7B-Chat-AWQ
Afterward, you can query the model using the openai Python package.
import openai
# Modify OpenAI's API key and API base to use vLLM's API server.
openai.api_key = "EMPTY"
openai.api_base = "http://localhost:8000/v1"
# First model
models = openai.Model.list()
model = models["data"][0]["id"]
query_message = "Parle-moi de toi-même."
# Chat completion API
chat_completion = openai.ChatCompletion.create(
model=model,
messages=[
{"role": "user", "content": query_message},
],
max_tokens=1024,
temperature=0.7,
)
print("Chat completion results:", chat_completion)
Limitations
Vigogne is still under development, and there are many limitations that have to be addressed. Please note that it is possible that the model generates harmful or biased content, incorrect information or generally unhelpful answers.