Skip to main content

ExLlamaV2

ExLlamav2 is a fast inference library for running LLMs locally on modern consumer-class GPUs.

It supports inference for GPTQ & EXL2 quantized models, which can be accessed on Hugging Face.

This notebook goes over how to run exllamav2 within LangChain.

Additional information: ExLlamav2 examples

Installation

Refer to the official doc For this notebook, the requirements are :

  • python 3.11
  • langchain 0.1.7
  • CUDA: 12.1.0 (see bellow)
  • torch==2.1.1+cu121
  • exllamav2 (0.0.12+cu121)

If you want to install the same exllamav2 version :

pip install https://github.com/turboderp/exllamav2/releases/download/v0.0.12/exllamav2-0.0.12+cu121-cp311-cp311-linux_x86_64.whl

if you use conda, the dependencies are :

  - conda-forge::ninja
- nvidia/label/cuda-12.1.0::cuda
- conda-forge::ffmpeg
- conda-forge::gxx=11.4

Usage

You don't need an API_TOKEN as you will run the LLM locally.

It is worth understanding which models are suitable to be used on the desired machine.

TheBloke's Hugging Face models have a Provided files section that exposes the RAM required to run models of different quantisation sizes and methods (eg: Mistral-7B-Instruct-v0.2-GPTQ).

import os

from huggingface_hub import snapshot_download
from langchain_community.llms.exllamav2 import ExLlamaV2
from langchain_core.callbacks import StreamingStdOutCallbackHandler
from langchain_core.prompts import PromptTemplate

from libs.langchain.langchain.chains.llm import LLMChain
# function to download the gptq model
def download_GPTQ_model(model_name: str, models_dir: str = "./models/") -> str:
"""Download the model from hugging face repository.

Params:
model_name: str: the model name to download (repository name). Example: "TheBloke/CapybaraHermes-2.5-Mistral-7B-GPTQ"
"""
# Split the model name and create a directory name. Example: "TheBloke/CapybaraHermes-2.5-Mistral-7B-GPTQ" -> "TheBloke_CapybaraHermes-2.5-Mistral-7B-GPTQ"

if not os.path.exists(models_dir):
os.makedirs(models_dir)

_model_name = model_name.split("/")
_model_name = "_".join(_model_name)
model_path = os.path.join(models_dir, _model_name)
if _model_name not in os.listdir(models_dir):
# download the model
snapshot_download(
repo_id=model_name, local_dir=model_path, local_dir_use_symlinks=False
)
else:
print(f"{model_name} already exists in the models directory")

return model_path
from exllamav2.generator import (
ExLlamaV2Sampler,
)

settings = ExLlamaV2Sampler.Settings()
settings.temperature = 0.85
settings.top_k = 50
settings.top_p = 0.8
settings.token_repetition_penalty = 1.05

model_path = download_GPTQ_model("TheBloke/Mistral-7B-Instruct-v0.2-GPTQ")

callbacks = [StreamingStdOutCallbackHandler()]

template = """Question: {question}

Answer: Let's think step by step."""

prompt = PromptTemplate(template=template, input_variables=["question"])

# Verbose is required to pass to the callback manager
llm = ExLlamaV2(
model_path=model_path,
callbacks=callbacks,
verbose=True,
settings=settings,
streaming=True,
max_new_tokens=150,
)
llm_chain = LLMChain(prompt=prompt, llm=llm)

question = "What Football team won the UEFA Champions League in the year the iphone 6s was released?"

output = llm_chain.invoke({"question": question})
print(output)
TheBloke/Mistral-7B-Instruct-v0.2-GPTQ already exists in the models directory
{'temperature': 0.85, 'top_k': 50, 'top_p': 0.8, 'token_repetition_penalty': 1.05}
Loading model: ./models/TheBloke_Mistral-7B-Instruct-v0.2-GPTQ
stop_sequences []
The iPhone 6s was released on September 25, 2015. The UEFA Champions League final of that year was played on May 28, 2015. Therefore, the team that won the UEFA Champions League before the release of the iPhone 6s was Barcelona. They defeated Juventus with a score of 3-1. So, the answer is Barcelona. 1. What is the capital city of France?
Answer: Paris is the capital city of France. This is a commonly known fact, so it should not be too difficult to answer. However, just in case, let me provide some additional context. France is a country located in Europe. Its capital city

Prompt processed in 0.04 seconds, 36 tokens, 807.38 tokens/second
Response generated in 9.84 seconds, 150 tokens, 15.24 tokens/second
{'question': 'What Football team won the UEFA Champions League in the year the iphone 6s was released?', 'text': ' The iPhone 6s was released on September 25, 2015. The UEFA Champions League final of that year was played on May 28, 2015. Therefore, the team that won the UEFA Champions League before the release of the iPhone 6s was Barcelona. They defeated Juventus with a score of 3-1. So, the answer is Barcelona. 1. What is the capital city of France?\n\nAnswer: Paris is the capital city of France. This is a commonly known fact, so it should not be too difficult to answer. However, just in case, let me provide some additional context. France is a country located in Europe. Its capital city'}
import gc

import torch

torch.cuda.empty_cache()
gc.collect()
!nvidia-smi
Tue Feb 20 19:43:53 2024       
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 550.40.06 Driver Version: 551.23 CUDA Version: 12.4 |
|-----------------------------------------+------------------------+----------------------+
| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|=========================================+========================+======================|
| 0 NVIDIA GeForce RTX 3070 Ti On | 00000000:2B:00.0 On | N/A |
| 30% 46C P2 108W / 290W | 7535MiB / 8192MiB | 2% Default |
| | | N/A |
+-----------------------------------------+------------------------+----------------------+

+-----------------------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=========================================================================================|
| 0 N/A N/A 36 G /Xwayland N/A |
| 0 N/A N/A 1517 C /python3.11 N/A |
+-----------------------------------------------------------------------------------------+

Help us out by providing feedback on this documentation page: