Skip to content
This repository has been archived by the owner on Sep 26, 2024. It is now read-only.

Latest commit

 

History

History
241 lines (161 loc) · 6.69 KB

mining.md

File metadata and controls

241 lines (161 loc) · 6.69 KB

Full instructions for setup

Welcome to S19 Mining 🔥

Contents:

Overview

A miner consists of several parts, fitting into two categories:

  • Proxy
  • Workers

The proxy server is the server which has your hotkey, and spins up the axon (should probably be on a CPU close to your GPU worker servers). The worker servers are the GPU workers which perform the tasks!

I would advise starting with the worker servers. The simplest way to get started is to spin up four worker servers:

  • One for the mixtral chat model
  • One for the image generation stuff
  • One for the llama3 chat model

Worker server setup

I recommend using bare metal where possible. The documentation for bare metal can be found in here https://github.com/namoray/vision-workers

Below is the documentation for a template provider, such as runpod:

⚠️USE 11.8 <= CUDA <=12.2 For an easier life! ⚠️

Template setup

I'll use runpod as example, as it's easy with them

Chat (LLM) servers

FOR MORE INFO ON BARE METAL SETUP; READ HERE

Model configs

Mixtral

MODEL TheBloke/Nous-Hermes-2-Mixtral-8x7B-DPO-GPTQ
HALF_PRECISION true
REVISION gptq-8bit-128g-actorder_True

Llama-3

MODEL casperhansen/llama-3-70b-instruct-awq
HALF_PRECISION true
TOKENIZER tau-vision/llama-3-tokenizer-fix

It's very important that port 6919 is exposed here, as well as TCP port 22 ( so you can ssh in if you need to)

  • Create a GPU pod with this instance
  • Pick a GPU, pick a template, and you're off! and you're off! NOTE: I advise you use a estup with at least 80GB vram (H100 or A100 for example)

NOTE DONT USE THE TEMPLATE BELOW, ITS JUST AN EXAMPLE :D With template selected

Image worker

Very similar steps, just a slightly different template!

For detailed info on setup; read this image_server_docs Image server template

I would recommend either a RTX4090 or H100 or A100

Getting the address of the servers

Navigate to one of the pods, click 'connect', and then click the orange button, and copy the URL in your browser HTTP button

Proxy server setup

Get a CPU VM (Digital Ocean Droplet, OVH, Vultr, etc) - make sure you have an open port if you want to run a organic API server. Note: Runpod CPU's don't seem to be the best If you pick a runpod cpu, i would advise running this

ulimit -n 4096

Setup environment

Note: if you're using a provider such as runpod or vast, make sure you expose the necessary ports first, e.g.: https://docs.runpod.io/docs/expose-ports#:~:text=If%20your%20pod%20supports%20a,address%20to%20access%20your%20service. Follow the "Symmetrical port mapping" step :)

Clone the repo

git clone https://github.com/namoray/vision.git
cd vision

Install system dependencies

If you are in a container such as runpod, run these:

### Install pm2 & jq
apt update && apt upgrade -y
apt install nodejs npm -y
npm i -g pm2
apt-get install -y jq

### Install nano so you can edit the config easily
apt-get update
apt-get install nano

If you are on a bare metal machine (e.g. Vast) where you require sudo, use the following:

### Install pm2 & jq
sudo apt update && apt upgrade -y
sudo apt install nodejs npm -y
sudo npm i -g pm2
sudo apt-get install -y jq

### Install nano so you can edit the config easily
sudo apt-get update
sudo apt-get install nano

Install python dependencies

Make sure you have installed the correct python version (python 3.10). If you don't, try this:

sudo apt update && sudo apt install -y software-properties-common && \
sudo add-apt-repository ppa:deadsnakes/ppa && sudo apt install -y python3.10 \
python3.10-venv && python3.10 -m venv venv && source venv/bin/activate && echo "source venv/bin/activate">>~/.bashrc
### Install the local python environment
pip install --upgrade pip
pip install -e .

If for some reason that doesn't work, you may need to use pip3;

### Install the local python environment
pip3 install --upgrade pip
pip3 install -e .

Get hot and coldkeys onto your machine

I trust we can do this at this point ;D

Creating the database

Used to store concurrency info

sudo curl -fsSL -o /usr/local/bin/dbmate https://github.com/amacneil/dbmate/releases/latest/download/dbmate-linux-amd64
sudo chmod +x /usr/local/bin/dbmate

dbmate --url "sqlite:vision_database.db" up

Create the config

vision create-config

If you get the error message vision not found, you should make sure that requirements are correctly installed

Configure the task_config & task_concurrency_config json's

For each hotkey, there was default task configuration created in the sqlite db.

View this with

./peer_at_sql_db.sh

For a GUI, or

sudo apt install sqlite3
sqlite3 vision_database.db

For no gui

The default values for volumes are 1/2 of the maximum allowed values. These are filled EITHER:

  • When you use vision create-config
  • run python set_miner_defaults.py

Task config

Here we defined the capacitity (or volume) for each task, for that miner. This is the maximum amount of 'work' that a hotkey can do in a 1 hour period. To calculate that work, you can use calculate_volumes_example.py tool.

Concurrency groups

We also define the concurrency groups. All the tasks belong to a concurrency group, and you can configure a maximum number of concurrent requests for that group, which is shared between all the tasks in that group

Start miners

Autoupdates

Announcements will still be made ahead of time. If you choose to run autoupdates as a miner, make sure your ./start_miners.sh script is up to date & working, and I would advise monitoring around update releases regardless.

You're of course free to change or use whatever autoupdater you like!

pm2 start --name run_miner_auto_update "python run_miner_auto_update.py"

IF that doesn't start the Miner pm2 process, try this instead

nohup python run_miner_auto_update.py </dev/null &>miner_autoupdate.log &

No autoupdates

You can either use the command

./start_miners.sh