Skip to content

Latest commit

 

History

History
68 lines (49 loc) · 3.51 KB

CustomRunner.md

File metadata and controls

68 lines (49 loc) · 3.51 KB
Raw

QLI Custom Runner (DRAFT!)

You have the possibility to use your own runner. A runner in our terms is a worker process which does the AI training. By Default, qli-Client checks your local configuration and loads the current runner for the acutal epoch and starts this as a worker process as needed. A custom runner is useful if you code your own trainer or if you want to experiment.

  1. Prerequisites
  2. Communication
  3. Qubic/Qiner Protocol
    1. Command Line Arguments
    2. STDOUT

Why do i need to run the qli-Client?

The qli-Client ensures a secure communcation with the qli backend API. The qli-Client takes care that you always run on the correct ID and reports the current state of the worker. Without prior authorization, it is not allowed to use a custum Client for the communication with the API.

By Default, the Client transmits the following information to the backend:

  • Current Training Task Information (ID, it/s)
  • Solutions
  • CPU Capability: AVX2/AVX512
  • OS: Windows, Linux, OSX
  • Alias: The name of the client instance

Prerequisites

You need the qli-Client Version 1.3.9 or greater to use all the described options.

The Configuration Options for a custom runner are:

Option Default Value Description
customRunner false Set this to true to use your custom runner
trainerBinary "qli-runner" With this option you can change the name of the runner.
serviceLock false Set this to true if you want to use Qiner as a runner

To use your own runner, set customRunner to true.

Communication

For the communication between the runner and the client you can use the standard Qubic/Qiner protocol: The qli-Client starts the worker process and observes it's STDOUT.

if you need a custom solution ask us in discord.

Qubic/Qiner Protocol

With this, your custom runner just needs to behave like the official Qiner from Qubic. (https://github.com/Qubic-World/qubic-miner-cpu)

Important: the stdout needs to be flushed regularly. In C++ you can use fflush(stdout); after every printout of the current state. (In Orignal Qunier this is right after printf("| %d-%d%d-%d%d %d%d:%d%d:%d%d | %d it/s | %d solutions | %.10s... |\n", systemTime.wYear, systemTime.wMonth / 10, systemTime.wMonth % 10, systemTime.wDay / 10, systemTime.wDay % 10, systemTime.wHour / 10, systemTime.wHour % 10, systemTime.wMinute / 10, systemTime.wMinute % 10, systemTime.wSecond / 10, systemTime.wSecond % 10, (numberOfMiningIterations - prevNumberOfMiningIterations) * 1000 / delta, numberOfFoundSolutions, argv[2]);)

If you use the Qiner protocol you can run only one instance of qli-Client per maschine.

Limitation: only one instance of qli-Client can run with Quiner protocol

Command Line Arguments Qiner

Your runner must accept the following commands: <IP> <ID> <NUMBEROFTHREADS where

  • will be the IP Address to report the found solution to
  • The Identity/ID to use for the training
  • The Number of Threads to be used for the training

STDOUT Qiner

Your custom runner must ouput at least the following:

  • X solutions => X ist the amount of found solutions
  • Y it/s => Y are the current iterations per seconds

these two parameters will be parsed by the client and displayed in the unified qli format. Including an average of it/s over the last 100 messages.

Solution Deliviery

Use the standard Quiner protocol.