Adds flag --neuron.use_compute_horde for the validator, to perform computations on the Compute Horde instead of locally.

Usage

Requires some new COMPUTE_HORDE_ settings in .env:

COMPUTE_HORDE_VALIDATOR_HOTKEY=<your hotkey>
# optional, defaults to production
# COMPUTE_HORDE_FACILITATOR_URL=
# optional, defaults to LLM A6000
# COMPUTE_HORDE_EXECUTOR_CLASS=
# This is this repo's Docker image, so you can docker build . -t <your image> and use that
COMPUTE_HORDE_JOB_DOCKER_IMAGE=backenddevelopersltd/slawek-test:v0-latest
COMPUTE_HORDE_JOB_NAMESPACE=SN15.0

To run the validator using Compute Horde, register as validator on SN15 (see the README for validators) and run:

poetry run python neurons/validator.py --netuid 15     --subtensor.network finney     --wallet.name default     --wallet.hotkey default     --logging.debug     --logging.trace     --axon.port 3000     --neuron.model_ids 'gpt-4o,gpt-4o-mini,mistral-7b,Claude-3.5,command-r-plus' --neuron.axon_off --neuron.num_task_examples=1 --neuron.disable_set_weights --neuron.use_compute_horde

You can also use pm2 as recommended.

To run a single job (miner model) on the Compute Horde without starting a validator loop – this doesn't require registration and acts as a "sanity check":

poetry run python scripts/compute_horde_e2e_test.py --hf-id=AltEinstein/emc9 --hotkey=5G9BymEWsPac2CXZsar9DpScZAjacr69SgPRNpSDb1fPza8M --coldkey=5H4MtMvFdK2UKdwXMwpSc6h95K7MsK5n6BttYw7Gen93Ck4e

(replace the flag values with any model you want to validate)

Implementation

The flow looks roughly like this:

1. A Docker image is created of this codebase
2. The validator loop is running as usual
3. When a miner model is to be validated, instead of doing that locally, a Job is sent to the Compute Horde with Docker image from step 1 and miner model as an input
4. Validation is performed on the Compute Horde (runs the same code as it would locally, but the entrypoint is different, neurons/compute_horde_entrypoint.py)
5. Results are returned to the validator loop (step 2) and processed further as usual (ranking, setting weights)

TODO

• Install compute-horde-sdk from PyPI rather than git