The ComputeHorde SDK enables Bittensor subnet owners to improve the quality of their subnets by providing validators with cost-effective, scalable compute resources. Instead of requiring validators to maintain their own physical GPUs, ComputeHorde offers on-demand decentralized access to trustworthy hardware, reducing costs and increasing validation power.
✔️ Lower Costs – Validators don’t need to buy, manage, or maintain GPUs.
✔️ Massive Scalability – Instant access to as many trustworthy GPUs as needed.
✔️ Faster Validation – Increased compute power leads to better validation.
✔️ Secure by Design – Only computation tasks are offloaded; private keys & weight setting remain on validator machines.
Subnet owners prepare validation code that can run both on ComputeHorde and on physical GPUs. This ensures validators can seamlessly opt in to use ComputeHorde’s computing power.
By enabling ComputeHorde support in their subnet, subnet owners benefit by:
✔️ Attracting More Validators – Lower costs and easier maintenance make validators more likely to participate.
✔️ Increasing Available Compute Power – More GPU resources mean faster, higher-quality validation.
✔️ Improving Subnet Quality – Stronger validation enhances the reliability and competitiveness of the subnet’s commodity.
Each validator chooses whether to use ComputeHorde. To gain access to ComputeHorde’s stake-based compute power, a validator must:
✔️ Be or become a ComputeHorde validator.
✔️ Partner with an existing ComputeHorde validator.
We actively support subnet owners and validators in integrating the SDK—both technically and business-wise. If you need guidance, reach out to us on the ComputeHorde Discord channel, and we’ll ensure your validator is ComputeHorde-ready.
For more details, see the ComputeHorde README.
To get started, install the ComputeHorde SDK with:
pip install compute-horde-sdk
For detailed API documentation, visit the ComputeHorde SDK Reference.
Important
This package uses ApiVer, make sure to import compute_horde_sdk.v1.
The ComputeHorde SDK allows validators to request ComputeHorde execution effortlessly. Below are examples demonstrating basic job execution, advanced job configuration, cross-validation, and job management.
This minimal example shows how to submit a job to ComputeHorde with basic parameters:
import asyncio
import bittensor_wallet
from compute_horde_sdk.v1 import ComputeHordeClient, ComputeHordeJobSpec, ExecutorClass
wallet = bittensor_wallet.Wallet(name="...", hotkey="...")
compute_horde_client = ComputeHordeClient(
hotkey=wallet.hotkey,
compute_horde_validator_hotkey="...", # usually the ss58_address of the hotkey above
)
async def main():
# Define your job
job_spec = ComputeHordeJobSpec(
executor_class=ExecutorClass.always_on__llm__a6000,
job_namespace="SN123.0",
docker_image="my-username/my-image:latest",
)
# Run the job
job = await compute_horde_client.run_until_complete(job_spec)
print(job.status) # should be "Completed".
asyncio.run(main())This example demonstrates how to submit a job with additional parameters, including:
- arguments & environment variables
- input & output volume configuration
- artifact (results) storage
For a full list of available parameters and detailed descriptions, see the ComputeHorde SDK Reference.
import asyncio
import bittensor_wallet
from compute_horde_sdk.v1 import (
ComputeHordeClient,
ComputeHordeJobSpec,
ExecutorClass,
HTTPInputVolume,
HTTPOutputVolume,
HuggingfaceInputVolume,
InlineInputVolume,
)
wallet = bittensor_wallet.Wallet(name="...", hotkey="...")
compute_horde_client = ComputeHordeClient(
hotkey=wallet.hotkey,
compute_horde_validator_hotkey="...", # usually the ss58_address of the hotkey above
)
async def main():
# Define your job
job_spec = ComputeHordeJobSpec(
executor_class=ExecutorClass.always_on__llm__a6000,
job_namespace="SN123.0",
docker_image="my-username/my-image:latest",
args=["main.py", "--block", "10000"],
env={"HF_HUB_ENABLE_HF_TRANSFER": "1"},
artifacts_dir="/artifacts",
input_volumes={
"/volume/models/model01": HuggingfaceInputVolume(
repo_id="my-username/my-model",
),
"/volume/data/version.txt": InlineInputVolume(
contents="dmVyc2lvbj0y",
),
"/volume/data/dataset.json": HTTPInputVolume(
url="https://my-dataset-bucket.s3.amazonaws.com/sample-dataset/data.json",
),
},
output_volumes={
"/output/image.png": HTTPOutputVolume(
http_method="PUT",
url="https://my-image-bucket.s3.amazonaws.com/images/image.png",
),
},
)
# Run the job
job = await compute_horde_client.run_until_complete(
job_spec=job_spec,
timeout=300, # retry/wait for up to 300 seconds for the job to complete
max_attempts=5, # try at most 5 times, if the job fails
)
print(job.status) # should be "Completed".
print(job.result)
asyncio.run(main())ComputeHorde lets you store output files externally using HTTP PUT or POST uploads to any S3-compatible service (e.g., AWS S3, Cloudflare R2).
⚠️ The SDK does not generate presigned URLs for you.
Instead, you must:
- Generate presigned POST/PUT URLs using your own code or SDK.
- Pass them into
output_volumesusingHTTPOutputVolume.
from compute_horde_sdk.v1 import HTTPOutputVolume
job_spec.output_volumes = {
"/output/image.png": HTTPOutputVolume(
http_method="PUT",
url="https://<your-s3-or-r2-presigned-url>",
)
}You can also include POST form fields using the form_fields dict parameter.
See send_hello_world_job.py for a working test that:
- Generates POST and PUT presigned URLs using
boto3 - Submits a job
- Verifies file upload succeeded
To ensure fairness and detect potential cheating, cross-validation should be performed on 1-2% of submitted jobs. This is done by executing a copy of a job on a trusted miner and comparing the results.
Flag on_trusted_miner=True tells create_job
to run the duplicate job on the ComputeHorde validator’s trusted miner instead of regular ComputeHorde miners.
If results differ, report the cheating miner using
report_cheated_job() method.
The details of the punishment are described in ComputeHorde's README.
Cross-validation is a crucial mechanism to ensure honest mining and maintain ComputeHorde’s reliability.
If you need to fetch a specific job, use get_job():
job = await client.get_job("7b522daa-e807-4094-8d96-99b9a863f960")To process all of your submitted jobs, use iter_jobs():
async for job in client.iter_jobs():
process(job)If job.status is "Completed", the job.result should be available.
If ComputeHorde is temporarily unavailable, you can run your jobs on a fallback cloud provider like RunPod using SkyPilot.
The SDK includes an optional FallbackClient that mirrors the standard ComputeHordeClient interface.
📦 Install with extras:
pip install compute-horde-sdk[fallback]
See the Fallback docs for usage examples and limitations.
This package uses Semantic Versioning.
TL;DR you are safe to use compatible release version specifier ~=MAJOR.MINOR in your pyproject.toml or requirements.txt.
Additionally, this package uses ApiVer to further reduce the risk of breaking changes.
This means, the public API of this package is explicitly versioned, e.g. compute_horde_sdk.v1, and will not change in a backwards-incompatible way even when compute_horde_sdk.v2 is released.
Internal packages, i.e. prefixed by compute_horde_sdk._ do not share these guarantees and may change in a backwards-incompatible way at any time even in patch releases.
Pre-requisites:
Ideally, you should run nox -t format lint before every commit to ensure that the code is properly formatted and linted.
Before submitting a PR, make sure that tests pass as well, you can do so using:
nox -t check # equivalent to `nox -t format lint test`
If you wish to install dependencies into .venv so your IDE can pick them up, you can do so using:
pdm install --dev
Run nox -s make_release -- X.Y.Z where X.Y.Z is the version you're releasing and follow the printed instructions.