Bittensor

Internet-scale Neural Networks

Discord • Docs • Network • Research • Code

This repository contains Bittensor's Python API, which can be used for the following purposes:

Querying the Bittensor network as a client.
Running and building Bittensor miners and validators for mining TAO.
Pulling network state information.
Managing TAO wallets, balances, transfers, etc.

Bittensor is a mining network, similar to Bitcoin, that includes built-in incentives designed to encourage miners to provide value by hosting trained or training machine learning models. These models can be queried by clients seeking inference over inputs, such as token-based text generations or numerical embeddings from a large foundation model like GPT-NeoX-20B.

Token-based incentives are designed to drive the network's growth and distribute the value generated by the network directly to the individuals producing that value, without intermediaries. The network is open to all participants, and no individual or group has full control over what is learned, who can profit from it, or who can access it.

To learn more about Bittensor, please read our paper.

1. Documentation
2. Install
3. Using Bittensor
4. Features
5. Release
6. License
7. Acknowledgments

1. Documentation

https://docs.bittensor.com/

2. Install

Three ways to install Bittensor

Through the installer:

$ /bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/opentensor/bittensor/master/scripts/install.sh)"

With pip:

$ pip3 install bittensor

From source:

$ git clone https://github.com/opentensor/bittensor.git
$ python3 -m pip install -e bittensor/

3. Using Bittensor

The following examples showcase how to use the Bittensor API for 3 separate purposes.

3.1. Client

Querying the network for generations.

import bittensor
wallet = bittensor.wallet().create_if_non_existent()
graph = bittensor.metagraph().sync()
print ( bittensor.dendrite( wallet = wallet ).generate
        ( 
            endpoints = graph.endpoints[graph.incentive.sort()[1][-1]],  // The highest ranked peer.
            prompt = "The quick brown fox jumped over the lazy dog", 
            num_to_generate = 20
        )
)

Querying the network for representations.

import bittensor
wallet = bittensor.wallet().create_if_non_existent()
graph = bittensor.metagraph().sync()
print ( bittensor.dendrite( wallet = wallet ).text_last_hidden_state
        (
            endpoints = graph.endpoints[graph.incentive.sort()[1][-1]],  // The highest ranked peer.
            inputs = "The quick brown fox jumped over the lazy dog"
        )
)
...
// Apply model. 
...
loss.backward() // Accumulate gradients on endpoints.

3.2. Server

Serving a custom model.

import bittensor
import torch
from transformers import GPT2Model, GPT2Config

model = GPT2Model( GPT2Config(vocab_size = bittensor.__vocab_size__, n_embd = bittensor.__network_dim__ , n_head = 8))
optimizer = torch.optim.SGD( [ {"params": model.parameters()} ], lr = 0.01 )

def forward_text( pubkey, inputs_x ):
    return model( inputs_x )
  
def backward_text( pubkey, inputs_x, grads_dy ):
    with torch.enable_grad():
        outputs_y = model( inputs_x.to(device) ).last_hidden_state
        torch.autograd.backward (
            tensors = [ outputs_y.to(device) ],
            grad_tensors = [ grads_dy.to(device) ]
        )
        optimizer.step()
        optimizer.zero_grad() 

wallet = bittensor.wallet().create().register()
axon = bittensor.axon (
    wallet = wallet,
    forward_text = forward_text,
    backward_text = backward_text
).start().serve()

3.3. Validator

Validating models by setting weights.

import bittensor
import torch

graph = bittensor.metagraph().sync()
dataset = bittensor.dataset()
chain_weights = torch.ones( [graph.n.item()], dtype = torch.float32 )

for batch in dataset.dataloader( 10 ):
    ...
    // Train chain_weights.
    ...
bittensor.subtensor().set_weights (
    weights = chain_weights,
    uids = graph.uids,
    wait_for_inclusion = True,
    wallet = bittensor.wallet(),
)

4. Features

4.1. CLI

Creating a new wallet.

$ btcli new_coldkey
$ btcli new_hotkey

Listing your wallets

$ btcli list

Registering a wallet

$ btcli register

Running a miner

$ btcli run

Checking balances

$ btcli overview

Checking the incentive mechanism.

$ btcli metagraph

Transfering funds

$ btcli transfer

Staking/Unstaking from a hotkey

$ btcli stake
$ btcli unstake

4.2. Selecting the network to join

There are two open Bittensor networks: staging (Nobunaga) and main (Nakamoto, Local).

Nobunaga (staging)
Nakamoto (main)
Local (localhost, mirrors nakamoto)

$ export NETWORK=local 
$ python (..) --subtensor.network $NETWORK
or
>> btcli run --subtensor.network $NETWORK

4.3. Running a template miner

The following command will run Bittensor's template miner

$ cd bittensor
$ python ./bittensor/_neuron/text/template_miner/main.py

>> import bittensor
>> bittensor.neurons.text.template_miner.neuron().run()

OR with customized settings

$ cd bittensor
$ python3 ./bittensor/_neuron/text/template_miner/main.py --wallet.name <WALLET NAME> --wallet.hotkey <HOTKEY NAME>

For the full list of settings, please run

$ python3 ~/.bittensor/bittensor/bittensor/_neuron/neurons/text/template_miner/main.py --help

4.4. Running a template server

The template server follows a similar structure as the template miner.

$ cd bittensor
$ python3 ./bittensor/_neuron/text/core_server/main.py --wallet.name <WALLET NAME> --wallet.hotkey <HOTKEY NAME>

>> import bittensor
>> bittensor.neurons.text.core_server.neuron().run()

For the full list of settings, please run

$ cd bittensor
$ python3 ./bittensor/_neuron/text/core_server/main.py --help

4.5. Serving an endpoint on the network

Endpoints are served to the bittensor network through the axon. The axon is instantiated via a wallet which holds an account on the Bittensor network.

import bittensor

wallet = bittensor.wallet().create().register()
axon = bittensor.axon (
    wallet = wallet,
    forward_text = forward_text,
    backward_text = backward_text
).start().serve()

4.6. Syncing with the chain/ Finding the ranks/stake/uids of other nodes

Information from the chain is collected/formated by the metagraph.

btcli metagraph

and

import bittensor

meta = bittensor.metagraph()
meta.sync()

# --- uid ---
print(meta.uids)

# --- hotkeys ---
print(meta.hotkeys)

# --- ranks ---
print(meta.R)

# --- stake ---
print(meta.S)

4.7. Finding and creating the endpoints for other nodes in the network

import bittensor

meta = bittensor.metagraph()
meta.sync()

### Address for the node uid 0
endpoint_as_tensor = meta.endpoints[0]
endpoint_as_object = meta.endpoint_objs[0]

4.8. Querying others in the network

import bittensor

meta = bittensor.metagraph()
meta.sync()

### Address for the node uid 0
endpoint_0 = meta.endpoints[0]

### Creating the wallet, and dendrite
wallet = bittensor.wallet().create().register()
den = bittensor.dendrite(wallet = wallet)
representations, _, _ = den.forward_text (
    endpoints = endpoint_0,
    inputs = "Hello World"
)

5. Release

The release manager should follow the instructions of the RELEASE_GUIDELINES.md document.

6. License

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

7. Acknowledgments

learning-at-home/hivemind

theeblardo/bittensor