SUBMIT A PRSUBMIT AN ISSUElast edit: Jul 02, 2024

Consensus-based Weights

This guide describes how to use the consensus-based weights feature (also called "liquid alpha").

With this feature, the dividends to a subnet validator are better correlated to the performance of the subnet miner on which the subnet validator is setting the weights. In this context, see also the documentation for the Commit Reveal feature, as both these features help the subnet validators in finding new subnet miners that perform well and bonding to them quickly.

Technical paper, blog

• See Amplifying the Weight-copying Penalty in Bittensor, a working paper (PDF).
• Blog post: Consensus-based Weights.
• Subtensor document section: Validator bonding.

Collab notebooks

A subnet owner can run the weight_copy/liquid_alpha_diagnostic.ipynb in the below Python notebook to experiment and choose the right values for the hyperparameters alpha_low, alpha_high, commit_reveal_interval.

• For commit reveal diagnostic: https://colab.research.google.com/github/opentensor/developer-docs/blob/main/static/weight_copy/commit_reveal_diagnostic.ipynb?authuser=5
• For liquid alpha diagnostic: https://colab.research.google.com/github/opentensor/developer-docs/blob/main/static/weight_copy/liquid_alpha_diagnostic.ipynb?authuser=5
• GitHub directory with Python notebooks.

Description

Currently while calculating the dividends to a subnet validator, a quantity called $\Delta B_{ij}$, defined as an instantaneous bond value of a subnet validator $i$ with the subnet miner $j$, is used.

A subnet validator maintains an instantaneous bond value on each subnet miner. However, while calculating a subnet validator's dividends, instead of directly using the instantaneous bond value $\Delta B_{ij}$, we use $B_{ij}^{(t)}$, an exponential moving average (EMA) of the bond value, weighted over the current epoch and the previous epoch. See the below equation for how this EMA is computed, where $t$ is the current epoch time and $t-1$ is the previous epoch time:

$$B_{ij}^{(t)} = \alpha\cdot\Delta B_{ij}^{(t)} + (1-\alpha)\cdot B_{ij}^{(t-1)}$$

This EMA, $B_{ij}^{(t)}$, helps in early discovery of promising subnet miners and prevents abrupt changes to the bond value. Normally any abrupt change in the bond value indicates exploitation.

Finally, the dividend $D_i$ to a subnet validator $i$ is calculated as:

$$D_i = \sum_j B_{ij} \cdot I_j$$

where $B_{ij}$ is the EMA bond value of the subnet validator $i$ with the subnet miner $j$, and $I_j$ is the subnet miner's incentive. See the subtensor document section, Validator bonding for a rigorous mathematical treatment of this topic.

What changed with this feature

Without the consensus-based weights feature, the $\alpha$ in the above equation is set to 0.9. With the consensus-based weights feature, this $\alpha$ value is made into a variable. An optimium value for the variable $\alpha$ is determined based on the current consensus (YC-2) in a given subnet. Hence this feature is called consensus-based weights.

Using the new subnet hyperparameters that are described below, a subnet owner should experiment and discover the optimum $\alpha$ for their subnet.

Installing the test consensus-based weights feature

The consensus-based weights feature is available only at the below testnet URL and the specified subtensor and bittensor repo branches:

• Testnet URL: wss://test.finney.opentensor.ai:443/.
• Subtensor repo: Make sure to checkout the tag 1.2.1-pre-release on the Subtensor repo https://github.com/opentensor/subtensor.
• Bittensor repo and branch:
  1. First make sure you do git pull master to pull the latest changes from the Bittensor repo https://github.com/opentensor/bittensor.
  2. Next, switch to rc_7.3.0 branch.
  3. Then install Bittensor by running the below command:

  pip install -e .

---

Using the test consensus-based weights feature

Summary steps

Here are summary steps to use the consensus-based weights feature. These steps are typically executed by a subnet owner:

1. To activate this feature, a subnet owner should set the liquid_alpha_enabled (bool) hyperparameter to True.
2. Next, the subnet owner should set the upper and lower bounds for $\alpha$ by using a single subnet hyperparameter, alpha_values (List[int]).

you must set both low and high alpha values together. See below.

---

Default values, allowed ranges and value format

Default values

• Default value for alpha_low is 0.7.
• Default value for alpha_high is 0.9.
• Default value for liquid_alpha_enabled is False.

Allowed ranges

• The range for both alpha_low and alpha_high hyperparameters is (0,1).
• However, until further notice, the alpha_high value must be greater than or equal to 0.8.
• The value of alpha_low must not be greater than or equal to alpha_high.

Value format

When you set the subnet hyperparameters alpha_low and alpha_high, you must pass their integer equivalents in u16. This applies whether you set these hyperparameters using the btcli command or in your Python code. These integer values are then converted by the subtensor into their corresponding decimal values in the u16 format.

Use the below conversion formula to determine the integer values for your desired decimal values for both alpha_low and alpha_high hyperparameters.

$$\text{Integer value} = \text{(your-desired-decimal-value)} \times 65536$$

Hence, for example:

• If you want alpha_low to be 0.1, then you would pass 6554, which is the rounded up value of 0.1 * 65536.
• If you want alpha_high to be 0.8, then you would pass 52429, which is the rounded up value of 0.8 * 65536.

---

Detailed steps using Python code

Method signatures

See below the Python definitions for the consensus-based weights feature:

import bittensor as bt
    # Enable consensus-based weights (liquid alpha) feature
    enabled_result = bt.subtensor.set_hyperparameter(
        wallet=wallet,
        netuid=netuid,
        parameter="liquid_alpha_enabled",
        value=value,
        wait_for_inclusion=True,
        wait_for_finalization=True,
    )
    # Set alpha_values as a list of integers passed to "value" parameter in this order: alpha_low, alpha_high
    alpha_low_high_result = bt.subtensor.set_hyperparameter(
        wallet=wallet,
        netuid=netuid,
        parameter="alpha_values",
        value=[],
        wait_for_inclusion=True,
        wait_for_finalization=True,
    )

Example Python code

Below is the example Python code showing how to use the above definitions for the commit reveal feature:

    import bittensor as bt
    # Enable consensus-based weights (liquid alpha) feature
    enabled_result = bt.subtensor.set_hyperparameter(
        wallet=wallet,
        netuid=netuid,
        parameter="liquid_alpha_enabled",
        value=True,
        wait_for_inclusion=True,
        wait_for_finalization=True,
    )
    # Set alpha low and high values
    alpha_low_high_result = bt.subtensor.set_hyperparameter(
        wallet=wallet,
        netuid=netuid,
        parameter="alpha_values",
        value=[6554, 52429], # decimal 0.1 for alpha_low and 0.8 for alpha_high
        wait_for_inclusion=True,
        wait_for_finalization=True,
    )

you must always set alpha_low and alpha_high together

You must set the values for both alpha_low and alpha_high together. Current functionality does not allow setting a value to only one of alpha_low or alpha_high. For example, if you want to set a new value to alpha_low but do not want to change the alpha_high value, you must set the new value to alpha_low and also set again the current, unchanging value to alpha_high.

---

Detailed steps using btcli

Set the subnet hyperparameters

1. Enable the consensus-based weights feature

Syntax

btcli sudo set hyperparameters --netuid <NETUID> --param liquid_alpha_enabled --value <True or False>

Example

For subnet 1 (netuid of 1):

btcli sudo set hyperparameters --netuid 1 --param liquid_alpha_enabled --value True

or you can also use,

btcli sudo set

and follow the terminal prompts.

tip

When you use btcli sudo set you can use 1 or 0 to enable or disable the liquid_alpha_enabled hyperparameter. You can also use true or True, or false or False.

2. Use the alpha_values to set the alpha_low and alpha_high

Syntax

btcli sudo set hyperparameters --netuid <NETUID> --param alpha_values --value <value-of-alpha_low, value-of-alpha_high>

Example

Setting the value of alpha_low to the decimal 0.1 (integer 6554) and alpha_high to the decimal 0.8 (integer 52429) for subnet 1 (netuid of 1):

btcli sudo set hyperparameters --netuid 1 --param alpha_values --value 6554, 52429

Now if you want to only change the alpha_high value from 0.8 to 0.85 (integer 55706) then:

First execute the btcli sudo set command and provide the netuid. The terminal will display the current values of alpha_low and alpha_high.

alert

When you use btcli sudo set the display will not show alpha_values parameter. It will only show alpha_low and alpha_high parameters.

Use the current value of alpha_low from the above display, and the new desired value of alpha_high and set both like below:

btcli sudo set hyperparameters --netuid 1 --param alpha_values --value 6554, 55706