Subnet Hyperparameters

Suggest Edits

In addition to miners and validator code, each subnet has a set of parameters (often referred to as hyper parameters) that define how it will interact with the bittensor network.

To see the parameters of a subnet, use the btcli su getor btcli s hyperparameters

Here are example parameters from Subnet 15 at the time of writing:

btcli sudo get
Enter netuid [0/1/2/3/4/5/6/7/8/9/10/11/12/13/14/15/16/17/18/19/20/21/22/23/24/25/26/27/28/29/30/31/32] (0): 15
  Subnet Hyperparameters - NETUID: 15 - finney  
 HYPERPARAMETER            VALUE                
 rho                       10                   
 kappa                     32767                
 immunity_period           1000                 
 min_allowed_weights       1                    
 max_weight_limit          65535                
 tempo                     360                  
 min_difficulty            18446744073709551615 
 max_difficulty            18446744073709551615 
 weights_version           0                    
 weights_rate_limit        100                  
 adjustment_interval       360                  
 activity_cutoff           5000                 
 registration_allowed      False                
 target_regs_per_interval  1                    
 min_burn                  1                    
 max_burn                  100000000000         
 bonds_moving_avg          900000               
 max_regs_per_block        1                    
 serving_rate_limit        50                   
 max_validators            64       
 adjustment_alpha          58000
 difficulty                18446744073709551615
  • rho: 10 for all subnets. Rho (p) is calculated based on the network's target inflation and actual neuron staking. It adjusts the emission rate of the TAO token to balance the network's economy and dynamics. The formula for Rho is defined as: p = (Staking_Target / Staking_Actual) * Inflation_Target. Here, Staking_Target and Staking_Actual represent the desired and actual total stakes in the network, while Inflation_Target is the predefined inflation rate goal.
  • kappa: 32767 for all subnets. Kappa (κ) is used in the calculation of neuron ranks, which determine their share of network incentives. It is derived from the softmax function applied to the inter-neuronal weights set by each neuron. The formula for Kappa is: κ_i = exp(w_i) / Σ(exp(w_j)), where w_i represents the weight set by neuron i, and the denominator is the sum of exponential weights set by all neurons. This mechanism ensures a normalized and probabilistic distribution of ranks based on relative weights.
  • Immunity Period: This parameter defines the duration during which new neurons are protected from certain network penalties or restrictions. (At 12 s/block, 1000 blocks is ~ 3.3 minutes.)
  • Min Allowed Weights: The minimum number of UIDs a subnet validator must set weights on, before the subnet validator is allowed to set weights on the blockhain.
  • Max Weight Limit: Highest weight value that can be set by a validator. This is a float value.
  • Tempo: Cadence of updates, in blocks.
  • mix/max difficulty: Obsolete - no longer used.
  • weights version: Sets the minimum version of validator code that a validator can use (and still set weights).
  • weights rate limit: time (in blocks) that a validator may update weights.
  • adjustment_interval: Time (in blocks) after which the node registration cost is re-evaluated. If the number of actual registrations that occurred in the last adjustment_interval is higher than the target_regs_per_interval, then the blockchain will raise the recycle register cost, by increasing the min_burn value by a certain amount, in order to slow down the actual registrations and bring them back to target_regs_per_interval value.
  • activity cutoff:
    • Expressed in number of blocks. If a subnet validator has not set weights on the blockchain for activity_cutoff duration, then the Yuma Consensus will consider this subnet validator as offline, i.e., turned off. The weights of this subnet validator are considered too old to be useful. The weights of this subnet validator slowly lose their impact over time and eventually will no longer be considered for consensus calculation.
  • registration_allowed: When a subnet updates or there is an issue - the subnet owner may pause registrations to allow for all miners and validators to update their servers without the risk of deregistration.
  • target_regs_per_interval: Number of nodes that can be registered per epoch. In the example above, just one node can be registered per interval.
  • min_burn/max_burn: Node registration costs (in rao - 1e-9 tao). The max listed above is 100 tao.
  • serving_rate_limit: Determines how often you can change your node's IP address on the blockchain. Expressed in number of blocks. Applies to both subnet validator and subnet miner nodes. Used when you move your node to a new machine.
  • Bonds Moving Avg :
    • This parameter controls how fast bonds will decay in the entire subnet. This is a unitless number. This number has a direct impact on subnet validator. The faster the bonds decay the quicker a subnet validator will lose its dividends after the subnet validator is out of the activity_cutoff.
      If this bonds_moving_avg value is low, then the moving average of the bonds will decay slowly. This will allow the subnet validator to become active again, start setting new weights and start earning new bonds.
      If this bonds_moving_avg value is high, then bonds in the subnet decay quickly. As a result, a subnet validator who has fallen out of the activity_cutoff and hence is running the risk of being viewed as "turned off", may not be able to become active again.
  • Max Validators: The number of validator slots available in the subnet.
  • Adjustment alpha: Larger numbers smooth the registration burn costs epoch to epoch.

Updated about 1 month ago