Struct Mamba2NetworkConfig 

pub struct Mamba2NetworkConfig {
    pub n_real_layers: usize,
    pub n_virtual_layers: Option<(usize, Schedule)>,
    pub vocab_size: usize,
    pub pad_vocab_size_multiple: usize,
    pub mamba_block: Mamba2Config,
    pub missing_lm_head: bool,
}

Configuration / factory for Mamba2Network.

Fields

n_real_layers: usize

Number of real (weight-bearing) Mamba-2 layers.

n_virtual_layers: Option<(usize, Schedule)>

Optional virtual-layer scheduling. See Mamba2Layers for details.

vocab_size: usize

The unpadded vocabulary size as specified by the tokenizer.

At initialisation this value is rounded up to the nearest multiple of pad_vocab_size_multiple to obtain the actual embedding / logit dimension padded_vocab_size.

pad_vocab_size_multiple: usize

Vocabulary size will be rounded up to a multiple of this value.

Set to 1 to disable rounding. Common values: 8, 16, 64.

mamba_block: Mamba2Config

Configuration shared by all Mamba-2 blocks.

missing_lm_head: bool

When true, the LM head weight is not allocated separately; instead the transposed embedding matrix is used directly (weight tying).

Implementations

impl Mamba2NetworkConfig

pub fn new( n_real_layers: usize, vocab_size: usize, pad_vocab_size_multiple: usize, mamba_block: Mamba2Config, missing_lm_head: bool, ) -> Self

Create a new instance of the config.

Arguments

Required Arguments

n_real_layers

Number of real (weight-bearing) Mamba-2 layers.

vocab_size

The unpadded vocabulary size as specified by the tokenizer.

At initialisation this value is rounded up to the nearest multiple of pad_vocab_size_multiple to obtain the actual embedding / logit dimension padded_vocab_size.

pad_vocab_size_multiple

Vocabulary size will be rounded up to a multiple of this value.

Set to 1 to disable rounding. Common values: 8, 16, 64.

mamba_block

Configuration shared by all Mamba-2 blocks.

missing_lm_head

When true, the LM head weight is not allocated separately; instead the transposed embedding matrix is used directly (weight tying).

Default Arguments

n_virtual_layers

Optional virtual-layer scheduling. See Mamba2Layers for details.

• Defaults to "None"

impl Mamba2NetworkConfig

pub fn with_n_virtual_layers( self, n_virtual_layers: Option<(usize, Schedule)>, ) -> Self

Sets the value for the field n_virtual_layers.

Optional virtual-layer scheduling. See Mamba2Layers for details.

• Defaults to "None"

impl Mamba2NetworkConfig

pub fn init<B: Backend>(&self, device: &B::Device) -> Mamba2Network<B>

Allocate and initialise the full network on device.

Trait Implementations

impl Clone for Mamba2NetworkConfig

fn clone(&self) -> Self

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Config for Mamba2NetworkConfig

fn load_binary(data: &[u8]) -> Result<Self, ConfigError>

Loads the configuration from a binary buffer.

impl Debug for Mamba2NetworkConfig

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl<'de> Deserialize<'de> for Mamba2NetworkConfig

fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>

where D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Display for Mamba2NetworkConfig

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Serialize for Mamba2NetworkConfig

fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>

where S: Serializer,

Serialize this value into the given Serde serializer.