Struct NemotronSsmWeights 

pub struct NemotronSsmWeights {
    pub in_proj: QuantizedWeight,
    pub out_proj: QuantizedWeight,
    pub conv1d_weight: DenseWeight,
    pub conv1d_bias: DenseWeight,
    pub a_log: DenseWeight,
    pub d_param: DenseWeight,
    pub dt_bias: DenseWeight,
    pub ssm_norm: DenseWeight,
}

Nemotron-H Mamba-2 SSM weights.

in_proj produces [z(d_inner), x(d_inner), B(n_groupsstate), C(n_groupsstate), dt(num_heads)].

Fields

in_proj: QuantizedWeight

in_proj: [in_proj_size, hidden_size] NVFP4.

out_proj: QuantizedWeight

out_proj: [hidden_size, d_inner] NVFP4.

conv1d_weight: DenseWeight

conv1d weight: [d_xBC, 1, conv_kernel] BF16.

conv1d_bias: DenseWeight

conv1d bias: [d_xBC] BF16.

a_log: DenseWeight

A_log: [mamba_num_heads] BF16 (cast to FP32 at runtime).

d_param: DenseWeight

D skip-connection: [mamba_num_heads] BF16.

dt_bias: DenseWeight

dt_bias: [mamba_num_heads] BF16.

ssm_norm: DenseWeight

SSM internal norm: [d_inner] BF16 (applied to y before gating with z).