Struct DeepseekV4DecoderConfig 

pub struct DeepseekV4DecoderConfig {
    pub rms_norm_eps: f32,
    pub hc_sinkhorn_iters: u32,
    pub hc_eps: f32,
    pub n_heads: usize,
    pub qk_head_dim: usize,
    pub v_head_dim: usize,
    pub qk_rope: usize,
    pub compress_rope_theta: f32,
    pub n_experts_active: usize,
    pub moe_renormalize: bool,
    pub compressor: LayerCompressor,
}

Fields

rms_norm_eps: f32

hc_sinkhorn_iters: u32

hc_eps: f32

n_heads: usize

qk_head_dim: usize

v_head_dim: usize

qk_rope: usize

compress_rope_theta: f32

n_experts_active: usize

moe_renormalize: bool

compressor: LayerCompressor

Which compressor this layer runs.

A mechanism, not a scalar ratio, and that is the correction: upstream ships a per-layer array (0, 0, 4, 128, 4, 128, 4, 0) where 0 means no compressor at all, 4 means CSA – overlapping blocks, a doubled projection width, and a Lightning Indexer – and 128 means HCA, non-overlapping and dense with no indexer. One uniform ratio across the stack builds an indexer on the HCA layers or none on the CSA ones, and gives half of them the wrong compressor width; all three failures run and produce numbers. LayerCompressor derives every one of those parameters from the mechanism, so they cannot be set inconsistently here.