Struct DeepseekV4CsaWeights 

pub struct DeepseekV4CsaWeights {
    pub role_proj: WeightMatrix,
    pub role_gate: WeightMatrix,
    pub indexer_key_proj: WeightMatrix,
    pub indexer_q_proj: WeightMatrix,
    pub indexer_head_weights: Vec<f32>,
    pub indexer_top_k: usize,
}

A CSA layer’s extras: the doubled role projection and the Lightning Indexer. Present only on a LayerCompressor::Csa layer, and its absence there is a configuration error rather than a default.

The doubled projection, which is the whole point of the split

On a CSA layer each raw token is projected twice: once for its role as the tail of the block ending at it, and once as the head of the next, overlapping block – two different learned projections of the same token, not one reused twice (llama.cpp load_arch_tensors’ coff = ratio == 4 ? 2 : 1, and build_overlap_compressed_kv_from_state’s GGML_ASSERT(kv_state->ne[0] == 2*n_embd_head)). A stack that applies one uniform ratio gives every CSA layer a single-width projection, which runs and produces numbers.

Fields

role_proj: WeightMatrix

qk_head_dim -> 2 * qk_head_dim: the head-role projection in the leading half, the tail-role projection in the trailing half.

role_gate: WeightMatrix

The same split for attn_comp_wgate’s scores, so every block row carries the gate belonging to the role it was projected for.

indexer_key_proj: WeightMatrix

One indexer key per compressed entry.

A skeleton simplification, named rather than hidden: upstream runs a separate compressor (indexer_comp_*) over the raw indexer projections, where this projects the already-compressed entry. The indexer keys are compressed representations either way, which is what the top-k selection needs; the second compressor’s own pooling is not reproduced here.

indexer_q_proj: WeightMatrix

The query side, qk_head_dim -> n_index_heads * index_head_dim.

indexer_head_weights: Vec<f32>

One weight per index head; its length is n_index_heads.

indexer_top_k: usize

How many compressed entries survive selection.