Struct Glm5NextDsaConfig 

pub struct Glm5NextDsaConfig {

    pub hidden: usize,
    pub index_heads: usize,
    pub index_head_dim: usize,
    pub index_kpool: usize,
    pub index_topk: usize,
    pub always_select_tail: bool,
    pub local_heads: usize,
    pub q_lora_rank: usize,
    pub kv_lora_rank: usize,
    pub qk_nope_head_dim: usize,
    pub qk_rope_head_dim: usize,
    pub v_head_dim: usize,
    pub max_context: usize,
}

2026-09-25: DSA geometry for one layer, from ModelConfig.

Attention head counts are per rank (local_heads); the indexer is replicated, so index_heads is the full count. See tp.

Fields

hidden: usize

index_heads: usize

index_head_dim: usize

index_kpool: usize

index_topk: usize

always_select_tail: bool

local_heads: usize

2026-09-25: Attention heads this rank owns.

q_lora_rank: usize

kv_lora_rank: usize

qk_nope_head_dim: usize

qk_rope_head_dim: usize

2026-09-25: 0 on GLM-5.3 (NoPE). build::absorb_q and build::absorb_o refuse any other value.

v_head_dim: usize

max_context: usize

2026-09-25: Tokens a sequence’s indexer cache is sized for: --max-seq-len, or 16,384 when it is unset (from_config). state::max_dsa_context rounds it down to whole pools.

Implementations

impl Glm5NextDsaConfig

pub fn from_config(config: &ModelConfig) -> Result<Self>

2026-09-25: local_heads is config.num_attention_heads, which the serve’s serve_phases::topology has already divided by tp_size. Returns the result of validate.

pub fn qk_head_dim(&self) -> usize

pub fn is_nope(&self) -> bool

2026-09-25: True when there is no RoPE section, as on GLM-5.3.

pub fn select_k(&self, n_pools: usize) -> usize

2026-09-25: Pools selected per query, capped by how many pools exist.

pub fn out_width(&self) -> usize

2026-09-25: Width of the emitted index row: index_topk, plus index_kpool - 1 tail slots when always_select_tail.

pub fn kv_cache_dim(&self) -> usize

2026-09-25: KV latent cache width. Under NoPE this is kv_lora_rank, 512 on GLM-5.3; the DeepSeek-V4-Flash cache carries a rope tail and is 576 wide (MLA_CACHE_DIM in kernels/gb10/deepseek-v4-flash/nvfp4/mla_paged_decode.cu).

pub fn validate(&self) -> Result<()>

Trait Implementations

impl Clone for Glm5NextDsaConfig

fn clone(&self) -> Glm5NextDsaConfig

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for Glm5NextDsaConfig

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

Formats the value using the given formatter.

impl PartialEq for Glm5NextDsaConfig

fn eq(&self, other: &Glm5NextDsaConfig) -> bool

Tests for self and other values to be equal, and is used by ==.

fn ne(&self, other: &Rhs) -> bool

Tests for !=. The default implementation is almost always sufficient, and should not be overridden without very good reason.

impl Copy for Glm5NextDsaConfig

impl Eq for Glm5NextDsaConfig

impl StructuralPartialEq for Glm5NextDsaConfig