Struct Glm5NextMlpConfig 

pub struct Glm5NextMlpConfig {

    pub hidden: usize,
    pub local_dense_intermediate: usize,
    pub moe_intermediate: usize,
    pub local_shared_intermediate: usize,
    pub num_experts: usize,
    pub local_experts: usize,
    pub ep_rank: usize,
    pub top_k: usize,
    pub routed_scale: f32,
    pub renormalize: bool,
    pub swiglu_limit: f32,
    pub router_bf16_ladder: bool,
    pub tp_world_size: usize,
    pub ep_world_size: usize,
}

2026-09-25: GLM MLP geometry for one rank, built by Self::from_config.

Fields

hidden: usize

local_dense_intermediate: usize

2026-09-25: intermediate_size / tp_world_size: this rank’s share of the dense FFN width.

moe_intermediate: usize

2026-09-25: moe_intermediate_size, one routed expert’s width. Not divided by TP: an expert is owned whole by one EP rank.

local_shared_intermediate: usize

2026-09-25: shared_expert_intermediate_size / tp_world_size: this rank’s share of the shared expert.

num_experts: usize

2026-09-25: The full routed-expert count, not this rank’s share.

local_experts: usize

2026-09-25: num_experts / ep_world_size; this rank owns ids [ep_rank * local_experts, (ep_rank + 1) * local_experts).

ep_rank: usize

top_k: usize

routed_scale: f32

2026-09-25: routed_scaling_factor. Applied to the top-k weights, not to the shared expert.

renormalize: bool

2026-09-25: norm_topk_prob: divide the top-k weights by their sum plus 1e-20.

swiglu_limit: f32

2026-09-25: The asymmetric SwiGLU clamp bound; see the module header.

router_bf16_ladder: bool

2026-09-25: True for Glm5NextRouterMode::VllmBf16: the router kernel then rounds the scores, the running sum and each weight to BF16, which can change which experts are selected. The parser yields HfFp32 (false) when the checkpoint names no router dtype.

tp_world_size: usize

2026-09-25: TP ranks the dense/shared FFN is split over. Above 1, the site output is a partial sum.

ep_world_size: usize

2026-09-25: EP ranks the routed experts are split over. Above 1, the routed sum is a partial sum.

Implementations

impl Glm5NextMlpConfig

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

2026-09-25: Divides the global widths by TP and the expert set by EP (a world size of 0 counts as 1), then runs Self::validate. Errors when a width or the expert count does not divide evenly.

pub fn local_expert_range(&self) -> Range<usize>

2026-09-25: The half-open global expert-id range this rank owns.

pub fn local_slot(&self, global_id: usize) -> Option<usize>

2026-09-25: Global expert id to local slot, or None when another rank owns it. A remote id contributes zero; experts is indexed by this slot, never by the global id.

pub fn needs_all_reduce(&self) -> bool

2026-09-25: Whether the site output leaves this rank as a partial sum needing all_reduce(SUM).

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

Trait Implementations

impl Clone for Glm5NextMlpConfig

fn clone(&self) -> Glm5NextMlpConfig

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for Glm5NextMlpConfig

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

Formats the value using the given formatter.

impl PartialEq for Glm5NextMlpConfig

fn eq(&self, other: &Glm5NextMlpConfig) -> 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 Glm5NextMlpConfig

impl StructuralPartialEq for Glm5NextMlpConfig