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,
}

GLM MLP geometry for one rank, read from the checkpoint config — never defaulted.

Fields

hidden: usize

local_dense_intermediate: usize

intermediate_size — the dense layers’ FFN width, this rank’s share.

moe_intermediate: usize

moe_intermediate_size — one routed expert’s width. Never TP-sharded: an expert is owned whole by one EP rank.

local_shared_intermediate: usize

n_shared_experts * moe_intermediate_size, this rank’s share.

num_experts: usize

FULL routed-expert count (288). Not the local count — see the module header.

local_experts: usize

Experts this rank owns, as a contiguous range [ep_rank * local, +local).

ep_rank: usize

top_k: usize

routed_scale: f32

routed_scaling_factor. 🪤 Applied to the top-k WEIGHTS, never to the shared expert.

renormalize: bool

norm_topk_prob — renormalise the top-k weights (with a 1e-20 epsilon, not 1e-6).

swiglu_limit: f32

🪤 Asymmetric. See the module header.

router_bf16_ladder: bool

True reproduces vLLM’s bf16 router ladder; false is HF 5.16.1’s fp32 semantics and the production default. Semantic, not precision — the two select different experts.

tp_world_size: usize

TP ranks the dense/shared FFN is split over. > 1 ⇒ the site output is a partial sum.

ep_world_size: usize

EP ranks the routed experts are split over. > 1 ⇒ the routed sum is a partial sum.

Implementations

impl Glm5NextMlpConfig

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

config carries GLOBAL MoE counts; the TP division of the dense/shared widths and the EP division of the expert set are applied here, in one place.

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

The half-open global expert-id range this rank owns.

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

Global expert id → local slot, or None when another rank owns it.

🪤 The whole EP scheme rests on this: a remote id must contribute zero, not be clamped into a local slot. Indexing a local array with a global id is the silent version of that mistake and yields a real expert’s weights for the wrong token.

pub fn needs_all_reduce(&self) -> bool

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