Enum Glm5NextRouterMode 

pub enum Glm5NextRouterMode {
    HfFp32,
    VllmBf16,
}

Which dtype ladder GLM-5.3’s MoE router runs in.

🔴 This is a SEMANTIC switch, not a precision preference. Slice 10 measured the two ladders selecting a different top-8 expert set on ~89–95 % of tokens (layers 3/23/44, T=2048), moving 20–26 % of routed weight mass onto experts the other ladder did not pick. Treating it as a harmless rounding choice is how a “faster router” silently becomes a different model.

Deliberately its OWN field, not derived from the quantization config or from PrecisionSchedule::router_dtype (which is a weight-STORAGE schedule with no compute meaning, and no consumers). Inferring a semantic from an unrelated knob is the defect this avoids.

GLM-scoped on purpose: no other Atlas model has a contested router ladder, and widening this into a cross-model routing refactor would be scope Atlas has not asked for.

Variants

HfFp32

CANONICAL / REFERENCE. HF transformers 5.16.1 semantics: F.linear(hidden.type(float32), weight.type(float32)), and sigmoid / correction bias / top-k / renormalisation all in fp32. This is the production default and must not change without review.

VllmBf16

COMPATIBILITY / ORACLE REPRODUCTION. Reproduces what vLLM currently does for glm5_next_text: GateLinear.out_dtype resolves to None (the fp32 special case in _get_moe_router_dtype fires only for glm_moe_dsa or an explicit moe_router_dtype), so the gate GEMM runs in the model dtype and grouped_topk does no upcast.

Exists so Atlas can reproduce the frozen vLLM oracle’s routing for A/B work. Never a production default.

Implementations

impl Glm5NextRouterMode

pub fn from_config_str(s: &str) -> Option<Self>

Parse the moe_router_dtype config field — the same name vLLM reads.

Absent ⇒ Self::HfFp32. That is the opposite of vLLM’s fallthrough, and deliberately so: absent means “the checkpoint did not say”, and the reference implementation’s answer for that case is fp32.

pub fn is_fp32(self) -> bool

True when router math must be carried in fp32.

Trait Implementations

impl Clone for Glm5NextRouterMode

fn clone(&self) -> Glm5NextRouterMode

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for Glm5NextRouterMode

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

Formats the value using the given formatter.

impl Default for Glm5NextRouterMode

fn default() -> Glm5NextRouterMode

Returns the “default value” for a type.

impl<'de> Deserialize<'de> for Glm5NextRouterMode

fn deserialize<__D>(__deserializer: __D) -> Result<Self, __D::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl PartialEq for Glm5NextRouterMode

fn eq(&self, other: &Glm5NextRouterMode) -> 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 Glm5NextRouterMode

impl Eq for Glm5NextRouterMode

impl StructuralPartialEq for Glm5NextRouterMode