Enum ExpertQuantization 

pub enum ExpertQuantization {
    Fp8,
    Nvfp4GateUp,
    Nvfp4,
}

2026-09-27: The expert precision tiers. Measured on GB10 (Qwen3.6-35B-A3B-FP8, canonical tiers, --mtp-gate force):

tier	BFCL echolp N=1004 overall/normalized	agentic-webserver	C16 tok/s, J/tok (dgx1)
fp8	84.96/86.03	pass, 535-546 s	315.5, 0.227
nvfp4-gate-up	84.86/85.33	pass 3/3, 504-533 s	359.6, 0.211
nvfp4	85.46/86.57	FAIL: 168 turns, 774 s over the 700 s ceiling	383.1, 0.201

Weights resident before the KV cache: fp8 38.4 GB, nvfp4-gate-up 49.6 GB, nvfp4 55.2 GB.

Variants

Fp8

2026-09-27: The checkpoint’s FP8 experts: most stable, still fast.

Nvfp4GateUp

2026-09-27: Routed gate and up projections NVFP4 in decode; routed down, the shared expert and prefill FP8. The stable speed lever.

Nvfp4

2026-09-27: Every routed projection NVFP4 in decode; the shared expert and prefill FP8. Fastest, and it changes the agent’s trajectories most.

Implementations

impl ExpertQuantization

pub const ALL: [Self; 3]

2026-09-27: Every tier, in the order the flag lists them.

pub fn name(self) -> &'static str

2026-09-27: The tier’s flag value and recipe value.

pub fn nvfp4_decode(self) -> bool

2026-09-27: Whether MoE decode takes the grouped NVFP4 path.

pub fn nvfp4_down(self) -> bool

2026-09-27: Whether the routed down projections decode as NVFP4, so their NVFP4 copy is built at load; under nvfp4-gate-up they decode from the FP8 experts.

Trait Implementations

impl Clone for ExpertQuantization

fn clone(&self) -> ExpertQuantization

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for ExpertQuantization

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

Formats the value using the given formatter.

impl Default for ExpertQuantization

fn default() -> ExpertQuantization

Returns the “default value” for a type.

impl PartialEq for ExpertQuantization

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

impl Eq for ExpertQuantization

impl StructuralPartialEq for ExpertQuantization