pub struct MoeDescriptor {
pub num_tokens: i32,
pub num_experts: i32,
pub top_k: i32,
pub d_model: i32,
pub d_expert: i32,
pub variant: MoeVariant,
pub block_format: Option<GgufBlockFormat>,
pub element: ElementKind,
pub is_prefill: bool,
}Expand description
Descriptor for an MoE forward op.
Fields§
§num_tokens: i32Total number of tokens to process.
num_experts: i32Number of experts in the MoE block.
top_k: i32Number of experts each token is routed to (top_k in routing).
d_model: i32Hidden dim of the activation / output (size_k in Fuel-speak).
d_expert: i32Per-expert output feature dim (size_n in Fuel-speak).
variant: MoeVariantWhich kernel variant to dispatch.
block_format: Option<GgufBlockFormat>GGUF block format — must be Some(...) for ScalarGguf /
WmmaGguf variants and None for the Wmma variant.
element: ElementKindActivation element type. F32 for ScalarGguf; F16 or Bf16
for Wmma / WmmaGguf.
is_prefill: boolis_prefill flag for the Wmma variant (selects between
prefill M=16 / N=16 / WARPS_N=2 and decode M=8 / N=32 / WARPS_N=1
tile geometries). Ignored by the other variants.
Trait Implementations§
Source§impl Clone for MoeDescriptor
impl Clone for MoeDescriptor
Source§fn clone(&self) -> MoeDescriptor
fn clone(&self) -> MoeDescriptor
Returns a duplicate of the value. Read more
1.0.0 (const: unstable) · Source§fn clone_from(&mut self, source: &Self)
fn clone_from(&mut self, source: &Self)
Performs copy-assignment from
source. Read moreimpl Copy for MoeDescriptor
Auto Trait Implementations§
impl Freeze for MoeDescriptor
impl RefUnwindSafe for MoeDescriptor
impl Send for MoeDescriptor
impl Sync for MoeDescriptor
impl Unpin for MoeDescriptor
impl UnsafeUnpin for MoeDescriptor
impl UnwindSafe for MoeDescriptor
Blanket Implementations§
Source§impl<T> BorrowMut<T> for Twhere
T: ?Sized,
impl<T> BorrowMut<T> for Twhere
T: ?Sized,
Source§fn borrow_mut(&mut self) -> &mut T
fn borrow_mut(&mut self) -> &mut T
Mutably borrows from an owned value. Read more