pub struct DsaDecodeInputs {
pub q: DevicePtr,
pub k_cache: DevicePtr,
pub v_cache: DevicePtr,
pub out: DevicePtr,
pub block_tables: DevicePtr,
pub seq_lens: DevicePtr,
pub sel_indices: DevicePtr,
pub k_scale: f32,
pub v_scale: f32,
}Expand description
Everything the decode reads, all caller-owned.
Fields§
§q: DevicePtr[num_q_heads * kv_lora_rank] BF16 — this rank’s absorbed queries.
k_cache: DevicePtrFP8 paged latent cache. In absorbed NoPE MLA K and V are the same buffer; both are taken so a caller that splits them is not forced to lie.
v_cache: DevicePtr§out: DevicePtr[num_q_heads * kv_lora_rank] BF16 output.
block_tables: DevicePtr[num_seqs, max_blocks_per_seq] i32.
seq_lens: DevicePtr[num_seqs] i32.
sel_indices: DevicePtr[num_seqs, sel_width] i32 — super::select::DsaSelectScratch::tokens.
k_scale: f32§v_scale: f32Trait Implementations§
Source§impl Clone for DsaDecodeInputs
impl Clone for DsaDecodeInputs
Source§fn clone(&self) -> DsaDecodeInputs
fn clone(&self) -> DsaDecodeInputs
Returns a duplicate of the value. Read more
1.0.0 · Source§fn clone_from(&mut self, source: &Self)
fn clone_from(&mut self, source: &Self)
Performs copy-assignment from
source. Read moreSource§impl Debug for DsaDecodeInputs
impl Debug for DsaDecodeInputs
impl Copy for DsaDecodeInputs
Auto Trait Implementations§
impl Freeze for DsaDecodeInputs
impl RefUnwindSafe for DsaDecodeInputs
impl Send for DsaDecodeInputs
impl Sync for DsaDecodeInputs
impl Unpin for DsaDecodeInputs
impl UnwindSafe for DsaDecodeInputs
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