pub struct ChunkedPrefillPageMetadata {
pub block_table: DevicePtr,
pub seq_len: DevicePtr,
pub block_capacity: usize,
pub uploaded_blocks: usize,
}Expand description
Per-sequence paged attention metadata for chunked prefill.
Positions and slots remain chunk-local, but the paged block table and running sequence length can persist across chunks so we only upload the changed tail instead of rebuilding the full page metadata every time.
Fields§
§block_table: DevicePtrDevice buffer holding the sequence block table as raw 32-bit entries.
seq_len: DevicePtrDevice buffer holding the running paged-prefill sequence length.
block_capacity: usizeTotal block-table entries allocated for this prompt.
uploaded_blocks: usizeNumber of block-table entries already uploaded to block_table.
Auto Trait Implementations§
impl Freeze for ChunkedPrefillPageMetadata
impl RefUnwindSafe for ChunkedPrefillPageMetadata
impl Send for ChunkedPrefillPageMetadata
impl Sync for ChunkedPrefillPageMetadata
impl Unpin for ChunkedPrefillPageMetadata
impl UnwindSafe for ChunkedPrefillPageMetadata
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