Struct BlockPagedLayout 

#[non_exhaustive]
pub struct BlockPagedLayout {
    pub page_size: u32,
    pub num_pages: u64,
    pub paged_axis: u32,
    pub num_seqs: u32,
    pub kv_role: KvRole,
    pub layer_index: Option<u32>,
    pub block_table_index_type: BlockTableIndexType,
}

Descriptor for the block-paged indirect layout.

Block-paged stores a KV-cache tensor whose paged axis is divided into fixed-size pages drawn from a shared page pool, with a block table mapping each logical page position to a physical page ID. It is the interchange form of a PagedAttention KV cache.

This layout is defined only for rank-3 tensors in this version of the specification: [total_tokens, num_heads, head_dim].

Three buffers are always required:

Buffer	Name	Description
0	page_pool	Flat pool of fixed-size pages.
1	block_table	Concatenated per-sequence page-ID lists.
2	seq_ptr	CSR-style offset array into block_table.

When the tensor carries quantization parameters, additional buffers appear at indices 3 and above per docs/spec/quantization.md § Buffer Table Placement Rules.

See docs/spec/layouts/block-paged.md for the full normative definition.

Examples

use hurray_core::layout::{BlockPagedLayout, BlockTableIndexType, KvRole, LayoutDescriptor};

let layout = BlockPagedLayout::new(16, 128, 0, 4, KvRole::Key, Some(3), BlockTableIndexType::U32);
assert_eq!(layout.page_size, 16);
assert_eq!(layout.num_pages, 128);
assert_eq!(layout.num_seqs, 4);
assert_eq!(layout.layer_index, Some(3));

let desc = LayoutDescriptor::BlockPaged(layout);
assert_eq!(desc.tag(), 0x0A);
assert_eq!(desc.buffer_count().map(|n| n.get()), Some(3));

Fields (Non-exhaustive)

Non-exhaustive structs could have additional fields added in future. Therefore, non-exhaustive structs cannot be constructed in external crates using the traditional Struct { .. } syntax; cannot be matched against without a wildcard ..; and struct update syntax will not work.

page_size: u32

Tokens per page. MUST be >= 1.

Common values are 16 and 32 (matching vLLM’s default block_size).

num_pages: u64

Number of physical pages in page_pool.

paged_axis: u32

The axis subdivided into pages. MUST be 0 in this spec version.

Kept as a field rather than a constant so that the wire format remains forward-compatible if a future spec version adds non-zero paged axes.

num_seqs: u32

Number of sequences in the batch. MAY be 0 for an empty batch.

kv_role: KvRole

KV-cache role for this tensor (key, value, or fused/generic).

layer_index: Option<u32>

Transformer layer index, or None if the tensor is not layer-scoped.

Wire value 0xFFFFFFFF maps to None; all other values map to Some(n).

block_table_index_type: BlockTableIndexType

Element type for both the block_table (buffer 1) and seq_ptr (buffer 2).

U32 is the default and addresses up to 4 billion pages. U64 is required for larger pools.

Implementations

impl BlockPagedLayout

pub fn new( page_size: u32, num_pages: u64, paged_axis: u32, num_seqs: u32, kv_role: KvRole, layer_index: Option<u32>, block_table_index_type: BlockTableIndexType, ) -> Self

Creates a new BlockPagedLayout.

Arguments

• page_size — tokens per page (MUST be validated >= 1 later via validate_against_shape).
• num_pages — number of physical pages in the pool.
• paged_axis — the axis subdivided into pages (MUST be 0 in this version).
• num_seqs — number of sequences in the batch.
• kv_role — KV-cache role for this tensor.
• layer_index — transformer layer index (None = not layer-scoped).
• block_table_index_type — element type for block_table and seq_ptr buffers.

This constructor does not validate page_size >= 1 or paged_axis == 0 because the tensor shape is not available at construction time. Call [LayoutDescriptor::validate_against_shape] to perform all invariant checks once the shape is known.

Examples

use hurray_core::layout::{BlockPagedLayout, BlockTableIndexType, KvRole};

// Layer-3 key cache: page_size=16, 64 pages, 2 sequences, uint32 indices.
let layout = BlockPagedLayout::new(
    16, 64, 0, 2,
    KvRole::Key,
    Some(3),
    BlockTableIndexType::U32,
);
assert_eq!(layout.page_size, 16);
assert_eq!(layout.layer_index, Some(3));

pub fn validate_quantization_compatibility( &self, scheme_tag: u8, quant_axis: u32, quant_block_size: u32, ) -> Result<()>

Validates that a quantization scheme is compatible with this block-paged layout.

Per docs/spec/layouts/block-paged.md § Quantization Compatibility:

• Per-tensor (scheme_tag = 0x01): always compatible.
• Per-channel (scheme_tag = 0x02): MUST NOT be applied on axis 0 (the paged / token axis). quant_axis must be 1 (num_heads) or 2 (head_dim).
• Per-block-affine (scheme_tag = 0x03): quant_axis MUST be 0 and quant_block_size MUST equal page_size.
• All other schemes: not validated here; returns Ok(()).

Call this at the typed-quantization layer once you have a decoded crate::QuantizationDescriptor. The raw-bytes descriptor layer cannot perform this check.

Arguments

• scheme_tag — the scheme_tag byte from the quantization descriptor header.
• quant_axis — the axis field from the quantization descriptor (relevant for per-channel and per-block-affine).
• quant_block_size — the block_size field from the per-block-affine descriptor (ignored for all other schemes).

Errors

Returns crate::Error::InvalidQuantization if the scheme/layout combination violates the spec rules above.

Examples

use hurray_core::layout::{BlockPagedLayout, BlockTableIndexType, KvRole};

let layout = BlockPagedLayout::new(16, 64, 0, 2, KvRole::Key, Some(0), BlockTableIndexType::U32);

// Per-tensor (0x01) is always valid.
assert!(layout.validate_quantization_compatibility(0x01, 0, 0).is_ok());

// Per-channel on axis 1 (num_heads) is valid.
assert!(layout.validate_quantization_compatibility(0x02, 1, 0).is_ok());

// Per-channel on axis 0 (paged axis) is forbidden.
assert!(layout.validate_quantization_compatibility(0x02, 0, 0).is_err());

// Per-block-affine with axis=0 and block_size==page_size is valid.
assert!(layout.validate_quantization_compatibility(0x03, 0, 16).is_ok());

// Per-block-affine with block_size != page_size is forbidden.
assert!(layout.validate_quantization_compatibility(0x03, 0, 32).is_err());

Trait Implementations

impl Clone for BlockPagedLayout

fn clone(&self) -> BlockPagedLayout

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for BlockPagedLayout

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

Formats the value using the given formatter.

impl Eq for BlockPagedLayout

impl Hash for BlockPagedLayout

fn hash<__H: Hasher>(&self, state: &mut __H)

Feeds this value into the given Hasher.

fn hash_slice<H>(data: &[Self], state: &mut H)

where H: Hasher, Self: Sized,

Feeds a slice of this type into the given Hasher.

impl PartialEq for BlockPagedLayout

fn eq(&self, other: &BlockPagedLayout) -> bool

Equality operator ==.

fn ne(&self, other: &Rhs) -> bool

Inequality operator !=.

impl StructuralPartialEq for BlockPagedLayout