For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /max/get-started.md).
Mojo struct
VBufferConfig
struct VBufferConfig[BN: Int, BK: Int, WN: Int, depth: Int]
V-role decode buffer config that iterates DRAM along the row axis.
The warp tile is [BK, depth // num_warps_n] (strip width by per-warp
depth slice), the block tile is [BK, depth], and iterator_axis = 0
advances tile_idx across row strips of the DRAM tile.
Parameters
- BN (
Int): Number of keys per block tile (extent along the key axis). - BK (
Int): Strip width along the MMA K dimension. - WN (
Int): Number of keys per warp tile. - depth (
Int): Per-head depth of the KV cache (head dimension).
Implemented traits
AnyType,
ImplicitlyDeletable,
KVBufferConfig
comptime members
btile_dim0
comptime btile_dim0 = BK
btile_dim1
comptime btile_dim1 = depth
iterator_axis
comptime iterator_axis = 0
wsize
comptime wsize = VBufferConfig[BN, BK, WN, depth].wtile_dim1
wtile_dim0
comptime wtile_dim0 = BK
wtile_dim1
comptime wtile_dim1 = (depth // (BN // WN))
Methods
get_wtile_coord
Was this page helpful?
Thank you! We'll create more content like this.
Thank you for helping us improve!