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 function
lora_qkv_plane_row_offset
def lora_qkv_plane_row_offset[splits: IndexList[Int(2)]](out_col: Int, plane_stride: Int) -> Int
Additive row offset into a 3-plane row-stacked activation [3*M, K].
Lets a grouped matmul read a different plane of the activation per
output-column region within a single launch, without a capturing closure
(which the warp-specialized kernel's GPU slicer cannot handle in the load
path). Used by the LoRA-B QKV expand to select the matching plane of the
planar shrink output P [3, M, R] for the Q / K / V output regions.
For output column out_col, the plane is 0 when out_col < splits[0], 1
when out_col < splits[1], else 2; the returned offset is
plane * plane_stride. splits == (0, 0) disables it (returns 0), so the
default is a no-op for every other caller.
Returns:
Was this page helpful?
Thank you! We'll create more content like this.
Thank you for helping us improve!