For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).
Mojo function
assert_sparse_row_packed
def assert_sparse_row_packed[sparse_row: SparseRowLayout](d_indices: OptionalReg[Pointer[Int32, MutAnyOrigin]], indices_stride: Int, q_token_idx: Int, visible: Int, selection_lag: Int, sparse_topk: Int)
Asserts that a token's sparse row is packed as sparse_row states.
Only runs with all assertions enabled. A producer that drops a causal key
makes the closed-form length overcount, and the softmax then scores the
trailing -1 slots as zero keys instead of masking them.
The message is a literal on purpose: formatting runtime values needs a
2 KB local buffer, and CUDA 13.1 ptxas segfaults compiling that callee
from a warp-specialized branch after setmaxnreg.
Parameters:
- sparse_row (
SparseRowLayout): The producer's row layout.
Args:
- d_indices (
OptionalReg[Pointer[Int32, MutAnyOrigin]]): The[total_q_tokens, indices_stride]sparse indices. - indices_stride (
Int): Row stride ofd_indices. - q_token_idx (
Int): Global index of the token whose row is checked. - visible (
Int): Number of keys the token sees. - selection_lag (
Int): Positions between the token the row was computed for and this one. - sparse_topk (
Int): The kernel's slot count for the sequence.