IMPORTANT: To view this page as Markdown, append `.md` to the URL (e.g. /get-started.md). For the complete documentation index, see llms.txt.
Skip to main content
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

ssd_chunk_state

def ssd_chunk_state[kernel_dtype: DType, SLT: TensorLayout, SEngine: TensorEngine, XFn: def[width: Int, alignment: Int](Int, Int, Int) -> SIMD[kernel_dtype, width] & RegisterPassable & ImplicitlyCopyable, BFn: def[width: Int, alignment: Int](Int, Int, Int) -> SIMD[kernel_dtype, width] & RegisterPassable & ImplicitlyCopyable, DSTATE: Int](slots_i: Int32, nheads_i: Int32, head_dim_i: Int32, ngroups_i: Int32, info: TileTensor[.int32, SLT, MutUntrackedOrigin, Engine=SEngine], cs: TileTensor[.float32, SLT, MutUntrackedOrigin, Engine=SEngine], delta_in: TileTensor[.float32, SLT, MutUntrackedOrigin, Engine=SEngine], states: TileTensor[.float32, SLT, MutUntrackedOrigin, Engine=SEngine], x_fn: XFn, b_fn: BFn)

Computes the state each chunk contributes by itself, on tensor cores.

states[slot, h, p, n] = sum_t exp2(cs_last - cs_t) dt_t x[t, p] B[t, n]. Grid (slots, nheads, head_dim / TILE), block OUT_THREADS (4 warps, 16 rows of p each). The reduction runs over t in tiles of 64 tokens that stay [t][p] and [t][n] in shared memory: ldmatrix.trans turns them into the MMA fragments.

Was this page helpful?