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

mha_decoding_single_batch_pipelined

def mha_decoding_single_batch_pipelined[q_type: DType, k_t: MHAOperand, v_t: MHAOperand, output_type: DType, mask_t: MHAMask, *, BM: Int, BN: Int, BK: Int, WM: Int, WN: Int, depth: Int, num_heads: Int, num_threads: Int, num_pipeline_stages: Int, group: Int = Int(1), decoding_warp_split_k: Bool = False, sink: Bool = False](q_ptr: Pointer[Scalar[q_type], ImmutAnyOrigin, _safe=False], k: k_t, v: v_t, output_ptr: Pointer[Scalar[output_type], MutAnyOrigin, _safe=False], exp_sum_ptr: Pointer[Scalar[get_accum_type[q_type]()], MutAnyOrigin, _safe=False], qk_max_ptr: Pointer[Scalar[get_accum_type[q_type]()], MutAnyOrigin, _safe=False], scale: Float32, num_keys: Int, num_partitions: Int, sink_weights: OptionalReg[LayoutTensor[q_type, Layout.row_major(Int(-1)), ImmutAnyOrigin]], mask: mask_t, batch_idx: Int)

Flash attention v2 decode kernel for a single batch element with pipelined multistage MMA.

Computes attention for the decoding (single-query) case using the FA2 online-softmax algorithm with multistage pipelining of K/V loads. When num_partitions exceeds 1, each block processes a contiguous slice of the key dimension and writes partial exp_sum and qk_max statistics for a subsequent mha_splitk_reduce pass.

Parameters:

q_type (DType): Element type of the query tensor (inferred).
k_t (MHAOperand): Key operand type backing the key tensor (inferred).
v_t (MHAOperand): Value operand type backing the value tensor (inferred).
output_type (DType): Element type of the output tensor (inferred).
mask_t (MHAMask): Attention mask type implementing MHAMask (inferred).
BM (Int): Number of query rows processed per thread block.
BN (Int): Number of key columns per thread block tile.
BK (Int): Tile size in the head-depth dimension.
WM (Int): Warp tile height in the query (M) dimension.
WN (Int): Warp tile width in the key (N) dimension.
depth (Int): Attention head depth (key/value dimension per head).
num_heads (Int): Total number of query heads.
num_threads (Int): Number of threads per thread block.
num_pipeline_stages (Int): Number of pipeline stages for the multistage MMA loads.
group (Int): GQA group size, query heads per key/value head (defaults to 1).
decoding_warp_split_k (Bool): Enable warp-level split-K for decode (defaults to False).
sink (Bool): Enable attention-sink mode where the first tokens always attend (defaults to False).

Args:

q_ptr (Pointer[Scalar[q_type], ImmutAnyOrigin, _safe=False]): Pointer to the query tensor for this batch element.
k (k_t): Key operand backed by a KV cache.
v (v_t): Value operand backed by a KV cache.
output_ptr (Pointer[Scalar[output_type], MutAnyOrigin, _safe=False]): Pointer to the output tensor for this batch element.
exp_sum_ptr (Pointer[Scalar[get_accum_type[q_type]()], MutAnyOrigin, _safe=False]): Pointer to the online-softmax exponent sum buffer for this batch.
qk_max_ptr (Pointer[Scalar[get_accum_type[q_type]()], MutAnyOrigin, _safe=False]): Pointer to the online-softmax running maximum buffer for this batch.
scale (Float32): Softmax temperature scale applied to Q·Kᵀ.
num_keys (Int): Total number of key/value entries (cache length) for this batch.
num_partitions (Int): Number of split-K partitions dividing the key dimension.
sink_weights (OptionalReg[LayoutTensor[q_type, Layout.row_major(Int(-1)), ImmutAnyOrigin]]): Optional sink-token weight tensor for attention sinks.
mask (mask_t): Mask instance used to apply the attention mask.
batch_idx (Int): Index of the batch element this block processes.