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

mha_decoding_single_batch

def mha_decoding_single_batch[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], k: k_t, v: v_t, output_ptr: Pointer[Scalar[output_type], MutAnyOrigin], exp_sum_ptr: Pointer[Scalar[get_accum_type[q_type]()], MutAnyOrigin], qk_max_ptr: Pointer[Scalar[get_accum_type[q_type]()], MutAnyOrigin], scale: Float32, num_keys: Int, num_partitions: Int, mask: mask_t, batch_idx: Int, sink_weights: OptionalReg[LayoutTensor[q_type, Layout.row_major(Int(-1)), ImmutAnyOrigin]])

Flash attention v2 algorithm.

Parameters:

• ​q_type (DType): Element type of the query tensor.
• ​k_t (MHAOperand): Key operand type (KV cache or dense tensor).
• ​v_t (MHAOperand): Value operand type (KV cache or dense tensor).
• ​output_type (DType): Element type of the output tensor.
• ​mask_t (MHAMask): Attention mask type implementing MHAMask.
• ​BM (Int): Number of query rows per thread block.
• ​BN (Int): Number of key columns per thread block.
• ​BK (Int): Tile size in the depth dimension for shared-memory tiles.
• ​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 software-pipeline stages for async copies.
• ​group (Int): GQA group size, query heads per key/value head (defaults to 1).
• ​decoding_warp_split_k (Bool): Enable warp-level split-K reduction (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]): Pointer to the query tensor in global memory.
• ​k (k_t): Key operand backed by a KV cache or dense tensor.
• ​v (v_t): Value operand backed by a KV cache or dense tensor.
• ​output_ptr (Pointer[Scalar[output_type], MutAnyOrigin]): Pointer to the output tensor in global memory.
• ​exp_sum_ptr (Pointer[Scalar[get_accum_type[q_type]()], MutAnyOrigin]): Pointer to the per-head online-softmax denominator (sum of exponentials) for cross-partition reduction.
• ​qk_max_ptr (Pointer[Scalar[get_accum_type[q_type]()], MutAnyOrigin]): Pointer to the per-head online-softmax running maximum for cross-partition reduction.
• ​scale (Float32): Softmax temperature scale applied to Q·Kᵀ.
• ​num_keys (Int): Number of valid key/value entries (cache length).
• ​num_partitions (Int): Number of split-K partitions along the key dimension.
• ​mask (mask_t): Mask instance used to apply the attention mask.
• ​batch_idx (Int): Index of the sequence within the batch.
• ​sink_weights (OptionalReg[LayoutTensor[q_type, Layout.row_major(Int(-1)), ImmutAnyOrigin]]): Optional sink-token weight tensor for attention sinks.