Mojo function

mla_prefill_decode_graph_bf16

mla_prefill_decode_graph_bf16[collection_t: KVCollectionT, //, mask_str: StringSlice[StaticConstantOrigin], kv_input_fn: def[width: Int](IndexList[2]) capturing -> SIMD[DType.bfloat16, width], target: StringSlice[StaticConstantOrigin] = StringSlice("cpu")](output: TileTensor[DType.bfloat16, linear_idx_type=output.linear_idx_type, element_size=output.element_size], q: TileTensor[DType.bfloat16, linear_idx_type=q.linear_idx_type, element_size=q.element_size], input_row_offsets: TileTensor[DType.uint32, linear_idx_type=input_row_offsets.linear_idx_type, element_size=input_row_offsets.element_size], freqs_cis: TileTensor[linear_idx_type=freqs_cis.linear_idx_type, element_size=freqs_cis.element_size], kv_norm_gamma: TileTensor[linear_idx_type=kv_norm_gamma.linear_idx_type, element_size=kv_norm_gamma.element_size], kv_collection: collection_t, layer_idx: UInt32, scale: Float32, epsilon: Float32, buffer_row_offsets: TileTensor[DType.uint32, linear_idx_type=buffer_row_offsets.linear_idx_type, element_size=buffer_row_offsets.element_size], cache_offsets: TileTensor[DType.uint32, linear_idx_type=cache_offsets.linear_idx_type, element_size=cache_offsets.element_size], buffer_length: Int, max_seq_len: Int, w_k: TileTensor[DType.bfloat16, linear_idx_type=w_k.linear_idx_type, element_size=w_k.element_size], w_uk: TileTensor[DType.bfloat16, linear_idx_type=w_uk.linear_idx_type, element_size=w_uk.element_size], w_uv: TileTensor[DType.bfloat16, linear_idx_type=w_uv.linear_idx_type, element_size=w_uv.element_size], scalar_args_buf: TileTensor[DType.int64, linear_idx_type=scalar_args_buf.linear_idx_type, element_size=scalar_args_buf.element_size], ctx: DeviceContext)

BF16 MLA prefill/decode graph.

Dispatches to prefill or decode based on max sequence length in the batch.