Mojo function

mla_prefill_decode_graph_fp8

mla_prefill_decode_graph_fp8[dtype: DType, fp8_dtype: DType, fp8_scale_dtype: DType, collection_t: KVCollectionT, //, m_scale_granularity: Int, n_scale_granularity: Int, k_scale_granularity: Int, mask_str: StringSlice[StaticConstantOrigin], score_mod_str: StringSlice[StaticConstantOrigin], target: StringSlice[StaticConstantOrigin] = "cpu"](output: TileTensor[dtype, output.LayoutType, output.origin, linear_idx_type=output.linear_idx_type, element_shape_types=output.element_shape_types], q: TileTensor[dtype, q.LayoutType, q.origin, linear_idx_type=q.linear_idx_type, element_shape_types=q.element_shape_types], input_row_offsets: TileTensor[DType.uint32, input_row_offsets.LayoutType, input_row_offsets.origin, linear_idx_type=input_row_offsets.linear_idx_type, element_shape_types=input_row_offsets.element_shape_types], freqs_cis: TileTensor[freqs_cis.dtype, freqs_cis.LayoutType, freqs_cis.origin, linear_idx_type=freqs_cis.linear_idx_type, element_shape_types=freqs_cis.element_shape_types], kv_norm_gamma: TileTensor[kv_norm_gamma.dtype, kv_norm_gamma.LayoutType, kv_norm_gamma.origin, linear_idx_type=kv_norm_gamma.linear_idx_type, element_shape_types=kv_norm_gamma.element_shape_types], kv_collection: collection_t, layer_idx: UInt32, scale: Float32, epsilon: Float32, buffer_row_offsets: TileTensor[DType.uint32, buffer_row_offsets.LayoutType, buffer_row_offsets.origin, linear_idx_type=buffer_row_offsets.linear_idx_type, element_shape_types=buffer_row_offsets.element_shape_types], cache_offsets: TileTensor[DType.uint32, cache_offsets.LayoutType, cache_offsets.origin, linear_idx_type=cache_offsets.linear_idx_type, element_shape_types=cache_offsets.element_shape_types], buffer_length: Int, max_seq_len: Int, w_k: TileTensor[fp8_dtype, w_k.LayoutType, w_k.origin, linear_idx_type=w_k.linear_idx_type, element_shape_types=w_k.element_shape_types], w_k_scale: TileTensor[fp8_scale_dtype, w_k_scale.LayoutType, w_k_scale.origin, linear_idx_type=w_k_scale.linear_idx_type, element_shape_types=w_k_scale.element_shape_types], w_uk: TileTensor[fp8_dtype, w_uk.LayoutType, w_uk.origin, linear_idx_type=w_uk.linear_idx_type, element_shape_types=w_uk.element_shape_types], w_uk_scale: TileTensor[fp8_scale_dtype, w_uk_scale.LayoutType, w_uk_scale.origin, linear_idx_type=w_uk_scale.linear_idx_type, element_shape_types=w_uk_scale.element_shape_types], w_uv: TileTensor[fp8_dtype, w_uv.LayoutType, w_uv.origin, linear_idx_type=w_uv.linear_idx_type, element_shape_types=w_uv.element_shape_types], w_uv_scale: TileTensor[fp8_scale_dtype, w_uv_scale.LayoutType, w_uv_scale.origin, linear_idx_type=w_uv_scale.linear_idx_type, element_shape_types=w_uv_scale.element_shape_types], ctx: DeviceContext)

This is a manually fused kernel that performs the following operations: - Perform MLA prefill or decode based on the maximum sequence length.