FastDeploy/fastdeploy/model_executor/layers/attention/append_attn_backend.py

"""
# Copyright (c) 2025 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""

from __future__ import annotations

import os
from dataclasses import dataclass, field
from typing import TYPE_CHECKING, List, Optional

import paddle

from fastdeploy.model_executor.layers.attention.ops import (
    append_attention,
    append_attention_with_output,
    get_block_shape_and_split_kv_block,
    init_kv_signal_per_query,
    init_signal_layerwise,
    open_shm_and_get_meta_signal,
)

if TYPE_CHECKING:
    from fastdeploy.model_executor.forward_meta import ForwardMeta

import numpy as np

from fastdeploy.config import FDConfig
from fastdeploy.model_executor.layers.attention.attention import Attention
from fastdeploy.model_executor.layers.attention.base_attention_backend import (
    AttentionBackend,
    AttentionMetadata,
)
from fastdeploy.model_executor.layers.attention.utils import init_rank_and_device_id
from fastdeploy.platforms import current_platform


@dataclass
class AppendAttentionMetadata(AttentionMetadata):
    """
    AppendAttentionMetadata
    """

    _dtype: paddle.dtype = paddle.bfloat16
    encoder_max_partition_size: int = 32768
    max_partition_size: int = 32768
    _fuse_kernel_compute_dtype: str = "bf16"

    # pd_disaggregation
    kv_signal_metadata: Optional[paddle.Tensor] = None
    kv_signal_data_list: List[Optional[paddle.Tensor]] = field(default_factory=list)


def allocate_launch_related_buffer(
    max_batch_size,
    max_model_len,
    encoder_block_shape_q,
    decoder_block_shape_q,
    decoder_step_token_num,
    num_heads,
    kv_num_heads,
    block_size,
):
    # Initialize AttentionBackend buffers
    assert num_heads % kv_num_heads == 0
    assert max_model_len % block_size == 0
    assert max_model_len % encoder_block_shape_q == 0
    group_size = num_heads // kv_num_heads

    # NOTE: (changwenbin) When using auto_chunk,
    # decode_max_tile_size must take into account the maximum case, where *1024 can cover 128K.
    decode_max_tile_size = (
        1024 * max_batch_size * (int)(np.ceil(decoder_step_token_num * group_size / decoder_block_shape_q))
    )
    encode_max_tile_size = max_batch_size * (max_model_len * group_size // encoder_block_shape_q)
    kv_max_tile_size = max_batch_size * (max_model_len // block_size)
    res = {}
    res["decoder_batch_ids"] = paddle.full([decode_max_tile_size], 0, dtype="int32")
    res["decoder_tile_ids_per_batch"] = paddle.full([decode_max_tile_size], 0, dtype="int32")
    if current_platform.is_maca():
        res["decoder_num_blocks_cpu"] = paddle.full([1], 0, dtype="int32").cpu()
    else:
        res["decoder_num_blocks_cpu"] = paddle.full([1], 0, dtype="int32").pin_memory()
    # NOTE: (changwenbin) MLA kernel only needs decoder_num_blocks_device in place of GPU tensor,
    # adapted to cudagraph.
    res["decoder_num_blocks_device"] = paddle.full([1], 0, dtype="int32")
    res["decoder_chunk_size_device"] = paddle.full([1], 64, dtype="int32")
    res["max_len_tensor_cpu"] = paddle.full([9], 0, dtype="int32").cpu()

    res["encoder_batch_ids"] = paddle.full([encode_max_tile_size], 0, dtype="int32")
    res["encoder_tile_ids_per_batch"] = paddle.full([encode_max_tile_size], 0, dtype="int32")
    res["encoder_num_blocks_x_cpu"] = paddle.full([1], 0, dtype="int32").cpu()

    res["kv_batch_ids"] = paddle.full([kv_max_tile_size], 0, dtype="int32")
    res["kv_tile_ids_per_batch"] = paddle.full([kv_max_tile_size], 0, dtype="int32")
    res["kv_num_blocks_x_cpu"] = paddle.full([1], 0, dtype="int32").cpu()
    return res


class AppendAttentionBackend(AttentionBackend):
    """
    AppendAttentionBackend backend implementation.
    """

    __infer_dynamic_dims_fields__ = ["attention_metadata"]
    attention_metadata: AppendAttentionMetadata

    def __init__(
        self,
        fd_config: FDConfig,
        kv_num_heads: int,
        num_heads: int,
        head_dim: int,
        encoder_block_shape_q: int = -1,
        decoder_block_shape_q: int = -1,
    ) -> None:
        """
        AppendAttentionBackend __init__
        """
        super().__init__()
        self.attention_metadata: AppendAttentionMetadata = None
        self.block_size: int = fd_config.cache_config.block_size
        self.max_seq_len: int = fd_config.model_config.max_model_len
        self.rope_theta: float = (
            10000.0 if fd_config.model_config.rope_theta is None else fd_config.model_config.rope_theta
        )
        self.rope_3d: bool = getattr(fd_config.model_config, "rope_3d", False) or getattr(
            fd_config.model_config, "use_3d_rope", False
        )
        if fd_config.speculative_config.model_type != "main":
            self.rope_3d = False
        self.causal: bool = getattr(fd_config.model_config, "causal", True)
        self.speculative_method: str = fd_config.speculative_config.method
        self.speculate_max_draft_token_num: int = fd_config.speculative_config.num_speculative_tokens
        self.keep_pd_step_flag: bool = fd_config.speculative_config.model_type == "mtp"
        self.num_layers_draft_model: int = int(fd_config.speculative_config.method in ["mtp"])

        self.kv_num_heads: int = kv_num_heads
        self.num_heads: int = num_heads
        self.group_size: int = self.num_heads // self.kv_num_heads
        self.head_dim: int = fd_config.model_config.head_dim
        self.num_layers: int = fd_config.model_config.num_hidden_layers
        self.max_partition_size: int = int(os.getenv("FLAGS_max_partition_size", 1024))
        # split kv still has bug in speculative decoding
        if self.speculative_method is not None:
            self.max_partition_size = self.max_seq_len
        self.encoder_block_shape_q: int = encoder_block_shape_q
        self.decoder_block_shape_q: int = decoder_block_shape_q

        self.pd_disaggregation_mode: str = fd_config.parallel_config.pd_disaggregation_mode

        self.start_layer_index: int = fd_config.model_config.start_layer_index

        if fd_config.parallel_config.expert_parallel_rank is None:
            fd_config.parallel_config.expert_parallel_rank = 0

        self.rank, self.device_id = init_rank_and_device_id(fd_config)
        self.use_output = not fd_config.graph_opt_config.full_cuda_graph
        self.fd_config = fd_config

    def init_attention_metadata(self, forward_meta: ForwardMeta):
        """Initialize attntion metadata hence all layers in the forward pass can reuse it."""
        metadata = AppendAttentionMetadata()
        metadata.max_partition_size = self.max_partition_size
        metadata.encoder_max_partition_size = self.max_seq_len
        metadata._dtype = paddle.get_default_dtype()
        if metadata._dtype == "bfloat16":
            metadata._fuse_kernel_compute_dtype = "bf16"
        elif metadata._dtype == "float16":
            metadata._fuse_kernel_compute_dtype = "fp16"
        elif metadata._dtype == "float32":
            metadata._fuse_kernel_compute_dtype = "fp32"

        # pd_disaggregation
        metadata.kv_signal_data_list = [None] * self.num_layers
        if self.pd_disaggregation_mode == "per_chunk":
            if not self.keep_pd_step_flag and not forward_meta.is_dummy_or_profile_run:
                init_kv_signal_per_query(
                    forward_meta.seq_lens_encoder,
                    forward_meta.seq_lens_this_time,
                    forward_meta.seq_lens_decoder,
                    self.rank,
                    self.num_layers + self.num_layers_draft_model,
                )
        elif self.pd_disaggregation_mode == "per_query":
            metadata.kv_signal_metadata = open_shm_and_get_meta_signal(
                self.rank, int(self.device_id), self.keep_pd_step_flag
            )

        self.attention_metadata: AttentionMetadata = metadata

    def get_attntion_meta(self) -> AttentionMetadata:
        """get_attntion_meta"""
        return self.attention_metadata

    def get_kv_cache_shape(
        self,
        max_num_blocks: int,
        kv_cache_quant_type: str = None,
    ):
        """
        Calculate kv cache shape
        """
        key_cache_shape = [max_num_blocks, self.kv_num_heads, self.block_size, self.head_dim]
        if kv_cache_quant_type is not None and kv_cache_quant_type == "int4_zp":
            key_cache_shape[-1] = self.head_dim // 2
        value_cache_shape = key_cache_shape
        return key_cache_shape, value_cache_shape

    def forward_mixed(
        self,
        q: paddle.Tensor,
        k: paddle.Tensor,
        v: paddle.Tensor,
        qkv: paddle.Tensor,
        compressed_kv: paddle.Tensor,
        k_pe: paddle.Tensor,
        layer: Attention,
        forward_meta: ForwardMeta,
    ) -> paddle.Tensor:
        """
        forward_mixed
        """
        metadata = self.attention_metadata
        sliding_window = layer.sliding_window

        if self.rope_3d:
            assert len(forward_meta.rotary_embs.shape) == 6
        else:
            assert len(forward_meta.rotary_embs.shape) == 5
            if layer.use_neox_rotary_style:
                assert forward_meta.rotary_embs.shape[0:4] == [2, 1, self.max_seq_len, 1]
                # 128 is qwen3
                # 32 is glm
                assert forward_meta.rotary_embs.shape[4] in [128, 32]
            else:
                assert forward_meta.rotary_embs.shape == [2, 1, self.max_seq_len, 1, 64]

        if self.pd_disaggregation_mode == "per_query":
            metadata.kv_signal_data_list[layer.layer_id] = init_signal_layerwise(
                metadata.kv_signal_metadata,
                layer.layer_id + self.start_layer_index,
            )
        cache_quant_type_str = getattr(layer, "cache_quant_type_str", "none")
        if cache_quant_type_str == "block_wise_fp8":
            cache_k = forward_meta.caches[4 * layer.layer_id]
            cache_v = forward_meta.caches[4 * layer.layer_id + 1]
            cache_k_scales = forward_meta.caches[4 * layer.layer_id + 2]
            cache_v_scales = forward_meta.caches[4 * layer.layer_id + 3]
        else:
            cache_k = forward_meta.caches[2 * layer.layer_id]
            cache_v = forward_meta.caches[2 * layer.layer_id + 1]
            cache_k_scales = getattr(layer, "cache_k_scale", None)
            cache_v_scales = getattr(layer, "cache_v_scale", None)

        if layer.layer_id == 0:
            # print(forward_meta.seq_lens_this_time)
            get_block_shape_and_split_kv_block(
                forward_meta.seq_lens_encoder,
                forward_meta.seq_lens_decoder,
                forward_meta.seq_lens_this_time,
                forward_meta.decoder_batch_ids,
                forward_meta.decoder_tile_ids_per_batch,
                forward_meta.decoder_num_blocks_cpu,
                forward_meta.decoder_num_blocks_device,
                forward_meta.decoder_chunk_size_device,
                forward_meta.max_len_tensor_cpu,
                forward_meta.encoder_batch_ids,
                forward_meta.encoder_tile_ids_per_batch,
                forward_meta.encoder_num_blocks_x_cpu,
                forward_meta.kv_batch_ids,
                forward_meta.kv_tile_ids_per_batch,
                forward_meta.kv_num_blocks_x_cpu,
                self.encoder_block_shape_q,
                self.decoder_block_shape_q,
                self.group_size,
                self.block_size,
            )

        if self.use_output:
            quant_max_bound = getattr(layer, "quant_max_bound", 0.0)
            cache_quant_type = getattr(layer, "cache_quant_type_str", "none")
            compute_type = metadata._fuse_kernel_compute_dtype
            out_scale = getattr(layer, "out_scale", -1.0)
            # 1. get output datatype
            qkv_dtype = qkv.dtype
            if qkv_dtype == paddle.float16:
                D_type = paddle.float16
            elif qkv_dtype == paddle.bfloat16:
                D_type = paddle.bfloat16
            elif qkv_dtype == paddle.int32:
                if compute_type == "bf16":
                    D_type = paddle.bfloat16
                elif compute_type == "fp16":
                    D_type = paddle.float16
                else:
                    raise NotImplementedError("Only supported attr of qkv_type in ['float16', 'bfloat16'].")
            else:
                raise NotImplementedError("Only supported attr of qkv_type in ['float16', 'bfloat16', 'int32'].")
            # 2.Extract related parameters
            token_nums = qkv.shape[0]
            head_dims = self.head_dim if cache_quant_type != "cache_int4_zp" else self.head_dim * 2
            q_num_heads = self.num_heads
            # 3. generate output tensor of different dtypes
            if out_scale > 0.0:
                if abs(quant_max_bound - 127) < 0.000001:
                    res = paddle.empty([token_nums, q_num_heads * head_dims], dtype="int8")
                elif abs(quant_max_bound - 448) < 0.000001:
                    res = paddle.empty([token_nums, q_num_heads * head_dims], dtype="float8_e4m3fn")
                else:
                    raise NotImplementedError("Only supported attr of quant_max_bound in ['127', '448'].")
            else:
                res = paddle.empty([token_nums, q_num_heads * head_dims], dtype=D_type)

            res = append_attention_with_output(
                qkv,
                cache_k,
                cache_v,
                forward_meta.seq_lens_encoder,
                forward_meta.seq_lens_decoder,
                forward_meta.seq_lens_this_time,
                forward_meta.batch_id_per_token,
                forward_meta.cu_seqlens_q,
                forward_meta.block_tables,
                forward_meta.encoder_batch_ids,
                forward_meta.encoder_tile_ids_per_batch,
                forward_meta.encoder_num_blocks_x_cpu,
                forward_meta.kv_batch_ids,
                forward_meta.kv_tile_ids_per_batch,
                forward_meta.kv_num_blocks_x_cpu,
                forward_meta.decoder_batch_ids,
                forward_meta.decoder_tile_ids_per_batch,
                forward_meta.decoder_num_blocks_cpu,
                forward_meta.max_len_tensor_cpu,
                res,
                forward_meta.rotary_embs,
                forward_meta.attn_mask,
                layer.qkv_bias,
                layer.qkv_scale,
                cache_k_scales,
                cache_v_scales,
                getattr(layer, "cache_k_out_scale", None),
                getattr(layer, "cache_v_out_scale", None),
                getattr(layer, "cache_k_zp", None),
                getattr(layer, "cache_v_zp", None),
                layer.linear_shift,
                layer.linear_smooth,
                forward_meta.attn_mask_offsets,
                metadata.kv_signal_data_list[layer.layer_id],
                getattr(layer, "q_norm_weight", None),
                getattr(layer, "k_norm_weight", None),
                getattr(layer, "sinks", None),
                getattr(layer, "rms_norm_eps", 1e-6),
                metadata._fuse_kernel_compute_dtype,
                getattr(layer, "cache_quant_type_str", "none"),
                layer.use_neox_rotary_style,
                self.rope_3d,
                self.max_seq_len,
                getattr(layer, "quant_max_bound", 0.0),
                getattr(layer, "quant_min_bound", 0.0),
                getattr(layer, "out_scale", -1.0),
                self.encoder_block_shape_q,
                self.decoder_block_shape_q,
                metadata.max_partition_size,
                metadata.encoder_max_partition_size,
                self.speculate_max_draft_token_num + 1,
                self.causal,
                self.speculative_method is not None,
                sliding_window,
            )
        else:
            res = append_attention(
                qkv,
                cache_k,
                cache_v,
                forward_meta.seq_lens_encoder,
                forward_meta.seq_lens_decoder,
                forward_meta.seq_lens_this_time,
                forward_meta.batch_id_per_token,
                forward_meta.cu_seqlens_q,
                forward_meta.block_tables,
                forward_meta.encoder_batch_ids,
                forward_meta.encoder_tile_ids_per_batch,
                forward_meta.encoder_num_blocks_x_cpu,
                forward_meta.kv_batch_ids,
                forward_meta.kv_tile_ids_per_batch,
                forward_meta.kv_num_blocks_x_cpu,
                forward_meta.decoder_batch_ids,
                forward_meta.decoder_tile_ids_per_batch,
                forward_meta.decoder_num_blocks_cpu,
                forward_meta.max_len_tensor_cpu,
                forward_meta.rotary_embs,
                forward_meta.attn_mask,
                layer.qkv_bias,
                layer.qkv_scale,
                cache_k_scales,
                cache_v_scales,
                getattr(layer, "cache_k_out_scale", None),
                getattr(layer, "cache_v_out_scale", None),
                getattr(layer, "cache_k_zp", None),
                getattr(layer, "cache_v_zp", None),
                layer.linear_shift,
                layer.linear_smooth,
                forward_meta.attn_mask_offsets,
                metadata.kv_signal_data_list[layer.layer_id],
                getattr(layer, "q_norm_weight", None),
                getattr(layer, "k_norm_weight", None),
                getattr(layer, "sinks", None),
                getattr(layer, "rms_norm_eps", 1e-6),
                metadata._fuse_kernel_compute_dtype,
                getattr(layer, "cache_quant_type_str", "none"),
                layer.use_neox_rotary_style,
                self.rope_3d,
                self.max_seq_len,
                getattr(layer, "quant_max_bound", 0.0),
                getattr(layer, "quant_min_bound", 0.0),
                getattr(layer, "out_scale", -1.0),
                self.encoder_block_shape_q,
                self.decoder_block_shape_q,
                metadata.max_partition_size,
                metadata.encoder_max_partition_size,
                self.speculate_max_draft_token_num + 1,
                self.causal,
                self.speculative_method is not None,
                sliding_window,
            )
        return res