mindspore 2.0.0rc1__cp38-none-any.whl → 2.2.0__cp38-none-any.whl

mindspore/ops/_op_impl/_custom_op/flash_attention/flash_attention_bwd.py

@@ -0,0 +1,467 @@
+# Copyright 2023 Huawei Technologies Co., Ltd
+#
+# 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.
+# ============================================================================
+"""The backward tik ops of flash attention"""
+from tbe import tik
+
+from mindspore.ops._op_impl._custom_op.flash_attention.tiling_strategy.strategy import TilingStrategy
+from mindspore.ops._op_impl._custom_op.flash_attention.attention import FlashAttention
+
+from mindspore.ops._op_impl._custom_op.flash_attention.constants import FP16
+from mindspore.ops._op_impl._custom_op.flash_attention.constants import FP32
+from mindspore.ops._op_impl._custom_op.flash_attention.constants import GM
+from mindspore.ops._op_impl._custom_op.flash_attention.constants import L1
+from mindspore.ops._op_impl._custom_op.flash_attention.constants import UB
+
+
+class FlashAttentionBwd(FlashAttention):
+    """The implementation of FlashAttention backward
+    This function contains the flash attention backward implementation used in flash attention (see paper)
+    `FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness <https://arxiv.org/pdf/2205.14135.pdf>`
+    """
+
+    def __init__(self, query, key, value, output, dO, rowsum, rowmax, attn_mask, dropout_mask, alibi_mask,
+                 prev_block_num,
+                 next_block_num,
+                 high_precision,
+                 kernel_name,
+                 tiling_stgy: TilingStrategy,
+                 disable_debug):
+        super().__init__(query, key, value, attn_mask, dropout_mask, alibi_mask, kernel_name,
+                         tiling_stgy, prev_block_num, next_block_num, high_precision, disable_debug)
+
+        if isinstance(query, dict):
+            self.dO_shape = dO["shape"]  # [B, Nq, d]
+        else:
+            self.dO_shape = dO.shape
+
+        self.dV_shape = self.v_shape
+        self.dQ_shape = self.q_shape
+        self.dK_shape = self.k_shape
+        self.dQ_gm = None
+        self.dK_gm = None
+        self.dV_gm = None
+        self.O_gm = None
+        self.dO_gm = None
+        self.l_gm = None
+        self.m_gm = None
+
+    def define_outputs(self):
+        """define output gm tensors"""
+        self.dQ_gm = self.tik_instance.Tensor(FP32, self.dQ_shape, name="dQ_gm", scope=GM, is_atomic_add=True)
+        self.dK_gm = self.tik_instance.Tensor(FP32, self.dK_shape, name="dK_gm", scope=GM, is_atomic_add=True)
+        self.dV_gm = self.tik_instance.Tensor(FP32, self.dV_shape, name="dV_gm", scope=GM, is_atomic_add=True)
+
+    def define_custom_inputs(self):
+        """define input gm tensors"""
+        self.O_gm = self.tik_instance.Tensor(FP16, self.O_shape, name="O_gm", scope=GM)
+        self.dO_gm = self.tik_instance.Tensor(FP16, self.dO_shape, name="dO_gm", scope=GM)
+        self.l_gm = self.tik_instance.Tensor(self.precision_type, self.l_shape, name="l_gm", scope=GM)
+        self.m_gm = self.tik_instance.Tensor(FP16, self.m_shape, name="m_gm", scope=GM)
+
+    def collect_inputs(self):
+        """collect all input gm tensors into input_gm_list,
+        the input list should keep order with the para order in Primitive and init
+        """
+        input_gm_list = [
+            self.Q_gm, self.K_gm, self.V_gm, self.O_gm, self.dO_gm, self.l_gm,
+            self.m_gm
+        ]
+        if self.has_attn_mask:
+            input_gm_list.append(self.att_mask_gm)
+        if self.has_drop_mask:
+            input_gm_list.append(self.drop_mask_gm)
+        if self.has_alibi_mask:
+            input_gm_list.append(self.alibi_mask_gm)
+        return input_gm_list
+
+    def prepare_global_ones(self):
+        """Prepare global ones tensor in L1 for cube impl row_sum"""
+        self.ones_l1 = self.tik_instance.Tensor(FP16, (self.d, 16), name="ones_l1", scope=L1)
+        with self.tik_instance.new_stmt_scope(disable_sync=False):
+            ones_ub = self.tik_instance.Tensor(FP16, (self.d, 16), name="ones_ub", scope=UB)
+            self.tik_instance.h_duplicate(ones_ub, 1.0)
+            self.cont_data_mv_1_bust(dst=self.ones_l1, src=ones_ub, burst=self.d)
+
+    def compute_Pij(self, Qi_l1_K1MK0_ed, KjT_l1_K1NK0_ed, m, k, n, lm_gm_offset, attn_mask_gm_offset,
+                    dropout_mask_gm_offset, alibi_mask_gm_offset):
+        """Refer to Algorithm 4 line11-14 in FlashAttention implement Pij computation"""
+        m_aligned = self.tik_ops_utils.up_align_to_K0(m)
+        n_aligned = self.tik_ops_utils.up_align_to_K0(n)
+        Sij_ub = self.tik_ops_utils.matmul_compute(Qi_l1_K1MK0_ed, KjT_l1_K1NK0_ed, m, k, n, N1MN0_to_MN=False)
+        Pij_drop_ed_ub = self.tik_instance.Tensor(FP16, (n_aligned // self.N0, m_aligned, self.N0),
+                                                  name="Pij_drop_ed_ub", scope=UB)
+
+        with self.tik_instance.new_stmt_scope(disable_sync=False):
+            if self.has_alibi_mask:
+                self.do_alibi_mask(Sij_ub, alibi_mask_gm_offset, m_aligned, n_aligned)
+            if self.has_attn_mask:
+                self.do_att_mask(Sij_ub, attn_mask_gm_offset, m, n, m_aligned, n_aligned)
+
+            # move li (ith block of l_gm) and mi (ith block of m_gm) from gm to ub
+            li_ub = self.tik_instance.Tensor(self.precision_type, (m_aligned,), name="li_ub", scope=UB)
+            mi_ub = self.tik_instance.Tensor(FP16, (m_aligned,), name="mi_ub", scope=UB)
+            self.tik_ops_utils.move_vector_from_gm_to_ub(li_ub, self.l_gm, lm_gm_offset, m)
+            self.tik_ops_utils.move_vector_from_gm_to_ub(mi_ub, self.m_gm, lm_gm_offset, m)
+            n1 = n_aligned // self.N0
+            with self.tik_instance.new_stmt_scope(disable_sync=False):
+                broadcast_mi_ub = self.tik_ops_utils.broadcast(mi_ub, (m, self.N0))
+                broadcast_mi_ub = broadcast_mi_ub.reshape((1, m, self.N0))
+                for idx in range(n1):
+                    self.tik_instance.h_sub(Sij_ub[idx, :, :], Sij_ub[idx, :, :], broadcast_mi_ub)
+            li_rec_ub = self.tik_ops_utils.calc_vec_rec(li_ub, m)
+            with self.tik_instance.new_stmt_scope(disable_sync=False):
+                if self.high_precision:
+                    # fp16 -> fp32
+                    Sij_ub_fp32 = self.tik_instance.Tensor(FP32, (n_aligned // self.N0, m_aligned, self.N0),
+                                                           name="Sij_ub_fp32", scope=UB)
+                    self.tik_instance.h_cast(Sij_ub_fp32, Sij_ub, "none")
+                    self.tik_instance.h_exp(Sij_ub_fp32, Sij_ub_fp32)
+                    cur_row_sum_rec = self.tik_instance.Tensor(FP32, (m_aligned, self.N0), name="cur_row_sum_rec",
+                                                               scope=UB)
+                    for i in range(m_aligned):
+                        src_scalar = self.tik_instance.Scalar(init_value=li_rec_ub[i], dtype=FP32)
+                        self.tik_instance.h_duplicate(cur_row_sum_rec[i, :], src_scalar)
+                    cur_row_sum_rec = cur_row_sum_rec.reshape((1, m_aligned, self.N0))
+                    with self.tik_instance.for_range(0, n_aligned // self.N0) as idx:
+                        self.tik_instance.h_mul(Sij_ub_fp32[idx, :, :], Sij_ub_fp32[idx, :, :], cur_row_sum_rec)
+                    # fp32 -> fp16
+                    self.tik_instance.h_cast(Sij_ub, Sij_ub_fp32, "none")
+                else:
+                    self.tik_instance.h_exp(Sij_ub, Sij_ub)
+                    broadcast_li_rec_ub = self.tik_ops_utils.broadcast(li_rec_ub, (m_aligned, self.N0))
+                    broadcast_li_rec_ub = broadcast_li_rec_ub.reshape((1, m_aligned, self.N0))
+                    for idx in range(n1):
+                        self.tik_instance.h_mul(Sij_ub[idx, :, :], Sij_ub[idx, :, :], broadcast_li_rec_ub)
+
+            if self.has_drop_mask:
+                self.do_dropout_mask(Sij_ub, dropout_mask_gm_offset, n_aligned, n, m_aligned, m,
+                                     workspace=Pij_drop_ed_ub)
+            else:
+                self.cont_data_mv_1_bust(dst=Pij_drop_ed_ub, src=Sij_ub, burst=m_aligned * n_aligned // 16)
+
+        return Sij_ub, Pij_drop_ed_ub
+
+    def compute_Di(self, Di_ub, dOi_ub, qo_gm_offset, q_blk_height):
+        """Refer to Algorithm 4 line19 in FlashAttention implement Di computation"""
+        q_blk_height_aligned = self.tik_ops_utils.up_align_to_K0(q_blk_height)
+        with self.tik_instance.new_stmt_scope(disable_sync=False):
+            Oi_ub = self.tik_instance.Tensor(FP16, (self.d // self.N0, q_blk_height_aligned, self.N0),
+                                             scope=UB, name="Oi_ub")
+            self.tik_instance.data_move(dst=Oi_ub, src=self.O_gm[qo_gm_offset],
+                                        sid=0, nburst=self.N1, burst=q_blk_height * self.N0 // 16,
+                                        src_stride=(self.Nq - q_blk_height) * self.N0 // 16, dst_stride=0)
+            self.tik_instance.h_mul(Oi_ub, dOi_ub, Oi_ub)
+            dOi_Oi_l1_K1MK0 = self.tik_instance.Tensor(FP16, (self.d // self.N0, q_blk_height_aligned, self.N0),
+                                                       name="dOi_Oi_l1_K1MK0", scope=L1)
+            self.cont_data_mv_1_bust(dst=dOi_Oi_l1_K1MK0, src=Oi_ub, burst=q_blk_height_aligned * self.d // 16)
+            self.tik_ops_utils.row_sum_cube_impl(dOi_Oi_l1_K1MK0, self.ones_l1, Di_ub, q_blk_height,
+                                                 self.actual_d, precision_type=FP16)
+
+    def compute_dSij(self, Pij_ub, dOi_l1_K1MK0_ed, VjT_K1NK0_ed, Di_ub, kv_blk_height, q_blk_height,
+                     dropout_mask_gm_offset):
+        """Refer to Algorithm 4 line20 in FlashAttention implement dSij computation"""
+        with self.tik_instance.new_stmt_scope(disable_sync=False):
+            dPij_ub = self.tik_ops_utils.matmul_compute(dOi_l1_K1MK0_ed, VjT_K1NK0_ed,
+                                                        q_blk_height, self.actual_d, kv_blk_height, N1MN0_to_MN=False)
+            q_blk_height_aligned = self.tik_ops_utils.up_align_to_K0(q_blk_height)
+            kv_blk_height_aligned = self.tik_ops_utils.up_align_to_K0(kv_blk_height)
+            # dropout_mask
+            if self.has_drop_mask:
+                self.do_dropout_mask(dPij_ub, dropout_mask_gm_offset, kv_blk_height_aligned, kv_blk_height,
+                                     q_blk_height_aligned, q_blk_height)
+            # dPij - Di
+            with self.tik_instance.new_stmt_scope(disable_sync=False):
+                broadcast_Di_ub = self.tik_ops_utils.broadcast(Di_ub, (q_blk_height_aligned, self.N0))
+                broadcast_Di_ub = broadcast_Di_ub.reshape((1, q_blk_height_aligned, self.N0))
+                n1 = kv_blk_height_aligned // self.N0
+                for idx in range(n1):
+                    self.tik_instance.h_sub(dPij_ub[idx, :, :], dPij_ub[idx, :, :], broadcast_Di_ub)
+            self.tik_instance.h_mul(Pij_ub, Pij_ub, dPij_ub)
+        return Pij_ub
+
+    def update_dVj(self,
+                   PijT_l1_K1MK0_ed,
+                   dOi_l1_K1NK0_ed,
+                   kv_gm_offset,
+                   kv_blk_height,
+                   q_blk_height):
+        """Refer to Algorithm 4 line16 in FlashAttention implement dVj update"""
+        with self.tik_instance.new_stmt_scope(disable_sync=False):
+            PijT_Oi_ub = self.tik_ops_utils.matmul_compute(PijT_l1_K1MK0_ed, dOi_l1_K1NK0_ed,
+                                                           kv_blk_height, q_blk_height,
+                                                           self.actual_d, N1MN0_to_MN=False,
+                                                           precision_type=FP32)
+            self.tik_instance.set_atomic_add(1)
+            self.tik_instance.data_move(dst=self.dV_gm[kv_gm_offset], src=PijT_Oi_ub, sid=0,
+                                        nburst=self.N1, burst=kv_blk_height * self.N0 // 8,
+                                        src_stride=0, dst_stride=(self.Nq - kv_blk_height) * self.N0 // 8)
+            self.tik_instance.set_atomic_add(0)
+
+    def update_dQi(self,
+                   dSij_l1_K1MK0_ed,
+                   Kj_l1_K1NK0_ed,
+                   qo_gm_offset,
+                   q_blk_height,
+                   kv_blk_height):
+        """Refer to Algorithm 4 line21 in FlashAttention implement dQi update"""
+        with self.tik_instance.new_stmt_scope(disable_sync=False):
+            dSij_Kj_ub = self.tik_ops_utils.matmul_compute(dSij_l1_K1MK0_ed, Kj_l1_K1NK0_ed,
+                                                           q_blk_height, kv_blk_height,
+                                                           self.actual_d, N1MN0_to_MN=False, precision_type=FP32)
+            self.tik_instance.set_atomic_add(1)
+            self.tik_instance.data_move(dst=self.dQ_gm[qo_gm_offset], src=dSij_Kj_ub, sid=0,
+                                        nburst=self.d // self.N0, burst=q_blk_height * self.N0 // 8,
+                                        src_stride=0, dst_stride=(self.Nq - q_blk_height) * self.N0 // 8)
+            self.tik_instance.set_atomic_add(0)
+
+    def update_dKj(self,
+                   dSijT_l1_K1MK0_ed,
+                   Qi_l1_K1NK0_ed,
+                   kv_gm_offset,
+                   kv_blk_height,
+                   q_blk_height):
+        """Refer to Algorithm 4 line22 in FlashAttention implement dKi update"""
+        with self.tik_instance.new_stmt_scope(disable_sync=False):
+            dSijT_Qi_ub = self.tik_ops_utils.matmul_compute(dSijT_l1_K1MK0_ed, Qi_l1_K1NK0_ed,
+                                                            kv_blk_height, q_blk_height,
+                                                            self.actual_d, N1MN0_to_MN=False, precision_type=FP32)
+            self.tik_instance.set_atomic_add(1)
+            self.tik_instance.data_move(dst=self.dK_gm[kv_gm_offset], src=dSijT_Qi_ub, sid=0,
+                                        nburst=self.d // self.N0, burst=kv_blk_height * self.N0 // 8,
+                                        src_stride=0, dst_stride=(self.Nq - kv_blk_height) * self.N0 // 8)
+            self.tik_instance.set_atomic_add(0)
+
+    def compute_in_each_kv_block(self, batch_start, batch_idx, kv_blk_idx, kv_blk_height,
+                                 core_idx_to_tr_info, core_idx):
+        """The backward computation in each outer loop"""
+        kv_blk_height_aligned = self.tik_ops_utils.up_align_to_K0(kv_blk_height)
+        kv_gm_offset = self.get_gm_offset(batch_start, batch_idx, self.N, self.d,
+                                          self.Bc, kv_blk_idx)
+        # load KjT
+        Kj_l1_1_K1MK0 = self.tik_instance.Tensor(FP16, (self.d // self.N0, kv_blk_height_aligned, self.N0),
+                                                 name="Kj_l1_1_K1MK0",
+                                                 scope=L1)
+        self.tik_instance.data_move(dst=Kj_l1_1_K1MK0, src=self.K_gm[kv_gm_offset],
+                                    sid=0, nburst=self.N1, burst=kv_blk_height_aligned * self.N0 // 16,
+                                    src_stride=(self.N - kv_blk_height_aligned) * self.N0 // 16, dst_stride=0)
+
+        # load Kj
+        Kj_l1_2 = self.tik_instance.Tensor(FP16, (kv_blk_height_aligned, self.d), name="Kj_l1_2",
+                                           scope=L1)
+        with self.tik_instance.new_stmt_scope(disable_sync=False):
+            Kj_ub = self.tik_instance.Tensor(FP16, (self.d // self.N0, kv_blk_height_aligned, self.N0),
+                                             name="Kj_ub", scope=UB)
+            self.tik_instance.data_move(dst=Kj_ub, src=self.K_gm[kv_gm_offset],
+                                        sid=0, nburst=self.N1, burst=kv_blk_height_aligned * self.N0 // 16,
+                                        src_stride=(self.N - kv_blk_height_aligned) * self.N0 // 16, dst_stride=0)
+            # (N1, K, N0) -> (K, N)
+            Kj_ub = self.tik_ops_utils.N1MN0_TO_MN(Kj_ub)
+            # (K, N) -> (K1, N, K0)
+            Kj_l1_2_K1NK0_ed = self.tik_ops_utils.KN_TO_K1NK0(Kj_ub, workspace_tensor=Kj_l1_2)
+
+        # load VjT
+        Vj_l1 = self.tik_instance.Tensor(FP16, (self.d // self.N0, kv_blk_height_aligned, self.N0), name="Vj_l1",
+                                         scope=L1)
+        self.tik_instance.data_move(dst=Vj_l1, src=self.V_gm[kv_gm_offset],
+                                    sid=0, nburst=self.N1, burst=kv_blk_height_aligned * self.N0 // 16,
+                                    src_stride=(self.N - kv_blk_height_aligned) * self.N0 // 16, dst_stride=0)
+
+        tr_start_s = self.tik_instance.Scalar("int32", name="tr_start_s")
+        tr_end_s = self.tik_instance.Scalar("int32", name="tr_end_s")
+        tr_start_s.set_as(core_idx_to_tr_info[core_idx, batch_start + batch_idx, 0])
+        tr_end_s.set_as(core_idx_to_tr_info[core_idx, batch_start + batch_idx, 1])
+        with self.tik_instance.for_range(tr_start_s, tr_end_s, name="q_blk_idx") as q_blk_idx:
+            with self.tik_instance.if_scope(tik.all(kv_blk_idx - self.next_block_num <= q_blk_idx,
+                                                    q_blk_idx <= kv_blk_idx + self.prev_block_num)):
+                with self.tik_instance.if_scope(q_blk_idx != self.Tr - 1):
+                    self.compute_in_each_q_block(Kj_l1_1_K1MK0,
+                                                 Kj_l1_2_K1NK0_ed,
+                                                 Vj_l1,
+                                                 batch_idx,
+                                                 batch_start,
+                                                 kv_gm_offset,
+                                                 kv_blk_height,
+                                                 self.Br,
+                                                 kv_blk_idx,
+                                                 q_blk_idx)
+                with self.tik_instance.else_scope():
+                    self.compute_in_each_q_block(Kj_l1_1_K1MK0,
+                                                 Kj_l1_2_K1NK0_ed,
+                                                 Vj_l1,
+                                                 batch_idx,
+                                                 batch_start,
+                                                 kv_gm_offset,
+                                                 kv_blk_height,
+                                                 self.last_Br,
+                                                 kv_blk_idx,
+                                                 q_blk_idx)
+
+    def compute_in_each_q_block(self, KjT_l1_K1NK0_ed, Kj_l1_K1NK0_ed, VjT_l1_K1NK0_ed,
+                                batch_idx, batch_start, kv_gm_offset, kv_blk_height,
+                                q_blk_height, kv_blk_idx, q_blk_idx):
+        """The backward computation in each inner loop"""
+        kv_blk_height_alig = self.tik_ops_utils.up_align_to_K0(kv_blk_height)
+        q_blk_height_alig = self.tik_ops_utils.up_align_to_K0(q_blk_height)
+
+        qo_gm_offset = self.get_gm_offset(batch_start, batch_idx, self.Nq, self.d, self.Br, q_blk_idx)
+        Qi_l1_K1MK0 = self.tik_instance.Tensor(FP16, (self.d // self.N0, q_blk_height_alig, self.N0),
+                                               name="Qi_l1_K1MK0",
+                                               scope=L1)
+        self.tik_instance.data_move(dst=Qi_l1_K1MK0, src=self.Q_gm[qo_gm_offset],
+                                    sid=0, nburst=self.N1, burst=q_blk_height_alig * self.N0 // 16,
+                                    src_stride=(self.Nq - q_blk_height_alig) * self.N0 // 16, dst_stride=0)
+
+        Qi_l1_right = self.tik_instance.Tensor(FP16, (q_blk_height_alig, self.d), name="Qi_l1_right",
+                                               scope=L1)
+        with self.tik_instance.new_stmt_scope(disable_sync=False):
+            Qi_ub = self.tik_instance.Tensor(FP16, (self.d // self.N0, q_blk_height_alig, self.N0),
+                                             name="Qi_ub", scope=UB)
+            self.tik_instance.data_move(dst=Qi_ub, src=self.Q_gm[qo_gm_offset],
+                                        sid=0, nburst=self.N1, burst=q_blk_height_alig * self.N0 // 16,
+                                        src_stride=(self.N - q_blk_height_alig) * self.N0 // 16, dst_stride=0)
+            # (N1, K, N0) -> (K, N)
+            Qi_ub = self.tik_ops_utils.N1MN0_TO_MN(Qi_ub)
+            # (K, N) -> (K1, N, K0)
+            Qi_l1_K1NK0_ed = self.tik_ops_utils.KN_TO_K1NK0(Qi_ub, workspace_tensor=Qi_l1_right)
+
+        lm_gm_offset = self.get_l_m_gm_offset(batch_start, batch_idx, self.Nq, self.Br, q_blk_idx)
+        attn_mask_gm_offset, dropout_mask_gm_offset, alibi_mask_gm_offset = None, None, None
+        if self.has_attn_mask:
+            attn_mask_gm_offset = self.get_attn_mask_gm_offset(batch_start, batch_idx, self.Nq, self.N,
+                                                               self.Br, q_blk_idx, self.Bc, kv_blk_idx)
+        if self.has_drop_mask:
+            dropout_mask_gm_offset = self.get_drop_mask_gm_offset(batch_start, batch_idx, self.Nq, self.N,
+                                                                  self.Br, q_blk_idx, self.Bc, kv_blk_idx)
+        if self.has_alibi_mask:
+            alibi_mask_gm_offset = self.get_alibi_gm_offset(batch_start, batch_idx, self.N, self.Bc, kv_blk_idx)
+        Pij_ub, Pij_drop_ed_ub = self.compute_Pij(Qi_l1_K1MK0, KjT_l1_K1NK0_ed,
+                                                  q_blk_height, self.actual_d, kv_blk_height,
+                                                  lm_gm_offset, attn_mask_gm_offset,
+                                                  dropout_mask_gm_offset, alibi_mask_gm_offset)
+
+        dOi_l1_right = self.tik_instance.Tensor(FP16, (q_blk_height_alig, self.d), name="dOi_l1_right",
+                                                scope=L1)
+        Di_ub = self.tik_instance.Tensor(FP16, (q_blk_height_alig,), name="Di_ub", scope=UB)
+        with self.tik_instance.new_stmt_scope(disable_sync=False):
+            dOi_ub = self.tik_instance.Tensor(FP16, (self.d // self.N0, q_blk_height_alig, self.N0),
+                                              name="dOi_ub", scope=UB)
+            self.tik_instance.data_move(dst=dOi_ub, src=self.dO_gm[qo_gm_offset],
+                                        sid=0, nburst=self.N1, burst=q_blk_height_alig * self.N0 // 16,
+                                        src_stride=(self.Nq - q_blk_height_alig) * self.N0 // 16, dst_stride=0)
+
+            self.compute_Di(Di_ub, dOi_ub, qo_gm_offset, q_blk_height)
+            # (N1, K, N0) -> (K, N)
+            dOi_ub = self.tik_ops_utils.N1MN0_TO_MN(dOi_ub)
+            # (K, N) -> (K1, N, K0)
+            dOi_l1_K1NK0_ed = self.tik_ops_utils.KN_TO_K1NK0(dOi_ub, workspace_tensor=dOi_l1_right)
+
+        dOi_l1_K1MK0 = self.tik_instance.Tensor(FP16, (self.d // self.N0, q_blk_height_alig, self.N0),
+                                                name="dOi_l1_K1MK0",
+                                                scope=L1)
+
+        self.tik_instance.data_move(dst=dOi_l1_K1MK0, src=self.dO_gm[qo_gm_offset],
+                                    sid=0, nburst=self.N1, burst=q_blk_height_alig * self.N0 // 16,
+                                    src_stride=(self.Nq - q_blk_height_alig) * self.N0 // 16, dst_stride=0)
+        Pij_l1 = self.tik_instance.Tensor(FP16, (q_blk_height_alig, kv_blk_height_alig), name="Pij_l1", scope=L1)
+        Pij_drop_ed_ub = self.tik_ops_utils.N1MN0_TO_MN(Pij_drop_ed_ub)
+        PijT_l1_K1MK0_ed = self.tik_ops_utils.KN_TO_K1NK0(Pij_drop_ed_ub, workspace_tensor=Pij_l1)
+        self.update_dVj(PijT_l1_K1MK0_ed, dOi_l1_K1NK0_ed,
+                        kv_gm_offset, kv_blk_height, q_blk_height)
+        # (L1: 512K)
+        dSij_l1_K1MK0_ed = self.tik_instance.Tensor(FP16, (kv_blk_height_alig // self.N0, q_blk_height_alig, self.N0),
+                                                    name="dSij_l1_1", scope=L1)
+        dSij_l1_2 = self.tik_instance.Tensor(FP16, (q_blk_height_alig, kv_blk_height_alig),
+                                             name="dSij_l1_2", scope=L1)
+        with self.tik_instance.new_stmt_scope(disable_sync=False):
+            dSij_ub = self.compute_dSij(Pij_ub,
+                                        dOi_l1_K1MK0,
+                                        VjT_l1_K1NK0_ed,
+                                        Di_ub,
+                                        kv_blk_height,
+                                        q_blk_height,
+                                        dropout_mask_gm_offset)
+            self.cont_data_mv_1_bust(dst=dSij_l1_K1MK0_ed, src=dSij_ub,
+                                     burst=kv_blk_height_alig * q_blk_height_alig // 16)
+            dSij_ub = self.tik_ops_utils.N1MN0_TO_MN(dSij_ub)
+            dSijT_l1_K1MK0_ed = self.tik_ops_utils.KN_TO_K1NK0(dSij_ub, workspace_tensor=dSij_l1_2)
+        self.update_dQi(dSij_l1_K1MK0_ed, Kj_l1_K1NK0_ed,
+                        qo_gm_offset, q_blk_height, kv_blk_height)
+        self.update_dKj(dSijT_l1_K1MK0_ed, Qi_l1_K1NK0_ed,
+                        kv_gm_offset, kv_blk_height, q_blk_height)
+
+    def compute_one_core(self, batch_start_sc, batch_num_sc, core_idx_to_tr_info, core_idx):
+        """The computation of FlashAttention backward on each core"""
+        with self.tik_instance.for_range(0, batch_num_sc, name="batch_index") as batch_idx:
+            with self.tik_instance.for_range(0, self.Tc, name="kv_blk_idx") as kv_blk_idx:
+                with self.tik_instance.if_scope(kv_blk_idx != self.Tc - 1):
+                    self.compute_in_each_kv_block(batch_start_sc, batch_idx, kv_blk_idx, self.Bc,
+                                                  core_idx_to_tr_info, core_idx)
+                with self.tik_instance.else_scope():
+                    self.compute_in_each_kv_block(batch_start_sc, batch_idx, kv_blk_idx, self.last_Bc,
+                                                  core_idx_to_tr_info, core_idx)
+
+    def collect_outputs(self):
+        """collect all output gm tensors into output_gm_list,
+        the output list should keep order with the para order in Primitive and init
+        """
+        output_gm_list = [self.dQ_gm, self.dK_gm, self.dV_gm]
+        return output_gm_list
+
+
+def flash_attention_grad(Query, Key, Value, Output, dO, rowsum, rowmax, attn_mask, dropout_mask, alibi_mask,
+                         dq, dk, dv,
+                         prev_block_num=65536,
+                         next_block_num=65536,
+                         high_precision=False,
+                         tiling_stgy_name='sparse',
+                         kernel_name="flash_attention_grad",
+                         disable_debug=True):
+    """
+    algorithm: flash_attention_backward
+
+    Parameters
+    ----------
+    Query : dict. shape and dtype of input, only support float16
+    Key : dict. shape and dtype of input, only support float16
+    Value: dict. shape and dtype of input, only support float16
+    Output: dict. shape and dtype of input, only support float16
+    dO: dict. shape and dtype of input, only support float16
+    rowsum: dict. shape and dtype of input, only support float16
+    rowmax: dict. shape and dtype of input, only support float16
+    dropout_mask: dict. shape and dtype of input, only support float16
+    dropout_mask: dict. shape and dtype of input, only support float16
+    alibi_mask: dict. shape and dtype of input, only support float16
+    dq: dict. shape and dtype of output, only support float16
+    dk: dict. shape and dtype of output, only support float16
+    dv: dict. shape and dtype of output, only support float16
+    prev_block_num: int. an attribute used to define sparse attention
+    next_block_num: int. an attribute used to define sparse attention
+    tiling_stgy_name: str. an attribute used to choose the tiling strategy
+    kernel_name: str. cce kernel name, default value is real_div
+    disable_debug: bool. whether disable debug
+
+    Returns
+    -------
+    tik_instance
+    """
+    fa_grad = FlashAttentionBwd(Query, Key, Value, Output, dO, rowsum, rowmax, attn_mask, dropout_mask,
+                                alibi_mask, prev_block_num=prev_block_num,
+                                next_block_num=next_block_num,
+                                high_precision=high_precision,
+                                kernel_name=kernel_name,
+                                tiling_stgy=TilingStrategy.from_strategy_name(tiling_stgy_name),
+                                disable_debug=disable_debug)
+    fa_grad.compute_process()
+    return fa_grad.tik_instance