mindstudio-probe 1.2.2__py3-none-any.whl → 8.1.0__py3-none-any.whl

msprobe/mindspore/api_accuracy_checker/bench_functions/flash_attention_score.py

@@ -0,0 +1,580 @@
+# Copyright (c) 2024-2024, Huawei Technologies Co., Ltd.
+# 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 collections import namedtuple
+import torch
+import torch.nn as nn
+import numpy as np
+
+from einops import rearrange
+
+
+from msprobe.pytorch.common.utils import logger
+
+GTYPE = torch.float64  # arm host必须选择float64，x86环境选择float32即可，64也行。arm计算很慢，s=8k的场景建议使用x86
+SOFTMAX_BUILD_MODE = "QKV"  # "MAX_SUM"
+
+FaForwardParams = namedtuple("FaForwardParams",
+                             ["q", "k", "v", "drop_mask", "attn_mask", "pse", "scalar_value", "keep_prob"])
+FaBackwardParams = namedtuple("FaBackwardParams",
+                              ["dx", "q", "k", "v", "softmax_res", "drop_mask", "pse", "scalar_value", "keep_prob"])
+RebuildSoftmaxParams = namedtuple("RebuildSoftmaxParams",
+                                  ["q", "k", "attn_mask", "pse", "scalar_value", "softmax_max", "softmax_sum"])
+
+
+def softmax_forward(x):
+    x_max = torch.max(x, dim=-1, keepdims=True)[0]
+    x_sub = x.sub(x_max)
+    y = torch.exp(x_sub)
+    x_sum = y.sum(dim=-1, keepdims=True)
+    res = y.div(x_sum)
+    return res, x_max, x_sum
+
+
+def softmax_grad(dp, softmax_res):
+    muls = dp * softmax_res
+    muls_r = muls.sum(dim=-1, keepdims=True)
+    sub_r = dp - muls_r
+    res = sub_r * softmax_res
+    return res
+
+
+def broadcast_kv(num_heads, num_kv_heads, kv_tensor, dtype):
+    # 检查维度
+    if kv_tensor.dim() != 4:
+        raise ValueError(f"broadcast_kv: kv_tensor 必须是 4 维 (B, N_kv, S, D)，但得到 {kv_tensor.shape}")
+    if num_kv_heads == 0 or num_kv_heads > num_heads:
+        raise ValueError("broadcast_kv: num_kv_heads 必须大于 0 且不超过 num_heads。")
+    if num_heads % num_kv_heads != 0:
+        raise ValueError(f"broadcast_kv: num_heads({num_heads}) 必须能被 num_kv_heads({num_kv_heads}) 整除。")
+
+
+    factor = num_heads // num_kv_heads
+    kv_shape = kv_tensor.shape
+    b = kv_shape[0]
+    s = kv_shape[2]
+    d = kv_shape[3]
+    kv_res = torch.zeros([b, num_heads, s, d]).to(dtype)
+    for i in range(num_heads):
+        j = i // factor
+        kv_res[:, i:i + 1, :, :] = kv_tensor[:, j:j + 1, :, :]
+    return kv_res
+
+
+def calculate_qk(q, k, attn_mask, pse, scalar_value):
+    # 基本形状检查
+    if q.dim() < 4 or k.dim() < 4:
+        raise ValueError(f"calculate_qk: q,k 必须至少 4 维，q={q.dim()}，k={k.dim()}")
+    # 检查 head_dim 一致性
+    if q.size(-1) != k.size(-1):
+        raise ValueError(f"calculate_qk: q.head_dim({q.size(-1)}) != k.head_dim({k.size(-1)})")
+
+    if k.dim() != 4:
+        raise ValueError(f"k tensor dimension must be 4, but got {k.dim()} dimensions (shape: {k.shape})")
+
+    if k.dim() == 3:
+        k = k.unsqueeze(1)  # 在head维度扩展
+
+    if pse is None or len(pse.shape) == 0:
+        qk = torch.matmul(q, k.permute(0, 1, 3, 2)).mul(scalar_value)
+    else:
+        qk = (torch.matmul(q, k.permute(0, 1, 3, 2)) + pse).mul(scalar_value)
+    if attn_mask is None or len(attn_mask.shape) == 0:
+        return qk
+    else:
+        qk = qk + attn_mask.bool() * (-40000.0)  # -10000
+    return qk
+
+
+def fusion_attention_forward(forward_params):
+    q = forward_params.q
+    k = forward_params.k
+    v = forward_params.v
+    drop_mask = forward_params.drop_mask
+    attn_mask = forward_params.attn_mask
+    pse = forward_params.pse
+    scalar_value = forward_params.scalar_value
+    keep_prob = forward_params.keep_prob
+
+    # 拦截 keep_prob 为 0 的情况，防止除零
+    if keep_prob == 0:
+        raise ValueError("fusion_attention_forward: keep_prob 不能为 0，避免除零错误。")
+
+    qk = calculate_qk(q, k, attn_mask, pse, scalar_value)
+    softmax_res, softmax_max, softmax_sum = softmax_forward(qk)
+    if drop_mask is None or len(drop_mask.shape) == 0:
+        drop_res = softmax_res
+    else:
+        drop_res = softmax_res * drop_mask * (1.0 / keep_prob)
+    y = torch.matmul(drop_res, v)
+    return y, softmax_max, softmax_sum
+
+
+def fusion_attention_backward(backward_params):
+    dx = backward_params.dx
+    q = backward_params.q
+    k = backward_params.k
+    v = backward_params.v
+    softmax_res = backward_params.softmax_res
+    drop_mask = backward_params.drop_mask
+    pse = backward_params.pse
+    scalar_value = backward_params.scalar_value
+    keep_prob = backward_params.keep_prob
+
+    # 拦截 keep_prob 为 0 的情况，防止除零
+    if keep_prob == 0:
+        raise ValueError("fusion_attention_backward: keep_prob 不能为 0，避免除零错误。")
+
+    dp = torch.matmul(dx, v.permute(0, 1, 3, 2))
+    if drop_mask is None or len(drop_mask.shape) == 0:
+        drop_res = softmax_res.permute(0, 1, 3, 2)
+        dp_drop = dp
+    else:
+        drop_res = softmax_res.mul(drop_mask).mul(1.0 / keep_prob).permute(0, 1, 3, 2)
+        dp_drop = dp * drop_mask * (1.0 / keep_prob)
+    dv = torch.matmul(drop_res, dx)
+    softmax_grad_res = (softmax_grad(dp_drop, softmax_res) * scalar_value)
+    dq = torch.matmul(softmax_grad_res, k)
+    dk = torch.matmul(softmax_grad_res.permute(0, 1, 3, 2), q)
+    return dq, dk, dv
+
+
+def parse_bsnd_args(query, key, head_num, input_layout):
+    supported_input_layout = ["BSH", "SBH", "BSND", "BNSD", "TND"]
+    b, s1, s2, n1, n2, d, h1, h2 = None, None, None, head_num, None, None, None, None
+
+    if not isinstance(input_layout, str) or input_layout not in supported_input_layout:
+        raise ValueError(f"Invalid input_layout arg which must be one of {supported_input_layout}.")
+
+    if input_layout == "TND":
+        raise ValueError(f"input_layout {input_layout} does not supported for now.")
+
+    # 防止 head_num 为 0
+    if n1 == 0:
+        raise ValueError("parse_bsnd_args: head_num (n1) 不能为 0，避免除零错误。")
+
+    try:
+        if input_layout == "BSH":
+            b, s1, h1 = query.shape
+            _, s2, h2 = key.shape
+            d = h1 // n1
+            # 拦截 d 为 0 的情况
+            if d == 0:
+                raise ValueError("parse_bsnd_args: 计算得到的 head_dim d 不能为 0。")
+            n2 = h2 // d
+        elif input_layout == "SBH":
+            s1, b, h1 = query.shape
+            s2, _, h2 = key.shape
+            d = h1 // n1
+            if d == 0:
+                raise ValueError("parse_bsnd_args: 计算得到的 head_dim d 不能为 0。")
+            n2 = h2 // d
+        elif input_layout == "BSND":
+            b, s1, n1, d = query.shape
+            _, s2, n2, _ = key.shape
+            if d == 0:
+                raise ValueError("parse_bsnd_args: head_dim d 不能为 0。")
+            h1 = n1 * d
+            h2 = n2 * d
+        elif input_layout == "BNSD":
+            b, n1, s1, d = query.shape
+            _, n2, s2, _ = key.shape
+            if d == 0:
+                raise ValueError("parse_bsnd_args: head_dim d 不能为 0。")
+            h1 = n1 * d
+            h2 = n2 * d
+    except Exception as e:
+        raise ValueError(f"query.shape: {query.shape}, key.shape: {key.shape}, parse_bsnd_args error: {e}") from e
+
+    ret = (b, s1, s2, n1, n2, d, h1, h2, query.dtype)
+    return ret
+
+
+def convert_from_bnsd(_input, input_layout):
+    """
+    transform qkv from bnsd to input_layout.
+    B: batch_size
+    S: sequence_length
+    N: num_heads
+    D: head_dim
+    Args:
+       _input (torch.Tensor): tensor of shape (B,N,S,D)
+        input_layout (str): "BSH" or "SBH" or "BSND" or "BNSD" or "TND"
+    Returns:
+        tensor of shape (B,N,S,D) or (B,S,N,D) or (S,B,H) or (B,S,H)
+    """
+    if input_layout == "BSH":
+        # (B,N,S,D)=>(B,S,N*D)
+        out = rearrange(_input, 'b n s d -> b s (n d)').contiguous()
+    elif input_layout == "SBH":
+        # (B,N,S,D)=>(S,B,N*D)
+        out = rearrange(_input, 'b n s d -> s b (n d)').contiguous()
+    elif input_layout == "BSND":
+        # (B,N,S,D)=>(B,S,N,D)
+        out = rearrange(_input, 'b n s d -> b s n d').contiguous()
+    elif input_layout == "TND":
+        raise ValueError(f"input_layout {input_layout} does not supported for now.")
+    else:
+        out = _input
+    return out
+
+
+def convert_to_bnsd(_input, n, input_layout):
+    """
+    transform qkv from input_layout to bnsd.
+    B: batch_size
+    S: sequence_length
+    N: num_heads
+    D: head_dim
+    Args:
+        _input (torch.Tensor): tensor of shape (B,N,S,D) or (B,S,N,D) or (S,B,H) or (B,S,H)
+        n (int): num_heads
+        input_layout (str):"BSH" or "SBH" or "BSND" or "BNSD" or "TND"
+    Returns:
+        tensor of shape (B,N,S,D)
+    """
+    if input_layout == "BSH":
+        # (B,S,N*D)=>(B,N,S,D)
+        out = rearrange(_input, 'b s (n d) -> b n s d', n=n)
+    elif input_layout == "SBH":
+        # (S,B,N*D)=>(B,N,S,D)
+        out = rearrange(_input, 's b (n d) -> b n s d', n=n)
+    elif input_layout == "BSND":
+        # (B,S,N,D)=>(B,N,S,D)
+        out = rearrange(_input, 'b s n d -> b n s d', n=n)
+    elif input_layout == "TND":
+        raise ValueError(f"input_layout {input_layout} does not supported for now.")
+    else:
+        out = _input
+    if out.dim() != 4:
+        raise ValueError(f"convert qkv format failed with input_layout {input_layout}.")
+    return out.to(GTYPE)
+
+
+def generate_attn_mask(*args):
+    """
+    # 当sparse_mode=2、3、4时小算子到融合算子会走这个优化，反过来看就要拆解回原来的基本实现
+    ===> attn_mask = torch.from_numpy(np.triu(np.ones([2048, 2048]), k=1)).to(dtype)
+    """
+
+    sparse_mode, attn_mask, b, n1, s1, s2, pre_tocken, next_tocken, dtype = args
+    shape = [s1, s2]
+
+    if attn_mask is not None:
+        # 当FA的输入已经包含attn_mask时，可以认为已经是转换之后的mask矩阵了，有三种特殊场景，即稀疏矩阵场景，需要进行逆向还原
+        if sparse_mode == 2 or sparse_mode == 3 or sparse_mode == 4:
+            logger.info(f"s1: {s1}, s2:{s2}, attn_mask.shape:{attn_mask.shape}, attn_mask.dtype:{attn_mask.dtype}")
+
+            if attn_mask.dim() == 2 and attn_mask.shape[0] == 2048 and attn_mask.shape[1] == 2048:
+                if attn_mask.equal(torch.from_numpy(np.triu(np.ones([2048, 2048]), k=1)).to(attn_mask.dtype)):
+                    if sparse_mode == 2:
+                        attn_mask = torch.from_numpy(np.triu(np.ones(shape), k=1))
+                    elif sparse_mode == 3:
+                        attn_mask = torch.from_numpy(np.triu(np.ones(shape), k=s2 - s1 + 1))
+                    elif sparse_mode == 4:
+                        attn_mask_u = torch.from_numpy(np.triu(np.ones(shape), k=next_tocken + 1))
+                        attn_mask_l = torch.from_numpy(np.tril(np.ones(shape), k=-pre_tocken - 1))
+                        attn_mask = attn_mask_u + attn_mask_l
+                    logger.debug(f"反向转换attn_mask {attn_mask.shape}")
+                    return attn_mask.to(dtype)
+
+        return attn_mask.to(dtype)
+
+    if attn_mask is not None:
+        if attn_mask.dim() == 2:
+            if attn_mask.shape[0] != s1 or attn_mask.shape[1] != s2:
+                raise ValueError(f"Invalid attn_mask shape `SS` {attn_mask.shape}")
+            shape = [s1, s2]
+        elif attn_mask.dim() == 4:
+            if attn_mask.shape[1] == 1:
+                shape = [b, 1, s1, s2] if b != 1 else [1, 1, s1, s2]
+            else:
+                shape = [b, n1, s1, s2] if b != 1 else [1, n1, s1, s2]
+
+    if sparse_mode == 0:
+        attn_mask_u = torch.from_numpy(np.triu(np.ones(shape), k=next_tocken + 1))
+        attn_mask_l = torch.from_numpy(np.tril(np.ones(shape), k=-pre_tocken - 1))
+        attn_mask = attn_mask_u + attn_mask_l
+    elif sparse_mode == 1:  # no sparse
+        attn_mask = torch.from_numpy(np.zeros(shape))
+    elif sparse_mode == 2:
+        attn_mask = torch.from_numpy(np.triu(np.ones(shape), k=1))
+    elif sparse_mode == 3:
+        attn_mask = torch.from_numpy(np.triu(np.ones(shape), k=s2 - s1 + 1))
+    elif sparse_mode == 4:
+        attn_mask_u = torch.from_numpy(np.triu(np.ones(shape), k=next_tocken + 1))
+        attn_mask_l = torch.from_numpy(np.tril(np.ones(shape), k=-pre_tocken - 1))
+        attn_mask = attn_mask_u + attn_mask_l
+    # 注:不会出现sparse_mode=5的情况，该情况要求必须要传入attn_mask，且attn_mask矩阵数据格式须为BNSS或B1SS，
+    # 因此可以认为FA的输入已经是正确的attn_mask了
+    return attn_mask.to(dtype)
+
+
+def generate_kv(key, value, n1, n2):
+    # N不等长适配by cdy
+    if not (n1 == n2):
+        k_new = broadcast_kv(n1, n2, key, key.dtype)
+        v_new = broadcast_kv(n1, n2, value, value.dtype)
+    else:
+        k_new = key
+        v_new = value
+    return k_new, v_new
+
+
+def rebuid_softmax_by_qkv(q, k, attn_mask, pse, scalar_value):
+    """
+    attention = softmax(QK^T/sqrt(d))V
+    softmax(x_i) = e^(x_i - x_max) / sum(e^(x_i - x_max))
+    """
+    logger.info("Using QKV to rebuild original softmax")
+    qk = calculate_qk(q, k, attn_mask, pse, scalar_value)
+    softmax_res, _, _ = softmax_forward(qk)
+    return softmax_res
+
+
+def rebuild_softmax_by_max_sum(softmax_params):
+    """
+    attention = softmax(QK^T/sqrt(d))V
+    softmax(x_i) = e^(x_i - x_max_i) / x_sum_i)
+    """
+    q = softmax_params.q
+    k = softmax_params.k
+    attn_mask = softmax_params.attn_mask
+    pse = softmax_params.pse
+    scalar_value = softmax_params.scalar_value
+    softmax_max = softmax_params.softmax_max
+    softmax_sum = softmax_params.softmax_sum
+    logger.info("Using softmax_max and softmax_sum to rebuild original softmax")
+
+    qk = calculate_qk(q, k, attn_mask, pse, scalar_value)
+    if softmax_max.shape[-1] == 0:
+        raise ValueError(f"softmax_max.shape[-1] must be non-zero, softmax_max.shape: {softmax_max.shape}")
+    repeat_dim = qk.shape[-1] // softmax_max.shape[-1]
+    softmax_res = torch.exp(qk.sub(softmax_max.repeat(1, 1, 1, repeat_dim))).div(
+        softmax_sum.repeat(1, 1, 1, repeat_dim))
+    return softmax_res
+
+
+def get_head_num(*args, **kwargs):
+    if kwargs.get("head_num", None):
+        head_num = kwargs.get("head_num")
+    elif len(args) >= 4:
+        head_num = args[3]
+    else:
+        raise ValueError(f"Unsupported npu_fusion_attention args {args}.")
+    return head_num
+
+
+def get_input_layout(*args, **kwargs):
+    if kwargs.get("input_layout", None):
+        input_layout = kwargs.get("input_layout")
+    elif len(args) >= 5:
+        input_layout = args[4]
+    else:
+        raise ValueError(f"Unsupported npu_fusion_attention args {args}.")
+    return input_layout
+
+
+def npu_fusion_attention_forward_patch(*args, **kwargs):
+    if len(args) < 2:
+        raise RuntimeError("npu_fusion_attention_forward_patch: length of args should be greater than or equal to 2.")
+
+    # query, key, value, head_num, input_layout
+    head_num = get_head_num(*args, **kwargs)
+    input_layout = get_input_layout(*args, **kwargs)
+
+    b, s1, s2, n1, n2, d, h1, h2, dtype = parse_bsnd_args(args[0], args[1], head_num, input_layout)
+    # 此处 d 已在 parse_bsnd_args 中检查为非零
+    if n1 == n2 and s1 == s2:
+        logger.debug(f"running case : BNSD = {b}_{n1}_{s1}_{d}, sparse = {kwargs.get('sparse_mode', 0)}")
+    else:
+        logger.debug(f"running case: BNSD = {b}_{n1}({n2})_{s1}({s2})_{d}, sparse = {kwargs.get('sparse_mode', 0)}")
+    if n2 == 0:
+        raise ValueError("n2 不能为 0，避免除零错误。")
+    if not (n1 % n2 == 0 and n1 >= n2):
+        raise ValueError(f"N1与N2不匹配,请检查: n1 = {n1}, n2 = {n2}.")
+
+    dims_kwargs = {
+        "b": b, "s1": s1, "s2": s2, "n1": n1, "n2": n2,
+        "d": d, "h1": h1, "h2": h2, "dtype": dtype
+    }
+    new_kwargs = {
+        "keep_prob": 1,  # 注意：如果外部传入 keep_prob 为 0，也会在 fusion_attention_forward 中捕获
+        "scalar_value": kwargs.get("scalar_value", 1 / (d ** 0.5)),
+        "sparse_mode": kwargs.get("sparse_mode", 0),
+        "prefix": kwargs.get("prefix"),
+        "pre_tockens": kwargs.get("pre_tockens", 2147483647),
+        "next_tockens": kwargs.get("next_tockens", 2147483647),
+        "pse": kwargs.get("pse"),
+        "padding_mask": kwargs.get("padding_mask"),
+        "attn_mask": kwargs.get("attn_mask")
+    }
+
+    return args, dims_kwargs, new_kwargs
+
+
+def npu_fusion_attention_backward_patch(*args, **kwargs):
+    if len(args) != 6:
+        raise ValueError(f"Unsupported npu_fusion_attention_grad args {args}.")
+
+    b, s1, s2, n1, n2, d, h1, h2, dtype = parse_bsnd_args(args[0], args[1], args[4], args[5])
+    # 此处 d 已在 parse_bsnd_args 中检查为非零
+    if n1 == n2 and s1 == s2:
+        logger.info(f"running case : bnsd = {b}_{n1}_{s1}_{d}, sparse = {kwargs.get('sparse_mode', 0)}")
+    else:
+        logger.info(f"running case: bnsd = {b}_{n1}({n2})_{s1}({s2})_{d}, sparse = {kwargs.get('sparse_mode', 0)}")
+    if n2 == 0:
+        raise ValueError("n2 不能为 0，避免除零错误。")
+    if not (n1 % n2 == 0 and n1 >= n2):
+        raise ValueError(f"N1与N2不匹配,请检查: n1 = {n1}, n2 = {n2}.")
+
+    dims_kwargs = {
+        "b": b, "s1": s1, "s2": s2, "n1": n1, "n2": n2,
+        "d": d, "h1": h1, "h2": h2, "dtype": dtype
+    }
+
+    new_kwargs = {
+        "keep_prob": 1,  # 同上，fusion_attention_backward 内会拦截 keep_prob 为 0 的情况
+        "scalar_value_value": kwargs.get("scalar_value_value", 1 / (d ** 0.5)),
+        "sparse_mode": kwargs.get("sparse_mode", 0),
+        "prefix": kwargs.get("prefix"),
+        "pre_tockens": kwargs.get("pre_tockens", 2147483647),
+        "next_tockens": kwargs.get("next_tockens", 2147483647),
+        "pse": kwargs.get("pse"),
+        "padding_mask": kwargs.get("padding_mask"),
+        "softmax_max": kwargs.get("softmax_max"),
+        "softmax_sum": kwargs.get("softmax_sum"),
+        "softmax_in": kwargs.get("softmax_in"),
+        "attention_in": kwargs.get("attention_in"),
+        "seed": kwargs.get("seed", 0),
+        "offset": kwargs.get("offset", 0),
+        "numels": kwargs.get("numels", 0),
+        "attn_mask": kwargs.get("attn_mask")
+    }
+
+    return args, dims_kwargs, new_kwargs
+
+
+class FlashAttentionScore(nn.Module):
+    def __init__(self):
+        super(FlashAttentionScore, self).__init__()
+        # You can initialize any parameters here if necessary
+
+    def forward(self, *inputs, **kwargs):
+        # Extract the inputs for the attention calculation
+        new_args, dims_kwargs, new_kwargs = npu_fusion_attention_forward_patch(*inputs, **kwargs)
+        query, key, value = new_args[0], new_args[1], new_args[2]
+
+        input_layout = get_input_layout(*inputs, **kwargs)
+
+        n1 = dims_kwargs.get("n1")
+        n2 = dims_kwargs.get("n2")
+        s1 = dims_kwargs.get("s1")
+        s2 = dims_kwargs.get("s2")
+        b = dims_kwargs.get("b")
+        dtype = dims_kwargs.get("dtype")
+        attn_mask = new_kwargs.get("attn_mask")
+        keep_prob = new_kwargs.get("keep_prob")
+        sparse_mode = new_kwargs.get("sparse_mode")
+        pre_tockens = new_kwargs.get("pre_tockens")
+        next_tockens = new_kwargs.get("next_tokens")
+        pse = new_kwargs.get("real_shift")
+        scalar_value = new_kwargs.get("scalar_value")
+
+        args_temp = [sparse_mode, attn_mask, b, n1, s1, s2, pre_tockens, next_tockens, dtype]
+
+        attn_mask = generate_attn_mask(*args_temp)
+        query = convert_to_bnsd(query, n1, input_layout)
+        key = convert_to_bnsd(key, n2, input_layout)
+        value = convert_to_bnsd(value, n2, input_layout)
+
+        forward_params = FaForwardParams(
+            q=query,
+            k=key,
+            v=value,
+            drop_mask=None,
+            attn_mask=attn_mask,
+            pse=pse,
+            scalar_value=scalar_value,
+            keep_prob=keep_prob
+        )
+
+        out_golden, softmax_max, softmax_sum = fusion_attention_forward(forward_params)
+
+        # If output dimension is 5, reshape accordingly
+        if out_golden.dim() == 5:
+            out_golden = out_golden.reshape(out_golden.size(0),
+                                            out_golden.size(1) * out_golden.size(2),
+                                            out_golden.size(3), out_golden.size(4))
+
+        out_golden = convert_from_bnsd(out_golden, input_layout)
+
+        # Ensure the output matches the desired layout
+        out_golden = out_golden.cpu(), softmax_max.repeat(1, 1, 1, 8).cpu(), softmax_sum.repeat(1, 1, 1, 8).cpu()
+
+        return out_golden
+
+    def backward(self, *inputs, **kwargs):
+        # The backward pass will be similar to what was described for the gradient computation
+        new_args, dims_kwargs, new_kwargs = npu_fusion_attention_backward_patch(*inputs, **kwargs)
+        query, key, value, dx, input_layout = new_args[0], new_args[1], new_args[2], new_args[3], new_args[5]
+        n1 = dims_kwargs.get("n1")
+        n2 = dims_kwargs.get("n2")
+        s1 = dims_kwargs.get("s1")
+        s2 = dims_kwargs.get("s2")
+        b = dims_kwargs.get("b")
+        dtype = dims_kwargs.get("dtype")
+        attn_mask = new_kwargs.get("attn_mask")
+        keep_prob = new_kwargs.get("keep_prob")
+        sparse_mode = new_kwargs.get("sparse_mode")
+        pre_tockens = new_kwargs.get("pre_tockens")
+        next_tockens = new_kwargs.get("next_tockens")
+        pse = new_kwargs.get("pse")
+        softmax_max = new_kwargs.get("softmax_max")
+        softmax_sum = new_kwargs.get("softmax_sum")
+        scalar_value = new_kwargs.get("scalar_value")
+
+        args_temp = [sparse_mode, attn_mask, b, n1, s1, s2, pre_tockens, next_tockens, dtype]
+        attn_mask = generate_attn_mask(*args_temp)
+
+        query = convert_to_bnsd(query, n1, input_layout)
+        dx = convert_to_bnsd(dx, n1, input_layout)
+        key = convert_to_bnsd(key, n2, input_layout)
+        value = convert_to_bnsd(value, n2, input_layout)
+
+        k_new, v_new = generate_kv(key, value, n1, n2)
+
+        if SOFTMAX_BUILD_MODE == "QKV":
+            softmax_res = rebuid_softmax_by_qkv(query, k_new, attn_mask, pse, scalar_value)
+        else:
+            softmax_params = RebuildSoftmaxParams(query, k_new, attn_mask, pse, scalar_value, softmax_max, softmax_sum)
+            softmax_res = rebuild_softmax_by_max_sum(softmax_params)
+
+        backward_params = FaBackwardParams(dx, query, k_new, v_new, softmax_res, None, pse, scalar_value, keep_prob)
+        dq, dk, dv = fusion_attention_backward(backward_params)
+
+        # Reshape as needed
+        if dq.dim() == 5:
+            dq = dq.reshape(dq.size(0), dq.size(1) * dq.size(2), dq.size(3), dq.size(4))
+        if dk.dim() == 5:
+            dk = dk.reshape(dk.size(0), dk.size(1) * dk.size(2), dk.size(3), dk.size(4))
+        if dv.dim() == 5:
+            dv = dv.reshape(dv.size(0), dv.size(1) * dv.size(2), dv.size(3), dv.size(4))
+
+        dq = convert_from_bnsd(dq, input_layout)
+        dk = convert_from_bnsd(dk, input_layout)
+        dv = convert_from_bnsd(dv, input_layout)
+
+        return dq.cpu(), dk.cpu(), dv.cpu()

msprobe/mindspore/api_accuracy_checker/bench_functions/fusion_operator.py

@@ -0,0 +1,41 @@
+# Copyright (c) 2024-2024, Huawei Technologies Co., Ltd.
+# 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 msprobe.mindspore.api_accuracy_checker.bench_functions.flash_attention_score import FlashAttentionScore
+
+
+class FusionOperator:
+    """
+    所有融合算子的父类，定义了通用的接口和属性。
+    """
+
+    # 初始化操作符字典
+    def __init__(self):
+        self.flash_attention_score = None  # 用于存放 FlashAttentionScore 操作符
+        self._register_operators()
+
+    def __getattr__(self, name):
+        """ 动态获取算子类 """
+        if hasattr(self, name):
+            return getattr(self, name)
+        else:
+            raise AttributeError(f"'FusionOperator' object has no attribute '{name}'")
+
+    def _register_operators(self):
+        """ 注册操作符到父类，以便通过 fusion.xxx 调用 """
+        self.flash_attention_score = FlashAttentionScore()
+
+
+fusion = FusionOperator()

msprobe/mindspore/api_accuracy_checker/cmd_parser.py

@@ -39,6 +39,8 @@ def add_api_accuracy_checker_argument(parser):
                         help="<optional> The ut task result out path.")
     parser.add_argument("-csv_path", "--result_csv_path", dest="result_csv_path", default="", type=str, required=False,
                         help="<optional> the exit csv for continue")
+    parser.add_argument('-save_error_data', dest="save_error_data", action="store_true",
+                        help="<optional> Save compare failed api output.", required=False)
 
 
 def multi_add_api_accuracy_checker_argument(parser):
@@ -49,6 +51,8 @@ def multi_add_api_accuracy_checker_argument(parser):
                         help="<optional> The ut task result out path.")
     parser.add_argument("-csv_path", "--result_csv_path", dest="result_csv_path", default="", type=str, required=False,
                         help="<optional> the exit csv for continue")
+    parser.add_argument('-save_error_data', dest="save_error_data", action="store_true",
+                        help="<optional> Save compare failed api output.", required=False)
     #以下属于多线程参数
     parser.add_argument("-d", "--device", dest="device_id", nargs='+', type=int,
                         help="<optional> set device id to run ut, must be unique and in range 0-7",

msprobe/mindspore/api_accuracy_checker/data_manager.py

@@ -16,12 +16,13 @@
 import os
 import csv
 
-from msprobe.core.common.const import Const, CompareConst, MsCompareConst
+from msprobe.core.common.const import Const, CompareConst
 from msprobe.core.common.file_utils import FileOpen, create_directory, write_csv, read_csv
 from msprobe.core.common.utils import add_time_as_suffix, MsprobeBaseException
 from msprobe.mindspore.api_accuracy_checker.base_compare_algorithm import compare_algorithms
 from msprobe.core.common.file_utils import check_file_or_directory_path
 from msprobe.mindspore.common.log import logger
+from msprobe.mindspore.common.const import MsCompareConst
 
 
 class ResultCsvEntry:
@@ -187,7 +188,7 @@ class DataManager:
 
     def record_exception_skip(self, api_name, forward_or_backward, err_msg):
         '''
-            record exception_skip infomation into self.record_exception_skip.
+            record exception_skip information into self.record_exception_skip.
             self.record_exception_skip: dict{str: dict{"forward": str/None, "backward": str/None}}
             string in key is api_name, string in value is err_msg
         '''
@@ -269,7 +270,7 @@ class DataManager:
                 entry.backward_pass_status,
                 overall_err_msg
             ]
-            # change row if this api has excption_skip infomation
+            # change row if this api has exception_skip information
             if api_name in self.results_exception_skip:
                 if self.results_exception_skip[api_name][Const.FORWARD] is not None:
                     row[1] = CompareConst.SKIP

msprobe/mindspore/api_accuracy_checker/generate_op_script/config_op.json

@@ -0,0 +1,9 @@
+{
+    "dump_json_path": "./dump.json",
+    "api_name": "Mint.split.1",
+    "extract_api_path": "Mint.split.1.json",
+    "propagation": "backward",
+    "data_mode": "random_data",
+    "random_seed": 1234,
+    "iter_times": 1
+}