开启main_grad的时候，怎么融合gradient accumulation？ about paddle HOT 7 CLOSED

# Copyright (c) 2023 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.
import paddle
from paddle import _C_ops
from paddle.framework import core
import numpy as np
from paddle.distributed.fleet.utils import mix_precision_utils

def is_fused_matmul_bias_supported():
    if paddle.is_compiled_with_cuda() and not paddle.is_compiled_with_rocm() or paddle.is_compiled_with_xpu():
        return hasattr(core.eager.ops.legacy, "fused_gemm_epilogue")
    else:
        return False


if is_fused_matmul_bias_supported():
    origin_linear = paddle.incubate.nn.functional.fused_linear
else:
    origin_linear = paddle.nn.functional.linear


class FusedLinearWithGradAdd(paddle.autograd.PyLayer):
    @staticmethod
    def forward(ctx, x, weight, bias=None, name=None):
        y = origin_linear(x, weight, bias)
        ctx.save_for_backward(x, weight, bias)
        return y

    @staticmethod
    def backward(ctx, y_grad):
        x, weight, bias = ctx.saved_tensor()
        x_grad = paddle.matmul(y_grad, weight, transpose_y=True)

        # _C_ops.fused_linear_param_grad_add(x, y_grad, dw, db, multi precision, has bias)
        if bias is None:
            if hasattr(weight, "main_grad"):
                weight.main_grad, _ = _C_ops.fused_linear_param_grad_add(
                    x, y_grad, weight.main_grad, None, True, False
                )
                return x_grad, None
            else:
                if weight.grad is not None:
                    weight.grad, _ = _C_ops.fused_linear_param_grad_add(x, y_grad, weight.grad, None, False, False)
                    return x_grad, None
                else:
                    weight_grad, _ = _C_ops.fused_linear_param_grad_add(x, y_grad, None, None, False, False)
                    return x_grad, weight_grad

        if hasattr(weight, "main_grad") and hasattr(bias, "main_grad"):
            weight.main_grad, bias.main_grad = _C_ops.fused_linear_param_grad_add(
                x, y_grad, weight.main_grad, bias.main_grad, True, True
            )
            return x_grad, None, None
        else:
            if weight.grad is not None:
                assert bias.grad is not None
                weight.grad, bias.grad = _C_ops.fused_linear_param_grad_add(
                    x, y_grad, weight.grad, bias.grad, False, True
                )
                return x_grad, None, None
            else:
                weight_grad, bias_grad = _C_ops.fused_linear_param_grad_add(x, y_grad, None, None, False, True)
                return x_grad, weight_grad, bias_grad


def mock_layers():
    paddle.nn.functional.linear = FusedLinearWithGradAdd.apply
    if is_fused_matmul_bias_supported():
        paddle.incubate.nn.functional.fused_linear = FusedLinearWithGradAdd.apply


mock_layers()

def create_optimizer(model, use_pure_bf16, use_main_grad):
    if use_main_grad:
        assert use_pure_bf16
        model = mix_precision_utils.MixPrecisionLayer(model, dtype="bfloat16")
    optimizer = paddle.optimizer.AdamW(
        parameters=model.parameters(),
        learning_rate=0.0001,
        multi_precision=use_pure_bf16,
    )
    if use_main_grad:
        optimizer = mix_precision_utils.MixPrecisionOptimizer(optimizer)

    return optimizer

class Net(paddle.nn.Layer):
    """Network use for recompute testing"""

    def __init__(self):
        super().__init__()
        self.layer = paddle.nn.Linear(32, 32)

    def forward(self, inp):
        out = self.layer(inp)
        return out

def main():
    paddle.seed(10)

    model = Net()
    optimizer = create_optimizer(model, use_pure_bf16=True, use_main_grad=True)
    model = paddle.amp.decorate(models=model, dtype="bfloat16", level='O2', master_grad=True)

    model.train()
    for _ in range(10):
        inp = paddle.normal(mean=0, std=0.01, shape=[1, 32, 32]).astype('float32')
        inp.stop_gradient = False
        with paddle.amp.auto_cast(True, level="O2", dtype="bfloat16"):
            out = model(inp)
        loss = out.mean()
        loss.backward()
        optimizer.step()
        optimizer.clear_grad()
        print(loss)

if __name__ == "__main__":
    main()