90 lines
3.8 KiB
C++
90 lines
3.8 KiB
C++
// 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.
|
|
|
|
#include "paddle/phi/backends/xpu/xpu_context.h"
|
|
#include "paddle/phi/core/kernel_registry.h"
|
|
|
|
namespace xpu = baidu::xpu::api;
|
|
|
|
namespace phi {
|
|
|
|
template <typename T, typename Context>
|
|
void MoECombineBackwardKernel(const Context& dev_ctx,
|
|
const DenseTensor& x,
|
|
const DenseTensor& combine_weights,
|
|
const DenseTensor& scatter_index,
|
|
const DenseTensor& grad_y,
|
|
DenseTensor* grad_x,
|
|
DenseTensor* grad_combine_weights_helper) {
|
|
int64_t seq_len = combine_weights.dims()[0];
|
|
int64_t k = combine_weights.dims()[1];
|
|
int64_t hidden_size = x.dims()[1];
|
|
|
|
using XPUType = typename XPUTypeTrait<T>::Type;
|
|
|
|
auto dy_data = reinterpret_cast<const XPUType*>(grad_y.data<T>());
|
|
auto x_data = reinterpret_cast<const XPUType*>(x.data<T>());
|
|
auto weight_data =
|
|
reinterpret_cast<const XPUType*>(combine_weights.data<T>());
|
|
auto index_data = scatter_index.data<int>();
|
|
auto dx_data = reinterpret_cast<XPUType*>(grad_x->data<T>());
|
|
auto dw_data =
|
|
reinterpret_cast<XPUType*>(grad_combine_weights_helper->data<T>());
|
|
int ret =
|
|
xpu::constant<XPUType>(dev_ctx.x_context(), dx_data, x.numel(), 0.0f);
|
|
PADDLE_ENFORCE_XDNN_SUCCESS(ret, "constant");
|
|
ret = xpu::moe_combine_grad<XPUType, int>(dev_ctx.x_context(),
|
|
dy_data,
|
|
x_data,
|
|
weight_data,
|
|
index_data,
|
|
dx_data,
|
|
dw_data,
|
|
seq_len,
|
|
k,
|
|
hidden_size);
|
|
PADDLE_ENFORCE_XDNN_SUCCESS(ret, "moe_combine_grad");
|
|
}
|
|
|
|
template <typename T, typename Context>
|
|
void MoeCombineGradKernel(const Context& dev_ctx,
|
|
const DenseTensor& x,
|
|
const DenseTensor& combine_weights,
|
|
const DenseTensor& scatter_index,
|
|
const DenseTensor& grad_y,
|
|
DenseTensor* grad_x,
|
|
DenseTensor* grad_combine_weights_helper) {
|
|
PD_CHECK(x.dims().size() == 2, "The shape of X must be 2.");
|
|
PD_CHECK(scatter_index.dtype() == DataType::INT32,
|
|
"MoE combine only supports int32 for scatter_index");
|
|
dev_ctx.template Alloc<T>(grad_x);
|
|
dev_ctx.template Alloc<T>(grad_combine_weights_helper);
|
|
MoECombineBackwardKernel<T, Context>(dev_ctx,
|
|
x,
|
|
combine_weights,
|
|
scatter_index,
|
|
grad_y,
|
|
grad_x,
|
|
grad_combine_weights_helper);
|
|
}
|
|
} // namespace phi
|
|
|
|
PD_REGISTER_KERNEL(moe_combine_grad,
|
|
XPU,
|
|
ALL_LAYOUT,
|
|
phi::MoeCombineGradKernel,
|
|
float,
|
|
phi::bfloat16,
|
|
phi::float16) {}
|