wehub-resource-sync
130 lines
4.6 KiB
C++
130 lines
4.6 KiB
C++
// Copyright (c) 2024 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.
|
|
|
|
#pragma once
|
|
|
|
#include "paddle/phi/core/dense_tensor.h"
|
|
#include "paddle/phi/core/device_context.h"
|
|
#include "paddle/phi/kernels/full_kernel.h"
|
|
|
|
namespace phi {
|
|
|
|
template <typename T, typename Context>
|
|
void SwiGLUGradKernelImpl(const Context &dev_ctx,
|
|
const T *x,
|
|
const T *y,
|
|
const T *dz,
|
|
T *dx,
|
|
T *dy,
|
|
int64_t m,
|
|
int64_t n);
|
|
|
|
template <typename T, typename Context>
|
|
void SwiGLUGradKernel(const Context &dev_ctx,
|
|
const DenseTensor &x,
|
|
const optional<DenseTensor> &y,
|
|
const DenseTensor &dz,
|
|
DenseTensor *dx,
|
|
DenseTensor *dy) {
|
|
if (dx && dx->numel() == 0) {
|
|
dev_ctx.template Alloc<T>(dx);
|
|
if (dy) {
|
|
Full<T, Context>(dev_ctx, dy->dims(), 0, dy);
|
|
}
|
|
return;
|
|
}
|
|
|
|
if (dy && dy->numel() == 0) {
|
|
dev_ctx.template Alloc<T>(dy);
|
|
if (dx) {
|
|
Full<T, Context>(dev_ctx, dx->dims(), 0, dx);
|
|
}
|
|
return;
|
|
}
|
|
|
|
const auto *x_ptr = x.data<T>();
|
|
const auto *dz_ptr = dz.data<T>();
|
|
auto *dx_ptr = dx ? dev_ctx.template Alloc<T>(dx) : nullptr;
|
|
auto *dy_ptr = y && dy ? dev_ctx.template Alloc<T>(dy) : nullptr;
|
|
const auto &dims = x.dims();
|
|
|
|
if (y) {
|
|
const auto &y_tensor = y.get();
|
|
const auto &y_dims = y_tensor.dims();
|
|
const auto &dz_dims = dz.dims();
|
|
PADDLE_ENFORCE_EQ(y_dims,
|
|
dims,
|
|
common::errors::InvalidArgument(
|
|
"The shape of Input(Y):[%s] must be equal "
|
|
"to the shape of Input(X):[%s].",
|
|
y_dims,
|
|
dims));
|
|
PADDLE_ENFORCE_EQ(dz_dims,
|
|
dims,
|
|
common::errors::InvalidArgument(
|
|
"The shape of Input(dz):[%s] must be equal "
|
|
"to the shape of Input(X):[%s].",
|
|
dz_dims,
|
|
dims));
|
|
SwiGLUGradKernelImpl<T, Context>(dev_ctx,
|
|
x_ptr,
|
|
y_tensor.data<T>(),
|
|
dz_ptr,
|
|
dx_ptr,
|
|
dy_ptr,
|
|
x.numel(),
|
|
1);
|
|
} else {
|
|
auto dims_2d = flatten_to_2d(dims, dims.size() - 1);
|
|
int64_t m = dims_2d[0], n = dims_2d[1];
|
|
PADDLE_ENFORCE_EQ(n % 2,
|
|
0,
|
|
common::errors::InvalidArgument(
|
|
"The last dim of Input(X) should be exactly divided "
|
|
"by 2 when Input(Y) is None, but got %d",
|
|
n));
|
|
const auto &dz_dims = dz.dims();
|
|
PADDLE_ENFORCE_EQ(
|
|
dz_dims.size(),
|
|
dims.size(),
|
|
common::errors::InvalidArgument(
|
|
"The rank of Input(dz):[%d] must be equal to the rank of "
|
|
"Input(X):[%d] when Input(Y) is None.",
|
|
dz_dims.size(),
|
|
dims.size()));
|
|
for (int i = 0; i < dims.size() - 1; ++i) {
|
|
PADDLE_ENFORCE_EQ(dz_dims[i],
|
|
dims[i],
|
|
common::errors::InvalidArgument(
|
|
"The shape of Input(dz):[%s] must be equal to "
|
|
"the shape of Input(X):[%s] except the last dim "
|
|
"when Input(Y) is None.",
|
|
dz_dims,
|
|
dims));
|
|
}
|
|
PADDLE_ENFORCE_EQ(
|
|
dz_dims[dz_dims.size() - 1],
|
|
n / 2,
|
|
common::errors::InvalidArgument(
|
|
"The last dim of Input(dz):[%d] must be equal to half of the "
|
|
"last dim of Input(X):[%d] when Input(Y) is None.",
|
|
dz_dims[dz_dims.size() - 1],
|
|
n));
|
|
SwiGLUGradKernelImpl<T, Context>(
|
|
dev_ctx, x_ptr, nullptr, dz_ptr, dx_ptr, nullptr, m, n / 2);
|
|
}
|
|
}
|
|
|
|
} // namespace phi
|