211 lines
8.4 KiB
Plaintext
211 lines
8.4 KiB
Plaintext
// 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.
|
|
|
|
#include "paddle/phi/backends/gpu/gpu_context.h"
|
|
#include "paddle/phi/backends/gpu/gpu_launch_config.h"
|
|
#include "paddle/phi/common/amp_type_traits.h"
|
|
#include "paddle/phi/core/enforce.h"
|
|
#include "paddle/phi/core/kernel_registry.h"
|
|
#include "paddle/phi/kernels/funcs/aligned_vector.h"
|
|
#include "paddle/phi/kernels/fusion/gpu/fused_rope_utils.h"
|
|
|
|
namespace phi {
|
|
namespace fusion {
|
|
|
|
template <typename T, typename Context>
|
|
void FusedRopeGradKernel(const Context& dev_ctx,
|
|
const optional<DenseTensor>& sin,
|
|
const optional<DenseTensor>& cos,
|
|
const optional<DenseTensor>& position_ids,
|
|
const DenseTensor& dout_q,
|
|
const optional<DenseTensor>& dout_k,
|
|
const optional<DenseTensor>& dout_v,
|
|
bool use_neox_rotary_style,
|
|
bool time_major,
|
|
float rotary_emb_base,
|
|
DenseTensor* dq,
|
|
DenseTensor* dk,
|
|
DenseTensor* dv) {
|
|
int64_t numel = dout_q.numel();
|
|
dev_ctx.template Alloc<T>(dq);
|
|
if (dout_k) dev_ctx.template Alloc<T>(dk);
|
|
if (dout_v) dev_ctx.template Alloc<T>(dv);
|
|
if (numel <= 0) return;
|
|
|
|
auto batch_size = time_major ? dout_q.dims()[1] : dout_q.dims()[0];
|
|
auto seq_len = time_major ? dout_q.dims()[0] : dout_q.dims()[1];
|
|
auto num_heads = dout_q.dims()[2];
|
|
auto head_dim = dout_q.dims()[3];
|
|
auto freqs_head_dim = head_dim;
|
|
PADDLE_ENFORCE_NE(head_dim % 2,
|
|
1,
|
|
common::errors::InvalidArgument(
|
|
"The head_dim of input must be a multiple of 2."));
|
|
|
|
auto stream = dev_ctx.stream();
|
|
const T* sin_data = sin.get_ptr() ? sin.get_ptr()->data<T>() : nullptr;
|
|
const T* cos_data = cos.get_ptr() ? cos.get_ptr()->data<T>() : nullptr;
|
|
const int64_t* position_ids_data =
|
|
position_ids.get_ptr() ? position_ids.get_ptr()->data<int64_t>()
|
|
: nullptr;
|
|
|
|
bool flag_sin_cos = (sin_data && cos_data);
|
|
if (flag_sin_cos) {
|
|
auto sin_dims = sin.get_ptr()->dims();
|
|
freqs_head_dim = sin_dims[sin_dims.size() - 1];
|
|
}
|
|
|
|
const int64_t warps_per_block = std::min(num_heads, static_cast<int64_t>(8));
|
|
PADDLE_ENFORCE_LE_UINT32_MAX(seq_len, "fused_rope_grad grid.x");
|
|
PADDLE_ENFORCE_LE_UINT32_MAX(batch_size, "fused_rope_grad grid.y");
|
|
PADDLE_ENFORCE_LE_UINT32_MAX(warps_per_block, "fused_rope_grad block.y");
|
|
dim3 grid(static_cast<uint32_t>(seq_len), static_cast<uint32_t>(batch_size));
|
|
dim3 block(32,
|
|
static_cast<uint32_t>(warps_per_block)); // 32 threads per warp
|
|
size_t shared_mem_size = 2 * head_dim * sizeof(float);
|
|
|
|
// Q
|
|
int64_t stride_s_q = time_major ? dout_q.strides()[0] : dout_q.strides()[1];
|
|
int64_t stride_b_q = time_major ? dout_q.strides()[1] : dout_q.strides()[0];
|
|
int64_t stride_h_q = dout_q.strides()[2];
|
|
int64_t stride_d_q = dout_q.strides()[3];
|
|
|
|
int64_t o_stride_s_q = time_major ? dq->strides()[0] : dq->strides()[1];
|
|
int64_t o_stride_b_q = time_major ? dq->strides()[1] : dq->strides()[0];
|
|
int64_t o_stride_h_q = dq->strides()[2];
|
|
int64_t o_stride_d_q = dq->strides()[3];
|
|
|
|
FusedRopeKernelLauncher(dout_q.data<T>(),
|
|
sin_data,
|
|
cos_data,
|
|
dq->data<T>(),
|
|
FusedRopeGradKernelImpl<T, int>,
|
|
FusedRopeGradKernelImpl<T, int64_t>,
|
|
position_ids_data,
|
|
flag_sin_cos,
|
|
use_neox_rotary_style,
|
|
num_heads,
|
|
head_dim,
|
|
freqs_head_dim,
|
|
stride_s_q,
|
|
stride_b_q,
|
|
stride_h_q,
|
|
stride_d_q,
|
|
o_stride_s_q,
|
|
o_stride_b_q,
|
|
o_stride_h_q,
|
|
o_stride_d_q,
|
|
rotary_emb_base,
|
|
seq_len,
|
|
batch_size,
|
|
numel,
|
|
stream);
|
|
|
|
// K
|
|
if (dk && dk->numel() > 0) {
|
|
auto k_num_heads = dk->dims()[2];
|
|
int64_t stride_s_k =
|
|
time_major ? dout_k->strides()[0] : dout_k->strides()[1];
|
|
int64_t stride_b_k =
|
|
time_major ? dout_k->strides()[1] : dout_k->strides()[0];
|
|
int64_t stride_h_k = dout_k->strides()[2];
|
|
int64_t stride_d_k = dout_k->strides()[3];
|
|
|
|
int64_t o_stride_s_k = time_major ? dk->strides()[0] : dk->strides()[1];
|
|
int64_t o_stride_b_k = time_major ? dk->strides()[1] : dk->strides()[0];
|
|
int64_t o_stride_h_k = dk->strides()[2];
|
|
int64_t o_stride_d_k = dk->strides()[3];
|
|
|
|
FusedRopeKernelLauncher(dout_k->data<T>(),
|
|
sin_data,
|
|
cos_data,
|
|
dk->data<T>(),
|
|
FusedRopeGradKernelImpl<T, int>,
|
|
FusedRopeGradKernelImpl<T, int64_t>,
|
|
position_ids_data,
|
|
flag_sin_cos,
|
|
use_neox_rotary_style,
|
|
k_num_heads,
|
|
head_dim,
|
|
freqs_head_dim,
|
|
stride_s_k,
|
|
stride_b_k,
|
|
stride_h_k,
|
|
stride_d_k,
|
|
o_stride_s_k,
|
|
o_stride_b_k,
|
|
o_stride_h_k,
|
|
o_stride_d_k,
|
|
rotary_emb_base,
|
|
seq_len,
|
|
batch_size,
|
|
numel,
|
|
stream);
|
|
}
|
|
|
|
// V
|
|
if (dv && dv->numel() > 0) {
|
|
auto v_num_heads = dv->dims()[2];
|
|
int64_t stride_s_v =
|
|
time_major ? dout_v->strides()[0] : dout_v->strides()[1];
|
|
int64_t stride_b_v =
|
|
time_major ? dout_v->strides()[1] : dout_v->strides()[0];
|
|
int64_t stride_h_v = dout_v->strides()[2];
|
|
int64_t stride_d_v = dout_v->strides()[3];
|
|
|
|
int64_t o_stride_s_v = time_major ? dv->strides()[0] : dv->strides()[1];
|
|
int64_t o_stride_b_v = time_major ? dv->strides()[1] : dv->strides()[0];
|
|
int64_t o_stride_h_v = dv->strides()[2];
|
|
int64_t o_stride_d_v = dv->strides()[3];
|
|
|
|
FusedRopeKernelLauncher(dout_v->data<T>(),
|
|
sin_data,
|
|
cos_data,
|
|
dv->data<T>(),
|
|
FusedRopeGradKernelImpl<T, int>,
|
|
FusedRopeGradKernelImpl<T, int64_t>,
|
|
position_ids_data,
|
|
flag_sin_cos,
|
|
use_neox_rotary_style,
|
|
v_num_heads,
|
|
head_dim,
|
|
freqs_head_dim,
|
|
stride_s_v,
|
|
stride_b_v,
|
|
stride_h_v,
|
|
stride_d_v,
|
|
o_stride_s_v,
|
|
o_stride_b_v,
|
|
o_stride_h_v,
|
|
o_stride_d_v,
|
|
rotary_emb_base,
|
|
seq_len,
|
|
batch_size,
|
|
numel,
|
|
stream);
|
|
}
|
|
}
|
|
|
|
} // namespace fusion
|
|
} // namespace phi
|
|
|
|
PD_REGISTER_KERNEL(fused_rotary_position_embedding_grad,
|
|
GPU,
|
|
ALL_LAYOUT,
|
|
phi::fusion::FusedRopeGradKernel,
|
|
float,
|
|
double,
|
|
phi::float16,
|
|
phi::bfloat16){};
|