Files
paddlepaddle--paddle/paddle/phi/kernels/gpu/global_scatter_kernel.cu
T
2026-07-13 12:40:42 +08:00

171 lines
6.0 KiB
Plaintext

// 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.
#include "paddle/phi/kernels/gpu/global_scatter_kernel.h"
#include "paddle/phi/core/distributed/utils.h"
#include "paddle/phi/core/kernel_registry.h"
#if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
#include "paddle/phi/core/distributed/nccl_comm_context.h"
#include "paddle/phi/core/platform/device/gpu/nccl_helper.h"
#endif
#include "paddle/phi/core/utils/data_type.h"
namespace phi {
template <typename Context, typename T>
struct GlobalScatterFunctor {
void operator()(const Context& dev_ctx,
const DenseTensor& x_in,
const DenseTensor& local_count_in,
const DenseTensor& global_count_in,
DenseTensor* out);
};
template <typename T>
struct GlobalScatterFunctor<GPUContext, T> {
void operator()(const GPUContext& dev_ctx,
const DenseTensor& x_in,
const DenseTensor& local_count_in,
const DenseTensor& global_count_in,
DenseTensor* out) {
#if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
#if NCCL_VERSION_CODE >= 2703
auto x = &x_in;
auto local_count = &local_count_in;
auto global_count = &global_count_in;
auto local_count_type = local_count->dtype();
auto global_count_type = global_count->dtype();
if (local_count_type != DataType::INT64) {
PADDLE_THROW(common::errors::InvalidArgument(
"Please use int64 type in local_count."));
}
if (global_count_type != DataType::INT64) {
PADDLE_THROW(common::errors::InvalidArgument(
"Please use int64 type in global_count."));
}
const int64_t* cpu_local_count_data;
const int64_t* cpu_global_count_data;
DenseTensor cpu_local_count;
if (local_count->place().GetType() == AllocationType::CPU) {
cpu_local_count_data = local_count->data<int64_t>();
} else {
Copy(dev_ctx, *local_count, CPUPlace(), true, &cpu_local_count);
cpu_local_count_data = cpu_local_count.data<int64_t>();
}
auto global_count_len = 0;
DenseTensor cpu_global_count;
if (global_count->place().GetType() == AllocationType::CPU) {
cpu_global_count_data = global_count->data<int64_t>();
global_count_len = global_count->numel();
} else {
Copy(dev_ctx, *global_count, CPUPlace(), true, &cpu_global_count);
cpu_global_count_data = cpu_global_count.data<int64_t>();
global_count_len = cpu_global_count.numel();
}
ncclDataType_t dtype = ToNCCLDataType(x->dtype());
gpuStream_t stream = nullptr;
stream = dev_ctx.stream();
distributed::NCCLCommContext* comm_ctx = nullptr;
int nranks = 0;
comm_ctx =
static_cast<distributed::NCCLCommContext*>(dev_ctx.GetCommContext());
PADDLE_ENFORCE_NE(comm_ctx,
nullptr,
common::errors::Unavailable(
"NCCLCommContext is nullptr, collective op should "
"has ring_id attr."));
nranks = comm_ctx->GetSize();
auto in_feat = x->dims()[1];
auto n_expert = local_count->dims()[0] / nranks;
int64_t fwd_count = 0;
for (auto i = 0; i < global_count_len; ++i) {
fwd_count += cpu_global_count_data[i];
}
DDim out_dims = make_ddim({fwd_count, in_feat});
int64_t* expert_ptr = new int64_t[n_expert * nranks];
expert_ptr[0] = 0;
auto tot_experts = n_expert * nranks;
for (auto i = 1; i < tot_experts; ++i) {
expert_ptr[i] = expert_ptr[i - 1] + cpu_local_count_data[i - 1];
}
auto recv_ptr = 0;
out->Resize(out_dims);
dev_ctx.template Alloc<T>(out);
for (auto i = 0; i < n_expert; ++i) {
comm_ctx->GroupStart();
for (auto j = 0; j < nranks; ++j) {
int idx = i + j * n_expert;
if (cpu_local_count_data[idx]) {
auto send_buf = distributed::GetPartialTensor(
*x,
expert_ptr[idx] * in_feat,
cpu_local_count_data[idx] * in_feat);
comm_ctx->Send(
send_buf, cpu_local_count_data[idx] * in_feat, j, stream);
}
if (cpu_global_count_data[idx]) {
auto recv_buf = distributed::GetPartialTensor(
*out, recv_ptr * in_feat, cpu_global_count_data[idx] * in_feat);
comm_ctx->Recv(
&recv_buf, cpu_global_count_data[idx] * in_feat, j, stream);
recv_ptr += cpu_global_count_data[idx];
}
}
comm_ctx->GroupEnd();
}
#else
PADDLE_THROW(
common::errors::Unavailable("NCCL version >= 2.7.3 is needed."));
#endif
#else
PADDLE_THROW(
common::errors::Unavailable("PaddlePaddle should compile with GPU."));
#endif
}
};
template <typename T, typename Context>
void GlobalScatterKernel(const Context& dev_ctx,
const DenseTensor& x,
const DenseTensor& local_count,
const DenseTensor& global_count,
DenseTensor* out) {
GlobalScatterFunctor<GPUContext, T> functor_;
functor_(dev_ctx, x, local_count, global_count, out);
}
} // namespace phi
PD_REGISTER_KERNEL(global_scatter,
GPU,
ALL_LAYOUT,
phi::GlobalScatterKernel,
float,
double,
int,
int64_t,
phi::float16) {
kernel->InputAt(1).SetDataType(phi::DataType::INT64);
kernel->InputAt(2).SetDataType(phi::DataType::INT64);
}