distributed_fused_lamb_op.cu

// Copyright (c) 2021 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 <cmath>
#include "paddle/fluid/memory/buffer.h"
#include "paddle/fluid/operators/amp/fp16_type_traits.h"
#include "paddle/fluid/operators/optimizers/cast_with_ptr.h"
#include "paddle/fluid/operators/optimizers/distributed_fused_lamb_op.h"
#include "paddle/fluid/operators/optimizers/multi_tensor_apply.h"
#include "paddle/fluid/operators/tensor_to_string.h"
#include "paddle/fluid/platform/aligned_vector.h"
#include "paddle/fluid/platform/collective_helper.h"
#include "paddle/fluid/platform/for_range.h"
#include "paddle/fluid/string/string_helper.h"
#include "paddle/phi/core/utils/data_type.h"

#ifdef __NVCC__
#include "cub/cub.cuh"
#include "math.h"  // NOLINT
#endif

#ifdef __HIPCC__
#include <hipcub/hipcub.hpp>
#include "math.h"  // NOLINT
namespace cub = hipcub;
#endif

namespace paddle {
namespace operators {

template <typename T>
using MasterT = typename details::MPTypeTrait<T>::Type;

template <typename T>
static void FillZeroWithPtr(T *x, size_t n, gpuStream_t stream) {
  static_assert(!std::is_same<T, void>::value, "T cannot be void.");
#ifdef PADDLE_WITH_HIP
  PADDLE_ENFORCE_GPU_SUCCESS(hipMemsetAsync(x, 0, n * sizeof(T), stream));
#else
  PADDLE_ENFORCE_GPU_SUCCESS(cudaMemsetAsync(x, 0, n * sizeof(T), stream));
#endif
}

template <typename T, int BlockDim, int VecSize>
struct L2NormFunctor {
  DEVICE void operator()(int tensor_id, int chunk_id, int offset, int size,
                         const T *x, MasterT<T> *y, int max_chunk_num) const {
    using MT = MasterT<T>;
    const T *ptr = x + offset;

    using BlockReduce = cub::BlockReduce<MT, BlockDim>;
    __shared__ typename BlockReduce::TempStorage storage;

    MT square_sum = static_cast<MT>(0);
    int i;
    for (i = threadIdx.x * VecSize; i + VecSize <= size;
         i += (BlockDim * VecSize)) {
      platform::AlignedVector<T, VecSize> tmp_vec;
      platform::Load(ptr + i, &tmp_vec);
#pragma unroll
      for (int j = 0; j < VecSize; ++j) {
        auto tmp = static_cast<MT>(tmp_vec[j]);
        square_sum += (tmp * tmp);
      }
    }

    for (; i < size; ++i) {
      auto tmp = static_cast<MT>(ptr[i]);
      square_sum += (tmp * tmp);
    }

    square_sum = BlockReduce(storage).Reduce(square_sum, cub::Sum());
    if (threadIdx.x == 0) {
      y[tensor_id * max_chunk_num + chunk_id] = square_sum;
    }
  }
};

template <typename InT, typename OutT, int BlockDim>
static __global__ void MultiTensorL2NormReduceAgainCUDAKernel(
    const InT *x, OutT *y, int max_chunk_num) {
  int tensor_id = blockIdx.x;
  x += (tensor_id * max_chunk_num);
  using BlockReduce = cub::BlockReduce<InT, BlockDim>;
  __shared__ typename BlockReduce::TempStorage storage;
  InT sum = static_cast<InT>(0);
  for (int i = threadIdx.x; i < max_chunk_num; i += BlockDim) {
    sum += x[i];
  }
  sum = BlockReduce(storage).Reduce(sum, cub::Sum());
  if (threadIdx.x == 0) {
    y[blockIdx.x] = static_cast<OutT>(sum);
  }
}

template <typename T>
static int GetChunkedVecSize(const T *ptr, int chunk_size) {
  static_assert(!std::is_same<T, void>::value, "T cannot be void.");

  constexpr int max_load_bits = 128;
  int valid_vec_size = max_load_bits / CHAR_BIT / sizeof(T);
  auto address = reinterpret_cast<uintptr_t>(ptr);
  constexpr int vec8 = alignof(platform::AlignedVector<T, 8>);
  constexpr int vec4 = alignof(platform::AlignedVector<T, 4>);
  constexpr int vec2 = alignof(platform::AlignedVector<T, 2>);
  chunk_size *= sizeof(T);
  if (address % vec8 == 0 && chunk_size % vec8 == 0) {
    return std::min(8, valid_vec_size);
  } else if (address % vec4 == 0 && chunk_size % vec4 == 0) {
    return std::min(4, valid_vec_size);
  } else if (address % vec2 == 0 && chunk_size % vec2 == 0) {
    return std::min(2, valid_vec_size);
  } else {
    return 1;
  }
}

#define PD_VEC_LAUNCH_KERNEL_CASE(__vec_size, ...) \
  case __vec_size: {                               \
    constexpr int kVecSize = __vec_size;           \
    __VA_ARGS__;                                   \
    break;                                         \
  }

#define PD_VEC_LAUNCH_KERNEL(__vec_size, ...)    \
  do {                                           \
    switch (__vec_size) {                        \
      PD_VEC_LAUNCH_KERNEL_CASE(8, __VA_ARGS__); \
      PD_VEC_LAUNCH_KERNEL_CASE(4, __VA_ARGS__); \
      PD_VEC_LAUNCH_KERNEL_CASE(2, __VA_ARGS__); \
      PD_VEC_LAUNCH_KERNEL_CASE(1, __VA_ARGS__); \
    }                                            \
  } while (0)

// TODO(zengjinle): which chunk_size is better?
template <typename InT, typename OutT, int MaxTensorNumPerLaunch = 160,
          int MaxChunkNumPerLaunch = 780>
static void MultiTensorL2Norm(const platform::CUDAPlace &place,
                              gpuStream_t stream, const InT *x,
                              const int *offsets, int n, OutT *y,
                              int chunk_size = 65536) {
  if (n <= 0) return;

  constexpr int kNumTensor = MaxTensorNumPerLaunch;
  constexpr int kNumChunk = MaxChunkNumPerLaunch;
  constexpr int kBlockDim = 512;

  int max_chunk_num = -1;
  int vec_size = 8;
  int total_chunk_num = 0;
  for (int i = 0; i < n; ++i) {
    vec_size = std::min(
        vec_size, GetChunkedVecSize(x + offsets[i] - offsets[0], chunk_size));
    int length = offsets[i + 1] - offsets[i];
    auto tmp_chunk_num = (length + chunk_size - 1) / chunk_size;
    max_chunk_num = std::max(max_chunk_num, tmp_chunk_num);
    total_chunk_num += tmp_chunk_num;
  }

  VLOG(1) << "MultiTensorL2Norm max_chunk_num = " << max_chunk_num
          << " , total_chunk_num = " << total_chunk_num
          << " , tensor_num = " << n;

  using MT = MasterT<InT>;
  memory::Buffer tmp_out(place);
  auto *tmp_out_ptr = tmp_out.Alloc<MT>(n * max_chunk_num);
  FillZeroWithPtr(tmp_out_ptr, n * max_chunk_num, stream);

#define PD_LAUNCH_MULTI_TENSOR_APPLY_L2_NORM_KERNEL                            \
  do {                                                                         \
    using FunctorT = L2NormFunctor<InT, kBlockDim, kVecSize>;                  \
    VLOG(10) << __func__ << " " << typeid(InT).name()                          \
             << " VecSize = " << kVecSize;                                     \
    MultiTensorApply<FunctorT, kNumTensor, kNumChunk>(                         \
        FunctorT(), stream, offsets, n, chunk_size, kBlockDim, x, tmp_out_ptr, \
        max_chunk_num);                                                        \
  } while (0)

  PD_VEC_LAUNCH_KERNEL(vec_size, PD_LAUNCH_MULTI_TENSOR_APPLY_L2_NORM_KERNEL);
#undef PD_LAUNCH_MULTI_TENSOR_APPLY_L2_NORM_KERNEL

  MultiTensorL2NormReduceAgainCUDAKernel<
      MT, OutT, kBlockDim><<<n, kBlockDim, 0, stream>>>(tmp_out_ptr, y,
                                                        max_chunk_num);
}

template <int LogLevel>
static void LogParamAndTrustRatioDivSquareNorm(
    const framework::ExecutionContext &ctx, const float *param_square_norm,
    const float *trust_ratio_div_square_norm) {
  if (!VLOG_IS_ON(LogLevel)) return;

  auto tensors = ctx.MultiInput<framework::Tensor>("Param");
  if (tensors.empty()) return;

  const auto *order = ctx.Input<framework::Tensor>("ParamOrder")->data<int>();

  size_t n = tensors.size();
  auto place = tensors[0]->place();

  auto pn_vec = ToVector(param_square_norm, n, place);
  auto tn_vec = ToVector(trust_ratio_div_square_norm, n, place);

  const auto &names = ctx.GetOp().Inputs("Param");
  for (size_t i = 0; i < n; ++i) {
    auto idx = order[i];
    VLOG(LogLevel) << "Param " << tensors[idx]->dtype() << " " << names[idx]
                   << " pn = " << pn_vec[i] << " , tn = " << tn_vec[i];
  }
}

static bool IsFinite(const platform::CUDADeviceContext &dev_ctx,
                     const float *ptr) {
  auto stream = dev_ctx.stream();
  float cpu_value;
#ifdef PADDLE_WITH_HIP
  PADDLE_ENFORCE_GPU_SUCCESS(hipMemcpyAsync(&cpu_value, ptr, sizeof(float),
                                            hipMemcpyDeviceToHost, stream));
  PADDLE_ENFORCE_GPU_SUCCESS(hipStreamSynchronize(stream));
#else
  PADDLE_ENFORCE_GPU_SUCCESS(cudaMemcpyAsync(&cpu_value, ptr, sizeof(float),
                                             cudaMemcpyDeviceToHost, stream));
  PADDLE_ENFORCE_GPU_SUCCESS(cudaStreamSynchronize(stream));
#endif
  LOG(INFO) << "NAN_INF indicator value: " << cpu_value;
  return isfinite(cpu_value);
}

template <typename T>
static const T *GetInputTensorPtr(const framework::ExecutionContext &ctx,
                                  const char *in_name,
                                  int64_t *numel = nullptr) {
  const auto *in_tensor = ctx.Input<framework::Tensor>(in_name);
  PADDLE_ENFORCE_NOT_NULL(in_tensor, platform::errors::InvalidArgument(
                                         "Input(%s) cannot be NULL.", in_name));
  if (in_tensor->IsInitialized()) {
    if (numel) *numel = in_tensor->numel();
    return in_tensor->data<T>();
  } else {
    if (numel) *numel = 0;
    return nullptr;
  }
}

template <typename T, bool AllowNotExist = false>
static T *GetSameInOutTensorPtr(const framework::ExecutionContext &ctx,
                                const platform::Place &place,
                                const char *in_name, const char *out_name,
                                int64_t *numel = nullptr) {
  const auto *in_tensor = ctx.Input<framework::Tensor>(in_name);
  if (in_tensor == nullptr || !in_tensor->IsInitialized()) {
    PADDLE_ENFORCE_EQ(AllowNotExist, true,
                      platform::errors::InvalidArgument(
                          "Input(%s) cannot be NULL.", in_name));
    if (numel) *numel = 0;
    return nullptr;
  }

  auto *out_tensor = ctx.Output<framework::Tensor>(out_name);
  PADDLE_ENFORCE_NOT_NULL(in_tensor, platform::errors::InvalidArgument(
                                         "Input(%s) cannot be NULL.", in_name));
  PADDLE_ENFORCE_NOT_NULL(out_tensor,
                          platform::errors::InvalidArgument(
                              "Output(%s) cannot be NULL.", out_name));
  const T *in_data = in_tensor->data<T>();
  T *out_data = out_tensor->mutable_data<T>(place);
  PADDLE_ENFORCE_EQ(in_data, out_data,
                    platform::errors::InvalidArgument(
                        "Input(%s) and Output(%s) must be the same Tensor.",
                        in_name, out_name));
  if (numel) *numel = out_tensor->numel();
  return out_data;
}

template <typename T>
struct SquareFunctor {
  HOSTDEVICE MasterT<T> operator()(T x) const {
    auto y = static_cast<MasterT<T>>(x);
    return y * y;
  }
};

template <typename T>
struct IsNanInfFunctor {
  HOSTDEVICE bool operator()(T x) const { return !isfinite(x); }
};

struct OrFunctor {
  HOSTDEVICE bool operator()(bool x, bool y) const { return x || y; }
};

struct AndFunctor {
  HOSTDEVICE bool operator()(bool x, bool y) const { return x && y; }
};

template <typename T1, typename T2, int VecSize>
static __global__ void ScaleCUDAKernel(const T1 *__restrict__ x,
                                       const T2 *__restrict__ scale,
                                       T1 *__restrict__ y, int num) {
  static_assert(sizeof(T1) <= sizeof(T2),
                "sizeof(T1) must be not greater than sizeof(T2).");
  T2 s = scale[0];

  int i = (threadIdx.x + blockIdx.x * blockDim.x) * VecSize;
  int stride = blockDim.x * gridDim.x * VecSize;

  for (; i + VecSize <= num; i += stride) {
    platform::AlignedVector<T1, VecSize> x_vec;
    platform::AlignedVector<T1, VecSize> y_vec;

    platform::Load(x + i, &x_vec);
#pragma unroll
    for (int j = 0; j < VecSize; ++j) {
      y_vec[j] = static_cast<T1>(static_cast<T2>(x_vec[j]) * s);
    }
    platform::Store(y_vec, y + i);
  }

  for (; i < num; ++i) {
    y[i] = static_cast<T1>(static_cast<T2>(x[i]) * s);
  }
}

template <typename T>
static __global__ void AddToCUDAKernel(const T *__restrict__ x,
                                       T *__restrict__ y) {
  y[0] += x[0];
}

// If clip before allreduce,
// coeff = global_scale * max_global_grad_norm / (1e-6 + sqrt(square_grad_norm)
// * rescale_grad)
// if coeff >= 1 or coeff is Nan/Inf, scale = 1.0
// else scale = coeff
template <typename T1, typename T2>
static __global__ void CalcGradNormClipBeforeAllReduceScale(
    const T1 *__restrict__ global_scale, T1 max_global_grad_norm,
    const T1 *__restrict__ square_grad_norm, T1 *__restrict__ out1,
    T2 *__restrict__ out2, T1 clip_rescale_grad) {
  T1 grad_norm = static_cast<T1>(sqrtf(*square_grad_norm)) * clip_rescale_grad;
  T1 scale = global_scale[0] * max_global_grad_norm / (1e-6 + grad_norm);
  bool found_nan_inf = !isfinite(scale);
  if (scale >= 1 || found_nan_inf) {
    scale = static_cast<T1>(1.0);
  }

  if (out1) {
    *out1 = scale;
  }
  if (out2) {
    *out2 = static_cast<T2>(scale);
  }
}

static __global__ void SetNanInfValueCUDAKernelOneFlag(const bool *in_flag_p,
                                                       float *out_p) {
  *out_p = (*in_flag_p) ? __int_as_float(0x7fffffffU) : 0.0f;
}

static __global__ void SetNanInfValueCUDAKernelTwoFlag(const bool *in_flag_p_1,
                                                       const bool *in_flag_p_2,
                                                       float *out_p) {
  *out_p =
      ((*in_flag_p_1) || (*in_flag_p_2)) ? __int_as_float(0x7fffffffU) : 0.0f;
}

template <typename T, typename GradT, int VecSize>
static __global__ void UpdateLambMomentAndTrustRatioDivCUDAKernel(
    const T *__restrict__ param_p, const GradT *__restrict__ grad_p,
    const T *__restrict__ square_grad_norm_p,
    const T *__restrict__ global_scale, const T *__restrict__ beta1pow_p,
    const T *__restrict__ beta2pow_p, T *__restrict__ mom1_p,
    T *__restrict__ mom2_p, T *__restrict__ trust_ratio_div_p, bool *found_inf,
    T weight_decay, int weight_decay_end_numel, T beta1, T beta2, T epsilon,
    T max_global_grad_norm, int num, T rescale_grad) {
  T square_grad_norm = *square_grad_norm_p;
  bool need_update_found_inf =
      (found_inf && threadIdx.x == 0 && blockIdx.x == 0);
  if (!isfinite(square_grad_norm)) {
    if (need_update_found_inf) *found_inf = true;
    return;
  } else if (need_update_found_inf) {
    *found_inf = false;
  }

  T scale = rescale_grad / global_scale[0];
  if (max_global_grad_norm > 0) {
    T clip_scale =
        max_global_grad_norm / (sqrtf(square_grad_norm) * scale + 1e-6);
    if (clip_scale < static_cast<T>(1)) {
      scale *= clip_scale;
    }
  }

  T one_minus_beta1pow = 1 - beta1pow_p[0];
  T one_minus_beta2pow = 1 - beta2pow_p[0];

  int i = (threadIdx.x + blockIdx.x * blockDim.x) * VecSize;
  int stride = blockDim.x * gridDim.x * VecSize;

  for (; i + VecSize <= num; i += stride) {
    platform::AlignedVector<T, VecSize> param_vec;
    platform::AlignedVector<GradT, VecSize> grad_vec;
    platform::AlignedVector<T, VecSize> mom1_vec;
    platform::AlignedVector<T, VecSize> mom2_vec;
    platform::AlignedVector<T, VecSize> trust_ratio_div_vec;

    T cur_weight_decay = (i < weight_decay_end_numel) * weight_decay;
    if (cur_weight_decay != static_cast<T>(0.0)) {
      platform::Load(param_p + i, &param_vec);
    } else {
#pragma unroll
      for (int j = 0; j < VecSize; ++j) {
        param_vec[j] = static_cast<T>(0);
      }
    }
    platform::Load(grad_p + i, &grad_vec);
    platform::Load(mom1_p + i, &mom1_vec);
    platform::Load(mom2_p + i, &mom2_vec);

#define PD_LAMB_MOM_TRUST_RATIO_DIV_UPDATE(__param, __grad, __mom1, __mom2,    \
                                           __trust_ratio_div, __idx)           \
  T p = __param[__idx];                                                        \
  T g = static_cast<T>(__grad[__idx]) * scale;                                 \
  T mom1 = __mom1[__idx];                                                      \
  T mom2 = __mom2[__idx];                                                      \
  mom1 = beta1 * mom1 + (1 - beta1) * g;                                       \
  mom2 = beta2 * mom2 + (1 - beta2) * g * g;                                   \
  T mom1_unbiased = mom1 / one_minus_beta1pow;                                 \
  T mom2_unbiased = mom2 / one_minus_beta2pow;                                 \
  __trust_ratio_div[__idx] =                                                   \
      mom1_unbiased / (sqrtf(mom2_unbiased) + epsilon) + cur_weight_decay * p; \
  __mom1[__idx] = mom1;                                                        \
  __mom2[__idx] = mom2;

#pragma unroll
    for (int j = 0; j < VecSize; ++j) {
      PD_LAMB_MOM_TRUST_RATIO_DIV_UPDATE(param_vec, grad_vec, mom1_vec,
                                         mom2_vec, trust_ratio_div_vec, j);
    }

    platform::Store(mom1_vec, mom1_p + i);
    platform::Store(mom2_vec, mom2_p + i);
    platform::Store(trust_ratio_div_vec, trust_ratio_div_p + i);
  }

  for (; i < num; ++i) {
    T cur_weight_decay = (i < weight_decay_end_numel) * weight_decay;
    PD_LAMB_MOM_TRUST_RATIO_DIV_UPDATE(param_p, grad_p, mom1_p, mom2_p,
                                       trust_ratio_div_p, i);
  }
}

template <typename T, typename GradT>
static void MultiTensorUpdateLambMomentAndTrustRatioDiv(
    const platform::CUDADeviceContext &dev_ctx, const int *offsets, int n,
    const T *param_p, const GradT *grad_p, const T *square_grad_norm_p,
    const T *global_scale, const T *beta1pow_p, const T *beta2pow_p, T *mom1_p,
    T *mom2_p, T *trust_ratio_div_p, bool *found_inf_p, T weight_decay,
    int weight_decay_end_idx, T beta1, T beta2, T epsilon,
    T max_global_grad_norm, T rescale_grad) {
  if (n <= 0) return;
  int numel = offsets[n] - offsets[0];
  PADDLE_ENFORCE_GE(weight_decay_end_idx, 0,
                    platform::errors::InvalidArgument(
                        "The weight decay end index should be >= 0."));
  PADDLE_ENFORCE_LE(weight_decay_end_idx, n,
                    platform::errors::InvalidArgument(
                        "The weight decay end index should be < %d.", n));
  auto weight_decay_end_numel = offsets[weight_decay_end_idx] - offsets[0];

  int vec_size = GetChunkedVecSize(param_p, 0);
  vec_size = std::min(vec_size, GetChunkedVecSize(grad_p, 0));
  vec_size = std::min(vec_size, GetChunkedVecSize(mom1_p, 0));
  vec_size = std::min(vec_size, GetChunkedVecSize(mom2_p, 0));
  vec_size = std::min(vec_size, GetChunkedVecSize(trust_ratio_div_p, 0));
  for (int i = 0; i < n; ++i) {
    auto length = offsets[i + 1] - offsets[i];
    while (length % vec_size != 0) {
      vec_size /= 2;
    }
  }

  VLOG(1) << __func__ << " VecSize = " << vec_size;

  auto stream = dev_ctx.stream();
  auto config = platform::GetGpuLaunchConfig1D(dev_ctx, numel, vec_size);

#define PD_LAUNCH_LAMB_MOM_TRUST_RATIO_DIV_KERNEL                      \
  do {                                                                 \
    UpdateLambMomentAndTrustRatioDivCUDAKernel<T, GradT, kVecSize><<<  \
        config.block_per_grid, config.thread_per_block, 0, stream>>>(  \
        param_p, grad_p, square_grad_norm_p, global_scale, beta1pow_p, \
        beta2pow_p, mom1_p, mom2_p, trust_ratio_div_p, found_inf_p,    \
        weight_decay, weight_decay_end_numel, beta1, beta2, epsilon,   \
        max_global_grad_norm, numel, rescale_grad);                    \
  } while (0)

  PD_VEC_LAUNCH_KERNEL(vec_size, PD_LAUNCH_LAMB_MOM_TRUST_RATIO_DIV_KERNEL);
#undef PD_LAUNCH_LAMB_MOM_TRUST_RATIO_DIV_KERNEL
}

template <typename T, bool NeedUpdate /*=true*/>
struct LambBetaPowUpdateOnceHelper {
  LambBetaPowUpdateOnceHelper(T *beta1pow, T *beta2pow, T beta1, T beta2) {
    PADDLE_ENFORCE_NOT_NULL(beta1pow,
                            platform::errors::InvalidArgument(
                                "The beta1pow should not be nullptr."));
    PADDLE_ENFORCE_NOT_NULL(beta2pow,
                            platform::errors::InvalidArgument(
                                "The beta2pow should not be nullptr."));
    beta1pow_ = beta1pow;
    beta2pow_ = beta2pow;
    beta1_ = beta1;
    beta2_ = beta2;
  }

  HOSTDEVICE void UpdateBetaPows() const {
    beta1pow_[0] *= beta1_;
    beta2pow_[0] *= beta2_;
  }

 private:
  T *__restrict__ beta1pow_;
  T *__restrict__ beta2pow_;
  T beta1_;
  T beta2_;
};

template <typename T>
struct LambBetaPowUpdateOnceHelper<T, false> {
  LambBetaPowUpdateOnceHelper(T *beta1pow, T *beta2pow, T beta1, T beta2) {
    PADDLE_ENFORCE_EQ(
        beta1pow, nullptr,
        platform::errors::InvalidArgument("The beta1pow should be nullptr."));
    PADDLE_ENFORCE_EQ(
        beta2pow, nullptr,
        platform::errors::InvalidArgument("The beta2pow should be nullptr."));
  }

  HOSTDEVICE void UpdateBetaPows() const {}
};

template <typename T, bool HasMasterParam /*=true*/>
struct LambParamHelper {
  LambParamHelper(T *param, MasterT<T> *master_param) {
    constexpr bool kIsSameType = std::is_same<T, MasterT<T>>::value;
    PADDLE_ENFORCE_EQ(kIsSameType, false,
                      platform::errors::InvalidArgument(
                          "T must not be the same with MasterT<T>."));
    PADDLE_ENFORCE_NOT_NULL(master_param,
                            platform::errors::InvalidArgument(
                                "Master parameter must be provided."));
    param_ = param;
    master_param_ = master_param;
  }

  HOSTDEVICE T *__restrict__ ParamPtr() { return param_; }

  HOSTDEVICE MasterT<T> *__restrict__ MasterParamPtr() { return master_param_; }

 private:
  T *__restrict__ param_;
  MasterT<T> *__restrict__ master_param_;
};

template <typename T>
struct LambParamHelper<T, false> {
  LambParamHelper(T *param, MasterT<T> *master_param) {
    constexpr bool kIsSameType = std::is_same<T, MasterT<T>>::value;
    PADDLE_ENFORCE_EQ(kIsSameType, true,
                      platform::errors::InvalidArgument(
                          "T must be the same with MasterT<T>."));
    if (master_param != nullptr) {
      PADDLE_ENFORCE_EQ(static_cast<void *>(param),
                        static_cast<void *>(master_param),
                        platform::errors::InvalidArgument(
                            "Master parameter must be nullptr or the same as "
                            "non-master parameter."));
    }
    param_ = param;
  }

  HOSTDEVICE T *__restrict__ ParamPtr() { return param_; }

  HOSTDEVICE constexpr MasterT<T> *MasterParamPtr() { return nullptr; }

 private:
  T *__restrict__ param_;
};

template <typename ParamT, bool HasMasterParam, bool NeedUpdateBetaPow,
          int VecSize>
struct LambUpdateParamAndBetaPowsFunctor {
  DEVICE void operator()(
      int tensor_id, int chunk_id, int offset, int size,
      LambParamHelper<ParamT, HasMasterParam> param_helper,
      const MasterT<ParamT> *trust_ratio_div, const MasterT<ParamT> *lr,
      const MasterT<ParamT> *param_square_norm,
      const MasterT<ParamT> *trust_ratio_div_square_norm, const bool *found_inf,
      LambBetaPowUpdateOnceHelper<MasterT<ParamT>, NeedUpdateBetaPow>
          betapow_helper) const {
    if (*found_inf) return;

    using MT = MasterT<ParamT>;

    MT p_square_norm = param_square_norm[tensor_id];
    MT t_square_norm = trust_ratio_div_square_norm[tensor_id];
    MT lr_value = *lr;
    MT ratio = (p_square_norm != static_cast<MT>(0) &&
                        t_square_norm != static_cast<MT>(0)
                    ? lr_value * sqrtf(p_square_norm / t_square_norm)
                    : lr_value);

    int i;
    int stride = blockDim.x * VecSize;

    ParamT *param = param_helper.ParamPtr() + offset;
    MT *master_param = HasMasterParam ? param_helper.MasterParamPtr() + offset
                                      : param_helper.MasterParamPtr();
    trust_ratio_div += offset;

    for (i = threadIdx.x * VecSize; i + VecSize <= size; i += stride) {
      platform::AlignedVector<MT, VecSize> trust_ratio_div_vec;
      platform::Load(trust_ratio_div + i, &trust_ratio_div_vec);
      if (HasMasterParam) {
        platform::AlignedVector<MT, VecSize> master_param_vec;
        platform::Load(master_param + i, &master_param_vec);
        platform::AlignedVector<ParamT, VecSize> param_vec;
#pragma unroll
        for (int j = 0; j < VecSize; ++j) {
          MT p = master_param_vec[j] - ratio * trust_ratio_div_vec[j];
          master_param_vec[j] = p;
          param_vec[j] = static_cast<ParamT>(p);
        }
        platform::Store(master_param_vec, master_param + i);
        platform::Store(param_vec, param + i);
      } else {
        platform::AlignedVector<ParamT, VecSize> param_vec;
        platform::Load(param + i, &param_vec);
#pragma unroll
        for (int j = 0; j < VecSize; ++j) {
          MT p = static_cast<MT>(param_vec[j]) - ratio * trust_ratio_div_vec[j];
          param_vec[j] = static_cast<ParamT>(p);
        }
        platform::Store(param_vec, param + i);
      }
    }

    for (; i < size; ++i) {
      if (HasMasterParam) {
        MT p = master_param[i] - ratio * trust_ratio_div[i];
        master_param[i] = p;
        param[i] = static_cast<ParamT>(p);
      } else {
        MT p = static_cast<MT>(param[i]) - ratio * trust_ratio_div[i];
        param[i] = static_cast<ParamT>(p);
      }
    }

    if (NeedUpdateBetaPow && threadIdx.x == 0 && blockIdx.x == 0) {
      betapow_helper.UpdateBetaPows();
    }
  }
};

// TODO(zengjinle): which block_dim and chunk_size would be better?
template <typename ParamT, int MaxTensorNumPerLaunch = 160,
          int MaxChunkNumPerLaunch = 780>
static void MultiTensorUpdateLambParamAndBetaPows(
    const platform::CUDADeviceContext &dev_ctx, const int *offsets, int n,
    const MasterT<ParamT> *trust_ratio_div, const MasterT<ParamT> *lr,
    const MasterT<ParamT> *param_square_norm,
    const MasterT<ParamT> *trust_ratio_div_square_norm, const bool *found_inf,
    ParamT *param, MasterT<ParamT> *master_param, MasterT<ParamT> *beta1pow,
    MasterT<ParamT> *beta2pow, MasterT<ParamT> beta1, MasterT<ParamT> beta2,
    int chunk_size = 65536) {
  constexpr bool kHasMasterParam =
      !(std::is_same<ParamT, MasterT<ParamT>>::value);

  bool has_beta_pow = (beta1pow != nullptr);
  if (has_beta_pow) {
    PADDLE_ENFORCE_NOT_NULL(beta2pow, platform::errors::InvalidArgument(
                                          "Beta2Pow should not be nullptr."));
  } else {
    PADDLE_ENFORCE_EQ(beta2pow, nullptr, platform::errors::InvalidArgument(
                                             "Beta2Pow should be nullptr."));
  }

  const int block_dim = 512;

  int vec_size = 8;
  for (int i = 0; i < n; ++i) {
    int offset = offsets[i] - offsets[0];
    vec_size =
        std::min(vec_size, GetChunkedVecSize(param + offset, chunk_size));
    if (kHasMasterParam) {
      vec_size = std::min(vec_size,
                          GetChunkedVecSize(master_param + offset, chunk_size));
    }
    vec_size = std::min(
        vec_size, GetChunkedVecSize(trust_ratio_div + offset, chunk_size));
  }

  VLOG(1) << __func__ << " VecSize = " << vec_size;

  constexpr auto kNumTensor = MaxTensorNumPerLaunch;
  constexpr auto kNumChunk = MaxChunkNumPerLaunch;

  auto stream = dev_ctx.stream();
#define PD_LAUNCH_MULTI_TENSOR_UPDATE_PARAM_BETAPOW(__has_beta_pow)            \
  do {                                                                         \
    using FunctorT =                                                           \
        LambUpdateParamAndBetaPowsFunctor<ParamT, kHasMasterParam,             \
                                          __has_beta_pow, kVecSize>;           \
    LambParamHelper<ParamT, kHasMasterParam> param_helper(param,               \
                                                          master_param);       \
    LambBetaPowUpdateOnceHelper<MasterT<ParamT>, __has_beta_pow>               \
        betapow_helper(beta1pow, beta2pow, beta1, beta2);                      \
    launcher.Launch(FunctorT(), param_helper, trust_ratio_div, lr,             \
                    param_square_norm, trust_ratio_div_square_norm, found_inf, \
                    betapow_helper);                                           \
  } while (0)

#define PD_LAUNCH_VEC_MULTI_TENSOR_UPDATE_PARAM_BETAPOW_CASE        \
  do {                                                              \
    auto callback = [&](                                            \
        const MultiTensorLauncher<kNumTensor, kNumChunk> &launcher, \
        int launch_n) {                                             \
      if (has_beta_pow && launch_n == 0) {                          \
        PD_LAUNCH_MULTI_TENSOR_UPDATE_PARAM_BETAPOW(true);          \
        beta1pow = nullptr;                                         \
        beta2pow = nullptr;                                         \
      } else {                                                      \
        PD_LAUNCH_MULTI_TENSOR_UPDATE_PARAM_BETAPOW(false);         \
      }                                                             \
    };                                                              \
    MultiTensorApplyWithCallback<kNumTensor, kNumChunk>(            \
        stream, offsets, n, chunk_size, block_dim, callback);       \
  } while (0)

  PD_VEC_LAUNCH_KERNEL(vec_size,
                       PD_LAUNCH_VEC_MULTI_TENSOR_UPDATE_PARAM_BETAPOW_CASE);

#undef PD_LAUNCH_MULTI_TENSOR_UPDATE_PARAM_BETAPOW
#undef PD_LAUNCH_VEC_MULTI_TENSOR_UPDATE_PARAM_BETAPOW_CASE
}

#if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
static bool CreatePreMulScaleOpIfSupported(ncclDataType_t dtype,
                                           ncclComm_t comm, const void *scale,
                                           ncclRedOp_t *op) {
#if NCCL_VERSION_CODE >= 21100
  int ver;
  PADDLE_ENFORCE_GPU_SUCCESS(platform::dynload::ncclGetVersion(&ver));
  if (ver >= 21100) {
    VLOG(10) << "ncclRedOpCreatePreMulSum is supported.";
    PADDLE_ENFORCE_GPU_SUCCESS(platform::dynload::ncclRedOpCreatePreMulSum(
        op, const_cast<void *>(scale), dtype, ncclScalarDevice, comm));
    return true;
  }
#endif
  VLOG(10) << "ncclRedOpCreatePreMulSum is not supported.";
  return false;
}

template <typename T1, typename T2>
static void LaunchScaleKernel(const platform::CUDADeviceContext &dev_ctx,
                              const T1 *x, const T2 *scale, T1 *y, int n,
                              gpuStream_t stream) {
  int vec_size = std::min(GetChunkedVecSize(x, 0), GetChunkedVecSize(y, 0));
  auto config = platform::GetGpuLaunchConfig1D(dev_ctx, n, vec_size);

#define PD_LAMB_VEC_SCALE_KERNEL_CASE                                          \
  do {                                                                         \
    ScaleCUDAKernel<T1, T2, kVecSize><<<config.block_per_grid,                 \
                                        config.thread_per_block, 0, stream>>>( \
        x, scale, y, n);                                                       \
  } while (0)

  PD_VEC_LAUNCH_KERNEL(vec_size, PD_LAMB_VEC_SCALE_KERNEL_CASE);
#undef PD_LAMB_VEC_SCALE_KERNEL_CASE
}

template <typename T>
static void NCCLReduceScatterWithScale(
    const T *sendbuff, T *recvbuff, size_t recvcount, size_t nranks,
    ncclComm_t comm, gpuStream_t stream,
    const platform::CUDADeviceContext &dev_ctx, const T *scale = nullptr) {
  static_assert(std::is_same<T, float>::value ||
                    std::is_same<T, platform::float16>::value,
                "T must be either float32 or float16.");
  if (recvcount == 0) return;

  if (comm == nullptr) {
    if (scale != nullptr) {
      PADDLE_ENFORCE_EQ(nranks, 1,
                        platform::errors::InvalidArgument(
                            "nranks must be 1 when scale != nullptr."));
      LaunchScaleKernel(dev_ctx, sendbuff, scale, recvbuff, recvcount * nranks,
                        stream);
    }
    return;
  }

  ncclRedOp_t op = ncclSum;
  ncclDataType_t dtype =
      std::is_same<T, float>::value ? ncclFloat32 : ncclFloat16;
  bool should_destroy_op =
      scale && CreatePreMulScaleOpIfSupported(dtype, comm, scale, &op);
  memory::Buffer buffer(dev_ctx.GetPlace());
  if (scale && !should_destroy_op) {
    size_t numel = recvcount * nranks;
    T *new_sendbuff = buffer.Alloc<T>(numel);
    LaunchScaleKernel(dev_ctx, sendbuff, scale, new_sendbuff, numel, stream);
    sendbuff = new_sendbuff;
  }

  PADDLE_ENFORCE_GPU_SUCCESS(platform::dynload::ncclReduceScatter(
      sendbuff, recvbuff, recvcount, dtype, op, comm, stream));

#if NCCL_VERSION_CODE >= 21100
  if (should_destroy_op) {
    VLOG(10) << "ncclRedOpDestroy starts";
    PADDLE_ENFORCE_GPU_SUCCESS(platform::dynload::ncclRedOpDestroy(op, comm));
    VLOG(10) << "ncclRedOpDestroy ends";
  }
#endif
}
#endif

template <typename InputIteratorT, typename OutputIteratorT, typename ReduceOpT,
          typename T>
static void CubDeviceReduce(InputIteratorT d_in, OutputIteratorT d_out,
                            int num_items, ReduceOpT reduction_op, T init,
                            gpuStream_t stream, memory::Buffer *buffer) {
  void *d_temp_storage = nullptr;
  size_t temp_storage_bytes = 0;
  PADDLE_ENFORCE_GPU_SUCCESS(
      cub::DeviceReduce::Reduce(d_temp_storage, temp_storage_bytes, d_in, d_out,
                                num_items, reduction_op, init, stream));
  d_temp_storage = buffer->Alloc<void>(temp_storage_bytes);
  VLOG(10) << "cub::DeviceReduce::Reduce needs " << temp_storage_bytes
           << " byte(s), ptr = " << d_temp_storage;
  PADDLE_ENFORCE_GPU_SUCCESS(
      cub::DeviceReduce::Reduce(d_temp_storage, temp_storage_bytes, d_in, d_out,
                                num_items, reduction_op, init, stream));
}

template <typename T>
static void GetSquareGradNormImpl(const T *grad, int n, float *square_norm,
                                  gpuStream_t stream,
                                  memory::Buffer *cub_tmp_buffer) {
  using Iterator =
      cub::TransformInputIterator<float, SquareFunctor<T>, const T *>;
  Iterator iter(grad, SquareFunctor<T>());
  CubDeviceReduce(iter, square_norm, n, cub::Sum(), static_cast<float>(0),
                  stream, cub_tmp_buffer);
}

// square_norm is of length 2 at least
static void GetSquareGradNorm(const float *fp32_grad, int fp32_numel,
                              const platform::float16 *fp16_grad,
                              int fp16_numel, float *square_norm,
                              gpuStream_t stream,
                              memory::Buffer *cub_tmp_buffer) {
  VLOG(10) << "GetSquareGradNorm starts, fp32_numel = " << fp32_numel
           << " , fp16_numel = " << fp16_numel;
  if (fp32_numel > 0) {
    GetSquareGradNormImpl(fp32_grad, fp32_numel, square_norm, stream,
                          cub_tmp_buffer);
    VLOG(10) << "FP32 square L2-Norm: "
             << FlattenToString(square_norm, 1, cub_tmp_buffer->GetPlace());
  }

  if (fp16_numel > 0) {
    float *fp16_square_norm = fp32_numel > 0 ? square_norm + 1 : square_norm;
    GetSquareGradNormImpl(fp16_grad, fp16_numel, fp16_square_norm, stream,
                          cub_tmp_buffer);
    VLOG(10) << "FP16 square L2-Norm: "
             << FlattenToString(fp16_square_norm, 1,
                                cub_tmp_buffer->GetPlace());
    if (fp32_numel > 0) {
      AddToCUDAKernel<<<1, 1, 0, stream>>>(fp16_square_norm, square_norm);
      VLOG(10) << "FP32+FP16 square L2-Norm: "
               << FlattenToString(square_norm, 1, cub_tmp_buffer->GetPlace());
    }
  }
  VLOG(10) << "GetSquareGradNorm ends, fp32_numel = " << fp32_numel
           << " , fp16_numel = " << fp16_numel;
}

template <typename T>
std::string NumToString(T x) {
  std::stringstream ss;
  ss << x;
  return ss.str();
}

template <typename T>
static std::string GetMinMaxStr(const T *x, size_t n,
                                const platform::Place &place) {
  PADDLE_ENFORCE_EQ(
      platform::is_gpu_place(place), true,
      platform::errors::InvalidArgument("Only support CUDAPlace currently."));

  auto *dev_ctx = static_cast<platform::CUDADeviceContext *>(
      platform::DeviceContextPool::Instance().Get(place));
  auto stream = dev_ctx->stream();

  memory::Buffer ret_buffer(place);
  T *ret = ret_buffer.Alloc<T>(2);

  if (n > 0) {
    memory::Buffer cub_buffer(place);
    CubDeviceReduce(x, ret, n, cub::Min(), std::numeric_limits<T>::max(),
                    stream, &cub_buffer);
    CubDeviceReduce(x, ret + 1, n, cub::Max(), std::numeric_limits<T>::lowest(),
                    stream, &cub_buffer);
    T ret_cpu[2];
#ifdef PADDLE_WITH_HIP
    PADDLE_ENFORCE_GPU_SUCCESS(hipMemcpyAsync(&ret_cpu[0], ret, 2 * sizeof(T),
                                              hipMemcpyDeviceToHost, stream));
    PADDLE_ENFORCE_GPU_SUCCESS(hipStreamSynchronize(stream));
#else
    PADDLE_ENFORCE_GPU_SUCCESS(cudaMemcpyAsync(&ret_cpu[0], ret, 2 * sizeof(T),
                                               cudaMemcpyDeviceToHost, stream));
    PADDLE_ENFORCE_GPU_SUCCESS(cudaStreamSynchronize(stream));
#endif
    return std::string("{\"min\": ") + NumToString(ret_cpu[0]) +
           " , \"max\": " + NumToString(ret_cpu[1]) + "}";
  } else {
    return "{\"min\": null, \"max\": null}";
  }
}

struct VisitDTypeFunctor {
  VisitDTypeFunctor(const framework::Tensor *x, std::string *s)
      : x_(x), s_(s) {}

  template <typename T>
  void apply() const {
    *s_ = GetMinMaxStr<T>(x_->template data<T>(), x_->numel(), x_->place());
  }

 private:
  const framework::Tensor *x_;
  std::string *s_;
};

static std::string GetMinMaxStr(const framework::Tensor *x) {
  if (x == nullptr) return "null";
  if (!x->IsInitialized()) return "not_inited";
  if (!platform::is_gpu_place(x->place())) return "CPUTensor";
  std::string str;
  VisitDTypeFunctor functor(x, &str);
  phi::VisitDataType(x->dtype(), functor);
  return str;
}

static void PrintAllMinMaxRange(const framework::ExecutionContext &ctx,
                                bool only_inputs) {
  if (!VLOG_IS_ON(1)) return;
  for (const auto &pair : ctx.GetOp().Inputs()) {
    const auto &key = pair.first;
    const auto tensors = ctx.MultiInput<framework::Tensor>(key);
    size_t n = tensors.size();
    for (size_t i = 0; i < n; ++i) {
      VLOG(1) << "Input(" << key + ")[" << i << "] = " << pair.second[i]
              << " , " << GetMinMaxStr(tensors[i]);
    }
  }

  if (only_inputs) return;
  for (const auto &pair : ctx.GetOp().Outputs()) {
    const auto &key = pair.first;
    const auto tensors = ctx.MultiOutput<framework::Tensor>(key);
    size_t n = tensors.size();
    for (size_t i = 0; i < n; ++i) {
      VLOG(1) << "Output(" << key + ")[" << i << "] = " << pair.second[i]
              << " , " << GetMinMaxStr(tensors[i]);
    }
  }
}

static void CheckHasNanInfGrad(const float *fp32_grad, int fp32_numel,
                               const platform::float16 *fp16_grad,
                               int fp16_numel, float *nan_inf_flag,
                               gpuStream_t stream,
                               memory::Buffer *cub_tmp_buffer) {
  bool *fp32_has_nan_inf = nullptr;
  bool *fp16_has_nan_inf = nullptr;
  if (fp32_numel > 0) {
    fp32_has_nan_inf = reinterpret_cast<bool *>(nan_inf_flag + 1);
    cub::TransformInputIterator<bool, IsNanInfFunctor<float>, const float *>
    iter(fp32_grad, IsNanInfFunctor<float>());
    CubDeviceReduce(iter, fp32_has_nan_inf, fp32_numel, OrFunctor(), false,
                    stream, cub_tmp_buffer);
  }

  if (fp16_numel > 0) {
    fp16_has_nan_inf = reinterpret_cast<bool *>(nan_inf_flag + 1) + 1;
    cub::TransformInputIterator<bool, IsNanInfFunctor<platform::float16>,
                                const platform::float16 *>
        iter(fp16_grad, IsNanInfFunctor<platform::float16>());
    CubDeviceReduce(iter, fp16_has_nan_inf, fp16_numel, OrFunctor(), false,
                    stream, cub_tmp_buffer);
  }

  if (fp32_has_nan_inf && fp16_has_nan_inf) {
    SetNanInfValueCUDAKernelTwoFlag<<<1, 1, 0, stream>>>(
        fp32_has_nan_inf, fp16_has_nan_inf, nan_inf_flag);
  } else if (fp32_has_nan_inf) {
    SetNanInfValueCUDAKernelOneFlag<<<1, 1, 0, stream>>>(fp32_has_nan_inf,
                                                         nan_inf_flag);
  } else {
    SetNanInfValueCUDAKernelOneFlag<<<1, 1, 0, stream>>>(fp16_has_nan_inf,
                                                         nan_inf_flag);
  }
}

template <typename T>
class DistributedFusedLambOpKernel<platform::CUDADeviceContext, T>
    : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext &ctx) const override {
#if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
    auto &dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
    auto stream = dev_ctx.stream();
    auto place = dev_ctx.GetPlace();

    // Step 1: Get fp16 param and grad tensors
    int64_t fp16_numel;
    auto *fp16_param = GetSameInOutTensorPtr<platform::float16, true>(
        ctx, place, "FP16FusedParam", "FP16FusedParamOut", &fp16_numel);
    bool has_fp16_param = (fp16_numel > 0);
    const platform::float16 *fp16_grad = nullptr;
    if (has_fp16_param) {
      fp16_grad = GetInputTensorPtr<platform::float16>(ctx, "FP16FusedGrad");
    } else {
      fp16_param = nullptr;
    }

    // Step 2: Get fp32 param and grad tensors
    int64_t fp32_numel = 0;
    auto *fp32_param = GetSameInOutTensorPtr<float, true>(
        ctx, place, "FP32FusedParam", "FP32FusedParamOut", &fp32_numel);
    PADDLE_ENFORCE_GE(fp32_numel, fp16_numel,
                      platform::errors::InvalidArgument(
                          "The element number in FP32FusedParam should be not "
                          "less than FP16FusedParam."));

    fp32_numel -= fp16_numel;  // the FP32FusedParam contains fp32 param and
                               // fp16 master weight
    bool has_fp32_param = (fp32_numel > 0);
    const float *fp32_grad = nullptr;
    if (has_fp32_param) {
      fp32_grad = GetInputTensorPtr<float>(ctx, "FP32FusedGrad");
    } else {
      PADDLE_ENFORCE_EQ(
          has_fp16_param, true,
          platform::errors::InvalidArgument(
              "Either FP32FusedGrad or FP16FusedGrad cannot be NULL."));
    }

    auto numel = fp32_numel + fp16_numel;
    VLOG(1) << "numel = " << numel << " , fp32_numel = " << fp32_numel
            << " , fp16_numel = " << fp16_numel;

    // The NVIDIA cub library does not support number > INT32_MAX
    PADDLE_ENFORCE_LE(numel, std::numeric_limits<int>::max(),
                      platform::errors::Unimplemented(
                          "Too many parameter number. Only <= %d is supported.",
                          std::numeric_limits<int>::max()));

    // Step 3: Get ParamInfo
    const auto *param_info_tensor = GetInputTensorPtr<int>(ctx, "ParamInfo");
    auto fp32_local_start_idx = param_info_tensor[0];
    auto fp32_local_param_num = param_info_tensor[1];
    auto fp32_global_param_num = param_info_tensor[2];
    auto fp32_weight_decay_end_idx = param_info_tensor[3];
    auto fp16_local_start_idx = param_info_tensor[4];
    auto fp16_local_param_num = param_info_tensor[5];
    auto fp16_global_param_num = param_info_tensor[6];
    auto fp16_weight_decay_end_idx = param_info_tensor[7];

    auto local_param_num = fp32_local_param_num + fp16_local_param_num;
    auto param_num = fp32_global_param_num + fp16_global_param_num;
    PADDLE_ENFORCE_LE(local_param_num, param_num,
                      platform::errors::InvalidArgument(
                          "The local parameter number should not exceed the "
                          "global parameter number."));
    VLOG(1) << "local_param_num = " << local_param_num
            << " , global_param_num = " << param_num
            << " , fp32_local_start_idx = " << fp32_local_start_idx
            << " , fp32_local_param_num = " << fp32_local_param_num
            << " , fp32_global_param_num = " << fp32_global_param_num
            << " , fp16_local_start_idx = " << fp16_local_start_idx
            << " , fp16_local_param_num = " << fp16_local_param_num
            << " , fp16_global_param_num = " << fp16_global_param_num;

    // Step 4: Get LearningRate, Moment1, Moment2, Beta1Pow, Beta2Pow,
    // GlobalScale, FoundInf
    const auto *global_scale = GetInputTensorPtr<float>(ctx, "GlobalScale");
    const auto *lr = GetInputTensorPtr<float>(ctx, "LearningRate");
    int64_t partial_numel = 0;
    auto *moment1 = GetSameInOutTensorPtr<float>(ctx, place, "Moment1",
                                                 "Moment1Out", &partial_numel);

    PADDLE_ENFORCE_EQ(numel % partial_numel, 0,
                      platform::errors::InvalidArgument(
                          "The total parameter number %d should be divided "
                          "exactly by the element number %d of Moment1.",
                          numel, partial_numel));

    int64_t num_devices = numel / partial_numel;
    VLOG(1) << "num_devices = " << num_devices
            << " , partial_numel = " << partial_numel;

    PADDLE_ENFORCE_EQ(fp32_numel % num_devices, 0,
                      platform::errors::InvalidArgument(
                          "The fp32 parameter number %d should be divided "
                          "exactly by the device number %d.",
                          fp32_numel, num_devices));
    PADDLE_ENFORCE_EQ(fp16_numel % num_devices, 0,
                      platform::errors::InvalidArgument(
                          "The fp16 parameter number %d should be divided "
                          "exactly by the device number %d.",
                          fp16_numel, num_devices));

    auto *moment2 =
        GetSameInOutTensorPtr<float>(ctx, place, "Moment2", "Moment2Out");
    auto *beta1pow =
        GetSameInOutTensorPtr<float>(ctx, place, "Beta1Pow", "Beta1PowOut");
    auto *beta2pow =
        GetSameInOutTensorPtr<float>(ctx, place, "Beta2Pow", "Beta2PowOut");

    auto *found_inf_t = ctx.Output<framework::Tensor>("FoundInf");
    found_inf_t->Resize({1});
    auto *found_inf = found_inf_t->mutable_data<bool>(place);

    // Step 5: Get attributes weight_decay, beta1, beta2, epsilon,
    // max_grad_norm, ring_id,
    // use_master_param_norm, is_grad_scaled_by_nranks
    auto weight_decay = ctx.Attr<float>("weight_decay");
    auto beta1 = ctx.Attr<float>("beta1");
    auto beta2 = ctx.Attr<float>("beta2");
    auto epsilon = ctx.Attr<float>("epsilon");
    auto max_global_grad_norm = ctx.Attr<float>("max_global_grad_norm");
    auto clip_after_allreduce = ctx.Attr<bool>("clip_after_allreduce");
    auto ring_id = ctx.Attr<int>("ring_id");
    auto use_master_param_norm = ctx.Attr<bool>("use_master_param_norm");
    auto is_grad_scaled_by_nranks = ctx.Attr<bool>("is_grad_scaled_by_nranks");
    VLOG(10) << "max_global_grad_norm = " << max_global_grad_norm
             << " , clip_after_allreduce = " << clip_after_allreduce
             << " , use_master_param_norm = " << use_master_param_norm
             << " , is_grad_scaled_by_nranks = " << is_grad_scaled_by_nranks;

    // Step 6: allreduce + global norm gradient clip
    int rank = 0;
    ncclComm_t comm = nullptr;
    if (num_devices > 1) {
      auto *nccl_comm_handle =
          platform::NCCLCommContext::Instance().Get(ring_id, place);
      comm = nccl_comm_handle->comm();
      rank = nccl_comm_handle->rank();
    }

    memory::Buffer grad_norm_square_buffer(place);
    auto *fp32_square_grad_norm = grad_norm_square_buffer.Alloc<float>(2);
    memory::Buffer cub_tmp_buffer(place);

    memory::Buffer sum_grad_buffer(place);
    float *fp32_sum_grad;
    platform::float16 *fp16_sum_grad;
    auto fp32_numel_each_device = fp32_numel / num_devices;
    auto fp16_numel_each_device = fp16_numel / num_devices;
    if (num_devices > 1 ||
        (max_global_grad_norm > 0 && !clip_after_allreduce)) {
      auto ptr = sum_grad_buffer.Alloc<uint8_t>(
          fp32_numel_each_device * sizeof(float) +
          fp16_numel_each_device * sizeof(platform::float16));
      fp32_sum_grad = has_fp32_param ? reinterpret_cast<float *>(ptr) : nullptr;
      fp16_sum_grad = has_fp16_param
                          ? reinterpret_cast<platform::float16 *>(
                                ptr + fp32_numel_each_device * sizeof(float))
                          : nullptr;
    } else {
      // NOTE: The const_cast here is not important. The fp32_sum_grad and
      // fp16_sum_grad would not be changed when num_devices == 1
      // But if I do not perform const_cast here, there would be more
      // if-else codes (num_devices > 1) when I write the following code.
      // So I prefer to use const_cast to unify the following code to reduce
      // the if-else codes.
      fp32_sum_grad = const_cast<float *>(fp32_grad);
      fp16_sum_grad = const_cast<platform::float16 *>(fp16_grad);
    }

    float rescale_grad = 1.0f;
    if (!is_grad_scaled_by_nranks) {
      rescale_grad /= num_devices;
    }

    if (max_global_grad_norm > 0) {
      if (clip_after_allreduce) {
        // (1) ReduceScater first
        NCCLReduceScatterWithScale(fp32_grad, fp32_sum_grad,
                                   fp32_numel_each_device, num_devices, comm,
                                   stream, dev_ctx);
        NCCLReduceScatterWithScale(fp16_grad, fp16_sum_grad,
                                   fp16_numel_each_device, num_devices, comm,
                                   stream, dev_ctx);
        // (2) Calculate the global grad norm
        GetSquareGradNorm(fp32_sum_grad, fp32_numel_each_device, fp16_sum_grad,
                          fp16_numel_each_device, fp32_square_grad_norm, stream,
                          &cub_tmp_buffer);
        VLOG(1) << "Grad square norm before all reduce: "
                << FlattenToString(fp32_square_grad_norm, 1, place);
        if (num_devices > 1) {
          PADDLE_ENFORCE_GPU_SUCCESS(platform::dynload::ncclAllReduce(
              fp32_square_grad_norm, fp32_square_grad_norm, 1, ncclFloat32,
              ncclSum, comm, stream));
        }
        VLOG(1) << "Grad square norm after all reduce: "
                << FlattenToString(fp32_square_grad_norm, 1, place);
      } else {
        // (1) Calculate the local grad norm
        GetSquareGradNorm(fp32_grad, fp32_numel, fp16_grad, fp16_numel,
                          fp32_square_grad_norm, stream, &cub_tmp_buffer);
        VLOG(1) << "Grad square norm before all reduce: "
                << FlattenToString(fp32_square_grad_norm, 1, place);
        // (2) Calculate the gradient clip scale
        float *fp32_scale = nullptr;
        platform::float16 *fp16_scale = nullptr;
        if (has_fp32_param && has_fp16_param) {
          auto *ptr = cub_tmp_buffer.Alloc<uint8_t>(sizeof(float) +
                                                    sizeof(platform::float16));
          fp32_scale = reinterpret_cast<float *>(ptr);
          fp16_scale =
              reinterpret_cast<platform::float16 *>(ptr + sizeof(float));
        } else if (has_fp32_param) {
          fp32_scale = cub_tmp_buffer.Alloc<float>(1);
        } else {
          fp16_scale = cub_tmp_buffer.Alloc<platform::float16>(1);
        }

        float clip_scale = 1.0f;
        if (is_grad_scaled_by_nranks) {
          clip_scale *= num_devices;
        }
        CalcGradNormClipBeforeAllReduceScale<
            float, platform::float16><<<1, 1, 0, stream>>>(
            global_scale, max_global_grad_norm, fp32_square_grad_norm,
            fp32_scale, fp16_scale, clip_scale);
        if (fp32_scale) {
          VLOG(1) << "Grad scale: " << FlattenToString(fp32_scale, 1, place);
        } else {
          VLOG(1) << "Grad scale: " << FlattenToString(fp16_scale, 1, place);
        }
        if (num_devices > 1) {
          PADDLE_ENFORCE_GPU_SUCCESS(platform::dynload::ncclAllReduce(
              fp32_square_grad_norm, fp32_square_grad_norm, 1, ncclFloat32,
              ncclSum, comm, stream));
        }
        // (3) Do ReduceScatter with scale
        NCCLReduceScatterWithScale(fp32_grad, fp32_sum_grad,
                                   fp32_numel_each_device, num_devices, comm,
                                   stream, dev_ctx, fp32_scale);
        NCCLReduceScatterWithScale(fp16_grad, fp16_sum_grad,
                                   fp16_numel_each_device, num_devices, comm,
                                   stream, dev_ctx, fp16_scale);
        // (4) mark max_global_grad_norm as 0, meaning that clip has been
        // already performed
        max_global_grad_norm = 0;
      }
    } else {
      NCCLReduceScatterWithScale(fp32_grad, fp32_sum_grad,
                                 fp32_numel_each_device, num_devices, comm,
                                 stream, dev_ctx);
      NCCLReduceScatterWithScale(fp16_grad, fp16_sum_grad,
                                 fp16_numel_each_device, num_devices, comm,
                                 stream, dev_ctx);
      CheckHasNanInfGrad(fp32_sum_grad, fp32_numel_each_device, fp16_sum_grad,
                         fp16_numel_each_device, fp32_square_grad_norm, stream,
                         &cub_tmp_buffer);
      if (num_devices > 1) {
        PADDLE_ENFORCE_GPU_SUCCESS(platform::dynload::ncclAllReduce(
            fp32_square_grad_norm, fp32_square_grad_norm, 1, ncclFloat32,
            ncclSum, comm, stream));
      }
      max_global_grad_norm = 0;
    }
    VLOG(10) << "ReduceScatter done";

    // Step 7: update the moment1, moment2. Calcuate the trust_ratio_div
    auto *fused_offsets_t = ctx.Input<framework::Tensor>("FusedParamOffsets");
    auto *fused_offsets = fused_offsets_t->data<int>();
    auto *fp32_partial_fused_offsets_t =
        ctx.Input<framework::Tensor>("FP32ShardFusedParamOffsets");
    const auto *fp32_partial_fused_offsets =
        fp32_partial_fused_offsets_t->data<int>();
    auto *fp16_partial_fused_offsets_t =
        ctx.Input<framework::Tensor>("FP16ShardFusedParamOffsets");
    const auto *fp16_partial_fused_offsets =
        fp16_partial_fused_offsets_t->data<int>();

    VLOG(1) << "FusedParamOffsets: "
            << FlattenToString(fused_offsets, fused_offsets_t->numel(),
                               fused_offsets_t->place());
    VLOG(1) << "FP32ShardFusedParamOffsets: "
            << FlattenToString(fp32_partial_fused_offsets,
                               fp32_partial_fused_offsets_t->numel(),
                               fp32_partial_fused_offsets_t->place());
    VLOG(1) << "FP16ShardFusedParamOffsets: "
            << FlattenToString(fp16_partial_fused_offsets,
                               fp16_partial_fused_offsets_t->numel(),
                               fp16_partial_fused_offsets_t->place());

    memory::Buffer trust_ratio_div_buffer(place);
    auto *trust_ratio_div = trust_ratio_div_buffer.Alloc<float>(partial_numel);
    auto fp32_offset = rank * fp32_numel_each_device;
    auto fp16_offset = rank * fp16_numel_each_device;
    if (has_fp32_param) {
      VLOG(10) << "Update FP32 Moment and TrustRatioDiv starts";
      MultiTensorUpdateLambMomentAndTrustRatioDiv(
          dev_ctx, fp32_partial_fused_offsets, fp32_local_param_num,
          fp32_param + fp32_offset, fp32_sum_grad, fp32_square_grad_norm,
          global_scale, beta1pow, beta2pow, moment1, moment2, trust_ratio_div,
          found_inf, weight_decay, fp32_weight_decay_end_idx, beta1, beta2,
          epsilon, max_global_grad_norm, rescale_grad);
      VLOG(10) << "Update FP32 Moment and TrustRatioDiv done";
    }
    float *master_param = nullptr;
    if (has_fp16_param) {
      master_param = fp32_param + fp32_numel;
      VLOG(10) << "Update FP16 Moment and TrustRatioDiv starts";
      auto tmp_found_inf = has_fp32_param ? nullptr : found_inf;
      MultiTensorUpdateLambMomentAndTrustRatioDiv(
          dev_ctx, fp16_partial_fused_offsets, fp16_local_param_num,
          master_param + fp16_offset, fp16_sum_grad, fp32_square_grad_norm,
          global_scale, beta1pow, beta2pow, moment1 + fp32_numel_each_device,
          moment2 + fp32_numel_each_device,
          trust_ratio_div + fp32_numel_each_device, tmp_found_inf, weight_decay,
          fp16_weight_decay_end_idx, beta1, beta2, epsilon,
          max_global_grad_norm, rescale_grad);
      VLOG(10) << "Update FP16 Moment and TrustRatioDiv done";
    }

    VLOG(10) << "Update Moment and TrustRatioDiv done hehahaha";

    // Step 8: calculate L2-Norm square of parameter and trust_ratio_div
    memory::Buffer square_norm_buffer(place);
    auto *param_square_norm = square_norm_buffer.Alloc<float>(2 * param_num);
    auto *trust_ratio_div_square_norm = param_square_norm + param_num;
    if (num_devices > 1) {
      if (use_master_param_norm) {
        FillZeroWithPtr(param_square_norm + fp32_global_param_num,
                        2 * param_num - fp32_global_param_num, stream);
      } else {
        FillZeroWithPtr(trust_ratio_div_square_norm, param_num, stream);
      }
    }
    MultiTensorL2Norm(place, stream, fp32_param, fused_offsets,
                      fp32_global_param_num, param_square_norm);
    if (use_master_param_norm) {
      MultiTensorL2Norm(place, stream, master_param + fp16_offset,
                        fp16_partial_fused_offsets, fp16_local_param_num,
                        param_square_norm + fp16_local_start_idx);
    } else {
      MultiTensorL2Norm(
          place, stream, fp16_param + fused_offsets[fp16_local_start_idx] -
                             fused_offsets[fp32_global_param_num],
          fused_offsets + fp16_local_start_idx, fp16_local_param_num,
          param_square_norm + fp16_local_start_idx);
    }

    MultiTensorL2Norm(place, stream, trust_ratio_div,
                      fp32_partial_fused_offsets, fp32_local_param_num,
                      trust_ratio_div_square_norm + fp32_local_start_idx);
    MultiTensorL2Norm(place, stream, trust_ratio_div + fp32_numel_each_device,
                      fp16_partial_fused_offsets, fp16_local_param_num,
                      trust_ratio_div_square_norm + fp16_local_start_idx);

    VLOG(1) << "TrustRatioDiv L2-Norm before allreduce: "
            << FlattenToString(trust_ratio_div_square_norm, param_num, place);
    if (num_devices > 1) {
      if (use_master_param_norm) {
        PADDLE_ENFORCE_GPU_SUCCESS(platform::dynload::ncclAllReduce(
            param_square_norm + fp32_global_param_num,
            param_square_norm + fp32_global_param_num,
            2 * param_num - fp32_global_param_num, ncclFloat32, ncclSum, comm,
            stream));
      } else {
        PADDLE_ENFORCE_GPU_SUCCESS(platform::dynload::ncclAllReduce(
            trust_ratio_div_square_norm, trust_ratio_div_square_norm, param_num,
            ncclFloat32, ncclSum, comm, stream));
      }
      VLOG(10) << "ncclAllReduce done";
    }

    LogParamAndTrustRatioDivSquareNorm<1>(ctx, param_square_norm,
                                          trust_ratio_div_square_norm);
    VLOG(10) << "Calculate L2-Norm of Param and TrustRatioDiv done";

    // Step 9: update parameter, beta1pow, beta2pow. All gather parameters.
    if (has_fp32_param) {
      MultiTensorUpdateLambParamAndBetaPows<float>(
          dev_ctx, fp32_partial_fused_offsets, fp32_local_param_num,
          trust_ratio_div, lr, param_square_norm + fp32_local_start_idx,
          trust_ratio_div_square_norm + fp32_local_start_idx, found_inf,
          fp32_param + fp32_offset, nullptr, beta1pow, beta2pow, beta1, beta2);
      if (num_devices > 1) {
        // ncclAllGather
        PADDLE_ENFORCE_GPU_SUCCESS(platform::dynload::ncclAllGather(
            fp32_param + fp32_offset, fp32_param, fp32_numel_each_device,
            ncclFloat32, comm, stream));
      }

      beta1pow = nullptr;
      beta2pow = nullptr;
    }
    if (has_fp16_param) {
      MultiTensorUpdateLambParamAndBetaPows<platform::float16>(
          dev_ctx, fp16_partial_fused_offsets, fp16_local_param_num,
          trust_ratio_div + fp32_numel_each_device, lr,
          param_square_norm + fp16_local_start_idx,
          trust_ratio_div_square_norm + fp16_local_start_idx, found_inf,
          fp16_param + fp16_offset, master_param + fp16_offset, beta1pow,
          beta2pow, beta1, beta2);
      if (num_devices > 1) {
        // ncclAllGather
        PADDLE_ENFORCE_GPU_SUCCESS(platform::dynload::ncclAllGather(
            fp16_param + fp16_offset, fp16_param, fp16_numel_each_device,
            ncclFloat16, comm, stream));
      }
    }
    VLOG(10) << "Update Param done";

    VLOG(1) << "IsFinite: " << IsFinite(dev_ctx, fp32_square_grad_norm);
#else
    PADDLE_THROW(platform::errors::Unimplemented(
        "distributed_fused_lamb op should be used with NCCL/RCCL."));
#endif
  }
};

}  // namespace operators
}  // namespace paddle

namespace plat = paddle::platform;
namespace ops = paddle::operators;

REGISTER_OP_CUDA_KERNEL(
    distributed_fused_lamb,
    ops::DistributedFusedLambOpKernel<plat::CUDADeviceContext, float>);