Merge branch 'fp32_residual_conn' into 'main'

Option to run residual connections in fp32

See merge request ADLR/megatron-lm!195

megatron/arguments.py

@@ -183,6 +183,9 @@ def parse_args(extra_args_provider=None, defaults={},
     # Mixed precision checks.
     if args.fp16_lm_cross_entropy:
         assert args.fp16, 'lm cross entropy in fp16 only support in fp16 mode.'
+    if args.fp32_residual_connection:
+        assert args.fp16, \
+            'residual connection in fp32 only supported when using fp16.'
     # Activation checkpointing.
     if args.distribute_checkpointed_activations:
         assert args.checkpoint_activations, \
@@ -197,6 +200,10 @@ def parse_args(extra_args_provider=None, defaults={},
     if args.scaled_masked_softmax_fusion:
         fused_kernels.load_scaled_masked_softmax_fusion_kernel()
 
+    # Load mixed precision fused layer norm.
+    if args.fp32_residual_connection:
+        fused_kernels.load_fused_mix_prec_layer_norm_kernel()
+
     _print_args(args)
     return args
 
@@ -435,6 +442,8 @@ def _add_mixed_precision_args(parser):
 
     group.add_argument('--fp16', action='store_true',
                        help='Run model in fp16 mode.')
+    group.add_argument('--fp32-residual-connection', action='store_true',
+                       help='Move residual connections to fp32.')
     group.add_argument('--apply-query-key-layer-scaling', action='store_true',
                        help='Scale Q * K^T by 1 / layer-number. If this flag '
                        'is set, then it will automatically set '

megatron/fused_kernels/__init__.py

@@ -98,3 +98,29 @@ def load_scaled_masked_softmax_fusion_kernel():
                            '--expt-relaxed-constexpr',
                            '--expt-extended-lambda',
                            '--use_fast_math'] + cc_flag)
+
+
+def load_fused_mix_prec_layer_norm_kernel():
+
+    # Check, if CUDA11 is installed for compute capability 8.0
+    cc_flag = []
+    _, bare_metal_major, _ = get_cuda_bare_metal_version(cpp_extension.CUDA_HOME)
+    if int(bare_metal_major) >= 11:
+        cc_flag.append('-gencode')
+        cc_flag.append('arch=compute_80,code=sm_80')
+
+    srcpath = pathlib.Path(__file__).parent.absolute()
+    buildpath = srcpath / 'build'
+
+    create_build_dir(buildpath)
+
+    fused_mix_prec_layer_norm_cuda = cpp_extension.load(
+        name='fused_mix_prec_layer_norm_cuda',
+        sources=[srcpath / 'layer_norm_cuda.cpp',
+                 srcpath / 'layer_norm_cuda_kernel.cu'],
+        build_directory=buildpath,
+        extra_cflags=['-O3'],
+        extra_cuda_cflags=['-O3',
+                           '-gencode', 'arch=compute_70,code=sm_70',
+                           '-maxrregcount=50',
+                           '--use_fast_math'] + cc_flag)

megatron/fused_kernels/compat.h

+/* coding=utf-8
+ * Copyright (c) 2020, NVIDIA CORPORATION.  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.
+ */
+
+/*This code is copied fron NVIDIA apex:
+ *     https://github.com/NVIDIA/apex
+ *     with minor changes. */
+
+
+
+#ifndef TORCH_CHECK
+#define TORCH_CHECK AT_CHECK
+#endif
+
+#ifdef VERSION_GE_1_3
+#define DATA_PTR data_ptr
+#else
+#define DATA_PTR data
+#endif

megatron/fused_kernels/layer_norm_cuda.cpp

+/* coding=utf-8
+ * Copyright (c) 2020, NVIDIA CORPORATION.  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.
+ */
+
+/*This code is copied fron NVIDIA apex:
+ *     https://github.com/NVIDIA/apex
+ *     with minor changes. */
+
+#include <torch/extension.h>
+#include <vector>
+#include <cassert>
+#include "compat.h"
+
+namespace {
+void compute_n1_n2(
+    at::Tensor input,
+    #ifdef VERSION_GE_1_1
+    at::IntArrayRef normalized_shape,
+    #else
+    at::IntList normalized_shape,
+    #endif
+    int& n1,
+    int& n2)
+{
+    int idiff = input.ndimension() - normalized_shape.size();
+    n2 = 1;
+    for (int i = 0;  i < (int)normalized_shape.size();  ++i) {
+	    assert( input.sizes()[i+idiff] == normalized_shape[i] );
+	    n2 *= normalized_shape[i];
+    }
+    n1 = 1;
+    for (int i = 0;  i < idiff;  ++i) {
+	    n1 *= input.sizes()[i];
+    }
+}
+
+void check_args(
+    #ifdef VERSION_GE_1_1
+    at::IntArrayRef normalized_shape,
+    #else
+    at::IntList normalized_shape,
+    #endif
+    at::Tensor gamma,
+    at::Tensor beta
+    )
+{
+    TORCH_CHECK(!gamma.defined() || gamma.sizes().equals(normalized_shape));
+    TORCH_CHECK(!beta.defined() || beta.sizes().equals(normalized_shape));
+}
+
+void check_args(
+    at::Tensor input,
+    #ifdef VERSION_GE_1_1
+    at::IntArrayRef normalized_shape,
+    #else
+    at::IntList normalized_shape,
+    #endif
+    int& n1,
+    int& n2
+    )
+{
+    int64_t normalized_ndim = normalized_shape.size();
+
+    if (normalized_ndim < 1) {
+      std::stringstream ss;
+      ss << "Expected normalized_shape to be at least 1-dimensional, i.e., "
+         << "containing at least one element, but got normalized_shape="
+         << normalized_shape;
+      throw std::runtime_error(ss.str());
+    }
+
+    auto input_shape = input.sizes();
+    auto input_ndim = input.dim();
+
+    if (input_ndim < normalized_ndim ||
+        !input_shape.slice(input_ndim - normalized_ndim).equals(normalized_shape)) {
+      std::stringstream ss;
+      ss << "Given normalized_shape=" << normalized_shape
+         << ", expected input with shape [*";
+      for (auto size : normalized_shape) {
+        ss << ", " << size;
+      }
+      ss << "], but got input of size" << input_shape;
+      throw std::runtime_error(ss.str());
+    }
+
+    compute_n1_n2(input,normalized_shape,n1,n2);
+}
+
+
+void check_args(
+    at::Tensor input,
+    #ifdef VERSION_GE_1_1
+    at::IntArrayRef normalized_shape,
+    #else
+    at::IntList normalized_shape,
+    #endif
+    at::Tensor gamma,
+    at::Tensor beta,
+    int& n1,
+    int& n2
+    )
+{
+    check_args(input,normalized_shape,n1,n2);
+    check_args(normalized_shape,gamma,beta);
+}
+}
+
+void cuda_layer_norm(
+    at::Tensor* output,
+    at::Tensor* mean,
+    at::Tensor* invvar,
+    at::Tensor* input,
+    int n1,
+    int n2,
+    #ifdef VERSION_GE_1_1
+    at::IntArrayRef normalized_shape,
+    #else
+    at::IntList normalized_shape,
+    #endif
+    at::Tensor* gamma,
+    at::Tensor* beta,
+    double epsilon);
+
+#define CHECK_CUDA(x) TORCH_CHECK(x.type().is_cuda(), #x " must be a CUDA tensor")
+#define CHECK_CONTIGUOUS(x) TORCH_CHECK(x.is_contiguous(), #x " must be contiguous")
+#define CHECK_INPUT(x) CHECK_CUDA(x); CHECK_CONTIGUOUS(x)
+
+std::vector<at::Tensor> layer_norm(
+    at::Tensor input,
+    #ifdef VERSION_GE_1_1
+    at::IntArrayRef normalized_shape,
+    #else
+    at::IntList normalized_shape,
+    #endif
+    double epsilon) {
+  CHECK_INPUT(input);
+  int n1,n2;
+  check_args(input,normalized_shape,n1,n2);
+  at::Tensor output = at::empty_like(input);
+  at::Tensor mean = at::empty({n1}, input.options().dtype(input.scalar_type()==at::ScalarType::Half ? at::ScalarType::Float : input.scalar_type()));
+  at::Tensor invvar = at::empty_like(mean);
+  cuda_layer_norm(&output,&mean,&invvar,&input,n1,n2,
+      normalized_shape,NULL,NULL,epsilon);
+  return {output, mean, invvar};
+}
+std::vector<at::Tensor> layer_norm_affine(
+    at::Tensor input,
+    #ifdef VERSION_GE_1_1
+    at::IntArrayRef normalized_shape,
+    #else
+    at::IntList normalized_shape,
+    #endif
+    at::Tensor gamma,
+    at::Tensor beta,
+    double epsilon) {
+  CHECK_INPUT(input);
+  CHECK_INPUT(gamma);
+  CHECK_INPUT(beta);
+  int n1,n2;
+  check_args(input,normalized_shape,gamma,beta,n1,n2);
+  at::Tensor output = at::empty_like(input, input.options().dtype(at::ScalarType::Half));
+  at::Tensor mean = at::empty({n1}, input.options().dtype(input.scalar_type()==at::ScalarType::Half ? at::ScalarType::Float : input.scalar_type()));
+  at::Tensor invvar = at::empty_like(mean);
+  cuda_layer_norm(&output,&mean,&invvar,&input,n1,n2,
+      normalized_shape,&gamma,&beta,epsilon);
+  return {output, mean, invvar};
+}
+
+void cuda_layer_norm_gradient(
+    at::Tensor* dout,
+    at::Tensor* mean,
+    at::Tensor* invvar,
+    at::Tensor* input,
+    int n1,
+    int n2,
+    #ifdef VERSION_GE_1_1
+    at::IntArrayRef normalized_shape,
+    #else
+    at::IntList normalized_shape,
+    #endif
+    at::Tensor* gamma,
+    at::Tensor* beta,
+    double epsilon,
+    at::Tensor* grad_input,
+    at::Tensor* grad_gamma,
+    at::Tensor* grad_beta
+    );
+
+at::Tensor layer_norm_gradient(
+    at::Tensor dout,
+    at::Tensor mean,
+    at::Tensor invvar,
+    at::Tensor input,
+    #ifdef VERSION_GE_1_1
+    at::IntArrayRef normalized_shape,
+    #else
+    at::IntList normalized_shape,
+    #endif
+    double epsilon) {
+  CHECK_INPUT(dout);
+  CHECK_INPUT(mean);
+  CHECK_INPUT(invvar);
+  CHECK_INPUT(input);
+  int n1,n2;
+  check_args(input,normalized_shape,n1,n2);
+  at::Tensor grad_input = at::empty_like(input);
+  cuda_layer_norm_gradient(&dout,&mean,&invvar,&input,n1,n2,
+      normalized_shape,NULL,NULL,epsilon,
+      &grad_input,NULL,NULL);
+  return grad_input;
+}
+std::vector<at::Tensor> layer_norm_gradient_affine(
+    at::Tensor dout,
+    at::Tensor mean,
+    at::Tensor invvar,
+    at::Tensor input,
+    #ifdef VERSION_GE_1_1
+    at::IntArrayRef normalized_shape,
+    #else
+    at::IntList normalized_shape,
+    #endif
+    at::Tensor gamma,
+    at::Tensor beta,
+    double epsilon) {
+  CHECK_INPUT(dout);
+  CHECK_INPUT(mean);
+  CHECK_INPUT(invvar);
+  CHECK_INPUT(input);
+  CHECK_INPUT(gamma);
+  CHECK_INPUT(beta);
+  int n1,n2;
+  check_args(input,normalized_shape,gamma,beta,n1,n2);
+  at::Tensor grad_input = at::empty_like(input);
+  at::Tensor grad_gamma = at::empty_like(gamma);
+  at::Tensor grad_beta = at::empty_like(beta);
+  cuda_layer_norm_gradient(&dout,&mean,&invvar,&input,n1,n2,
+      normalized_shape,&gamma,&beta,epsilon,
+      &grad_input,&grad_gamma,&grad_beta);
+  return {grad_input, grad_gamma, grad_beta};
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+  m.def("forward_affine", &layer_norm_affine, "LayerNorm forward (CUDA)");
+  m.def("forward", &layer_norm, "LayerNorm forward (CUDA)");
+  m.def("backward_affine", &layer_norm_gradient_affine, "LayerNorm backward (CUDA)");
+  m.def("backward", &layer_norm_gradient, "LayerNorm backward (CUDA)");
+}

megatron/fused_kernels/layer_norm_cuda_kernel.cu

+/* coding=utf-8
+ * Copyright (c) 2020, NVIDIA CORPORATION.  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.
+ */
+
+/*This code is copied fron NVIDIA apex:
+ *     https://github.com/NVIDIA/apex
+ *     with minor changes. */
+
+#include "ATen/ATen.h"
+#include "ATen/AccumulateType.h"
+#include "ATen/cuda/CUDAContext.h"
+#include <THC/THCDeviceUtils.cuh>
+
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include "type_shim.h"
+
+template<typename U> __device__
+void cuWelfordOnlineSum(
+  const U curr,
+  U& mu,
+  U& sigma2,
+  U& count)
+{
+  count = count + U(1);
+  U delta = curr - mu;
+  U lmean = mu + delta / count;
+  mu = lmean;
+  U delta2 = curr - lmean;
+  sigma2 = sigma2 + delta * delta2;
+}
+
+template<typename U> __device__
+void cuChanOnlineSum(
+  const U muB,
+  const U sigma2B,
+  const U countB,
+  U& mu,
+  U& sigma2,
+  U& count)
+{
+  U delta = muB - mu;
+  U nA = count;
+  U nB = countB;
+  count = count + countB;
+  U nX = count;
+  if (nX > U(0)) {
+    nA = nA / nX;
+    nB = nB / nX;
+    mu = nA*mu + nB*muB;
+    sigma2 = sigma2 + sigma2B + delta * delta * nA * nB * nX;
+  } else {
+    mu = U(0);
+    sigma2 = U(0);
+  }
+}
+
+template<typename T, typename U> __device__
+void cuWelfordMuSigma2(
+  const T* __restrict__ vals,
+  const int n1,
+  const int n2,
+  const int i1,
+  U& mu,
+  U& sigma2,
+  U* buf) 
+{
+  // Assumptions:
+  // 1) blockDim.x == warpSize
+  // 2) Tensor is contiguous
+  // 3) 2*blockDim.y*sizeof(U)+blockDim.y*sizeof(int) shared memory available.
+  //
+  // compute variance and mean over n2
+  U count = U(0);
+  mu= U(0);
+  sigma2 = U(0);
+  if (i1 < n1) {
+    // one warp normalizes one n1 index,
+    // synchronization is implicit
+    // initialize with standard Welford algorithm
+    const int numx = blockDim.x * blockDim.y;
+    const int thrx = threadIdx.x + threadIdx.y * blockDim.x;
+    const T* lvals = vals + i1*n2;
+    int l = 4*thrx;
+    for (;  l+3 < n2;  l+=4*numx) {
+      for (int k = 0;  k < 4;  ++k) {
+        U curr = static_cast<U>(lvals[l+k]);
+        cuWelfordOnlineSum<U>(curr,mu,sigma2,count);
+      }
+    }
+    for (;  l < n2;  ++l) {
+      U curr = static_cast<U>(lvals[l]);
+      cuWelfordOnlineSum<U>(curr,mu,sigma2,count);
+    }
+    // intra-warp reductions
+    for (int l = 0;  l <= 4;  ++l) {
+      int srcLaneB = (threadIdx.x+(1<<l))&31;
+      U muB = WARP_SHFL(mu, srcLaneB);
+      U countB = WARP_SHFL(count, srcLaneB);
+      U sigma2B = WARP_SHFL(sigma2, srcLaneB);
+      cuChanOnlineSum<U>(muB,sigma2B,countB,mu,sigma2,count);
+    }
+    // threadIdx.x == 0 has correct values for each warp
+    // inter-warp reductions
+    if (blockDim.y > 1) {
+      U* ubuf = (U*)buf;
+      U* ibuf = (U*)(ubuf + blockDim.y);
+      for (int offset = blockDim.y/2;  offset > 0;  offset /= 2) {
+        // upper half of warps write to shared
+        if (threadIdx.x == 0 && threadIdx.y >= offset && threadIdx.y < 2*offset) {
+          const int wrt_y = threadIdx.y - offset;
+          ubuf[2*wrt_y] = mu;
+          ubuf[2*wrt_y+1] = sigma2;
+          ibuf[wrt_y] = count;
+        }
+        __syncthreads();
+        // lower half merges
+        if (threadIdx.x == 0 && threadIdx.y < offset) {
+          U muB = ubuf[2*threadIdx.y];
+          U sigma2B = ubuf[2*threadIdx.y+1];
+          U countB = ibuf[threadIdx.y];
+          cuChanOnlineSum<U>(muB,sigma2B,countB,mu,sigma2,count);
+        }
+        __syncthreads();
+      }
+      // threadIdx.x = 0 && threadIdx.y == 0 only thread that has correct values
+      if (threadIdx.x == 0 && threadIdx.y == 0) {
+        ubuf[0] = mu;
+        ubuf[1] = sigma2;
+      }
+      __syncthreads();
+      mu = ubuf[0];
+      sigma2 = ubuf[1]/U(n2);
+      // don't care about final value of count, we know count == n2
+    } else {
+      mu = WARP_SHFL(mu, 0);
+      sigma2 = WARP_SHFL(sigma2/U(n2), 0);
+    }
+  }
+}
+
+template<> __device__
+void cuWelfordMuSigma2(
+  const at::Half* __restrict__ vals,
+  const int n1,
+  const int n2,
+  const int i1,
+  float& mu,
+  float& sigma2,
+  float* buf) 
+{
+  // Assumptions:
+  // 1) blockDim.x == warpSize
+  // 2) Tensor is contiguous
+  // 3) 2*blockDim.y*sizeof(U)+blockDim.y*sizeof(int) shared memory available.
+  //
+  // compute variance and mean over n2
+  float count = 0.0f;
+  mu= float(0);
+  sigma2 = float(0);
+  if (i1 < n1) {
+    // one warp normalizes one n1 index,
+    // synchronization is implicit
+    // initialize with standard Welford algorithm
+    const int numx = blockDim.x * blockDim.y;
+    const int thrx = threadIdx.x + threadIdx.y * blockDim.x;
+    const at::Half* lvals = vals + i1*n2;
+    int l = 8*thrx;
+    if ((((size_t)lvals)&3) != 0) {
+      // 16 bit alignment
+      // first thread consumes first point
+      if (thrx == 0) {
+        float curr = static_cast<float>(lvals[0]);
+        cuWelfordOnlineSum(curr,mu,sigma2,count);
+      }
+      ++l;
+    }
+    // at this point, lvals[l] are 32 bit aligned for all threads.
+    for (;  l+7 < n2;  l+=8*numx) {
+      for (int k = 0;  k < 8;  k+=2) {
+        float2 curr = __half22float2(*((__half2*)(lvals+l+k)));
+        cuWelfordOnlineSum(curr.x,mu,sigma2,count);
+	cuWelfordOnlineSum(curr.y,mu,sigma2,count);
+      }
+    }
+    for (;  l < n2;  ++l) {
+      float curr = static_cast<float>(lvals[l]);
+      cuWelfordOnlineSum(curr,mu,sigma2,count);
+    }
+    // intra-warp reductions
+    for (int l = 0;  l <= 4;  ++l) {
+      int srcLaneB = (threadIdx.x+(1<<l))&31;
+      float muB = WARP_SHFL(mu, srcLaneB);
+      float countB = WARP_SHFL(count, srcLaneB);
+      float sigma2B = WARP_SHFL(sigma2, srcLaneB);
+      cuChanOnlineSum(muB,sigma2B,countB,mu,sigma2,count);
+    }
+    // threadIdx.x == 0 has correct values for each warp
+    // inter-warp reductions
+    if (blockDim.y > 1) {
+      float* ubuf = (float*)buf;
+      float* ibuf = (float*)(ubuf + blockDim.y);
+      for (int offset = blockDim.y/2;  offset > 0;  offset /= 2) {
+        // upper half of warps write to shared
+        if (threadIdx.x == 0 && threadIdx.y >= offset && threadIdx.y < 2*offset) {
+          const int wrt_y = threadIdx.y - offset;
+          ubuf[2*wrt_y] = mu;
+          ubuf[2*wrt_y+1] = sigma2;
+          ibuf[wrt_y] = count;
+        }
+        __syncthreads();
+        // lower half merges
+        if (threadIdx.x == 0 && threadIdx.y < offset) {
+          float muB = ubuf[2*threadIdx.y];
+          float sigma2B = ubuf[2*threadIdx.y+1];
+          float countB = ibuf[threadIdx.y];
+          cuChanOnlineSum(muB,sigma2B,countB,mu,sigma2,count);
+        }
+        __syncthreads();
+      }
+      // threadIdx.x = 0 && threadIdx.y == 0 only thread that has correct values
+      if (threadIdx.x == 0 && threadIdx.y == 0) {
+        ubuf[0] = mu;
+        ubuf[1] = sigma2;
+      }
+      __syncthreads();
+      mu = ubuf[0];
+      sigma2 = ubuf[1]/float(n2);
+      // don't care about final value of count, we know count == n2
+    } else {
+      mu = WARP_SHFL(mu, 0);
+      sigma2 = WARP_SHFL(sigma2/float(n2), 0);
+    }
+  }
+}
+
+template<typename U> U rsqrt(U v) {
+  return U(1) / sqrt(v);
+}
+template<> float rsqrt(float v) {
+  return rsqrtf(v);
+}
+template<> double rsqrt(double v) {
+  return rsqrt(v);
+}
+
+namespace {
+// This is the un-specialized struct.  Note that we prevent instantiation of this
+// struct by putting an undefined symbol in the function body so it won't compile.
+//  template <typename T>
+//  struct SharedMemory
+//  {
+//      // Ensure that we won't compile any un-specialized types
+//      __device__ T *getPointer()
+//      {
+//          extern __device__ void error(void);
+//          error();
+//          return NULL;
+//      }
+//  };
+// https://github.com/NVIDIA/apex/issues/246
+template <typename T>
+struct SharedMemory;
+
+template <>
+struct SharedMemory <float>
+{
+    __device__ float *getPointer()
+    {
+        extern __shared__ float s_float[];
+        return s_float;
+    }
+};
+
+template <>
+struct SharedMemory <double>
+{
+    __device__ double *getPointer()
+    {
+        extern __shared__ double s_double[];
+        return s_double;
+    }
+};
+}
+
+template<typename T, typename U, typename V> __global__
+void cuApplyLayerNorm(
+  V* __restrict__ output_vals,
+  U* __restrict__ mean,
+  U* __restrict__ invvar,
+  const T* __restrict__ vals,
+  const int n1,
+  const int n2,
+  const U epsilon,
+  const V* __restrict__ gamma,
+  const V* __restrict__ beta
+  ) 
+{
+  // Assumptions:
+  // 1) blockDim.x == warpSize
+  // 2) Tensors are contiguous
+  //
+  for (auto i1=blockIdx.y; i1 < n1; i1 += gridDim.y) {
+    SharedMemory<U> shared;
+    U* buf = shared.getPointer();
+    U mu,sigma2;
+    cuWelfordMuSigma2(vals,n1,n2,i1,mu,sigma2,buf);
+    const T* lvals = vals + i1*n2;
+    V* ovals = output_vals + i1*n2;
+    U c_invvar = rsqrt(sigma2 + epsilon);
+    const int numx = blockDim.x * blockDim.y;
+    const int thrx = threadIdx.x + threadIdx.y * blockDim.x;
+    if (gamma != NULL && beta != NULL) {
+      for (int i = thrx;  i < n2;  i+=numx) {
+        U curr = static_cast<U>(lvals[i]);
+        ovals[i] = gamma[i] * static_cast<V>(c_invvar * (curr - mu)) + beta[i];
+      }
+    } else {
+      for (int i = thrx;  i < n2;  i+=numx) {
+        U curr = static_cast<U>(lvals[i]);
+        ovals[i] = static_cast<V>(c_invvar * (curr - mu));
+      }
+    }
+    if (threadIdx.x == 0 && threadIdx.y == 0) {
+      mean[i1] = mu;
+      invvar[i1] = c_invvar;
+    }
+  }
+}
+
+template<typename T, typename U, typename V> __device__
+void cuLoadWriteStridedInputs(
+    const int i1_block,
+    const int thr_load_row_off,
+    const int thr_load_col_off,
+    const int i2_off,
+    const int row_stride,
+    U* warp_buf1,
+    U* warp_buf2,
+    const T* input,
+    const V* dout,
+    const int i1_end,
+    const int n2,
+    const U* __restrict__ mean,
+    const U* __restrict__ invvar
+    )
+{
+  int i1 = i1_block+thr_load_row_off;
+  if (i1 < i1_end) {
+    U curr_mean = mean[i1];
+    U curr_invvar = invvar[i1];
+    for (int k = 0;  k < blockDim.y;  ++k) {
+      int i2 = i2_off + k;
+      int load_idx = i1*n2+i2;
+      int write_idx = thr_load_row_off*row_stride+thr_load_col_off+k;
+      if (i2<n2) {
+        U curr_input = static_cast<U>(input[load_idx]);
+	U curr_dout = static_cast<U>(dout[load_idx]);
+	warp_buf1[write_idx] = curr_dout;
+	warp_buf2[write_idx] = curr_dout * (curr_input - curr_mean) * curr_invvar;
+      } else {
+        warp_buf1[write_idx] = U(0);
+        warp_buf2[write_idx] = U(0);
+      }
+    }
+  } else {
+    for (int k = 0;  k < blockDim.y;  ++k) {
+      int write_idx = thr_load_row_off*row_stride+thr_load_col_off+k;
+      warp_buf1[write_idx] = U(0);
+      warp_buf2[write_idx] = U(0);
+    }
+  }
+}
+
+template<typename T, typename U, typename V> __device__
+void cuLoadAddStridedInputs(
+    const int i1_block,
+    const int thr_load_row_off,
+    const int thr_load_col_off,
+    const int i2_off,
+    const int row_stride,
+    U* warp_buf1,
+    U* warp_buf2,
+    const T* input,
+    const V* dout,
+    const int i1_end,
+    const int n2,
+    const U* __restrict__ mean,
+    const U* __restrict__ invvar
+    )
+{
+  int i1 = i1_block+thr_load_row_off;
+  if (i1 < i1_end) {
+    U curr_mean = mean[i1];
+    U curr_invvar = invvar[i1];
+    for (int k = 0;  k < blockDim.y;  ++k) {
+      int i2 = i2_off + k;
+      int load_idx = i1*n2+i2;
+      int write_idx = thr_load_row_off*row_stride+thr_load_col_off+k;
+      if (i2<n2) {
+        U curr_input = static_cast<U>(input[load_idx]);
+	U curr_dout = static_cast<U>(dout[load_idx]);
+	warp_buf1[write_idx] += curr_dout;
+	warp_buf2[write_idx] += curr_dout * (curr_input - curr_mean) * curr_invvar;
+      }
+    }
+  }
+}
+
+template<typename T, typename U, typename V> __global__
+void cuComputePartGradGammaBeta(
+    const V* __restrict__ dout,
+    const T* __restrict__ input,
+    const int n1,
+    const int n2,
+    const U* __restrict__ mean,
+    const U* __restrict__ invvar,
+    U epsilon,
+    U* part_grad_gamma,
+    U* part_grad_beta)
+{
+    const int numsegs_n1 = (n1+blockDim.y*blockDim.y-1) / (blockDim.y*blockDim.y);
+    const int segs_per_block = (numsegs_n1 + gridDim.y - 1) / gridDim.y;
+    const int i1_beg = blockIdx.y * segs_per_block * blockDim.y*blockDim.y;
+    const int i1_beg_plus_one = (blockIdx.y+1) * segs_per_block * blockDim.y*blockDim.y;
+    const int i1_end = i1_beg_plus_one < n1 ? i1_beg_plus_one : n1;
+    const int row_stride = blockDim.x+1;
+    const int thr_load_col_off = (threadIdx.x*blockDim.y)&(blockDim.x-1);
+    const int thr_load_row_off = (threadIdx.x*blockDim.y)/blockDim.x + threadIdx.y*blockDim.y;
+    const int i2_off = blockIdx.x * blockDim.x + thr_load_col_off;
+    SharedMemory<U> shared;
+    U* buf = shared.getPointer(); // buf has at least blockDim.x * blockDim.y * blockDim.y + (blockDim.y - 1)*(blockDim.x/blockDim.y) elements
+    U* warp_buf1 = (U*)buf;
+    U* warp_buf2 = warp_buf1 + blockDim.y * blockDim.y * row_stride;
+    // compute partial sums from strided inputs
+    // do this to increase number of loads in flight
+    cuLoadWriteStridedInputs(i1_beg,thr_load_row_off,thr_load_col_off,i2_off,row_stride,warp_buf1,warp_buf2,input,dout,i1_end,n2,mean,invvar);
+    for (int i1_block = i1_beg+blockDim.y*blockDim.y;  i1_block < i1_end;  i1_block+=blockDim.y*blockDim.y) {
+      cuLoadAddStridedInputs(i1_block,thr_load_row_off,thr_load_col_off,i2_off,row_stride,warp_buf1,warp_buf2,input,dout,i1_end,n2,mean,invvar);
+    }
+    __syncthreads();
+    // inter-warp reductions
+    // sum within each warp
+    U acc1 = U(0);
+    U acc2 = U(0);
+    for (int k = 0;  k < blockDim.y;  ++k) {
+      int row1 = threadIdx.y + k*blockDim.y;
+      int idx1 = row1*row_stride + threadIdx.x;
+      acc1 += warp_buf1[idx1];
+      acc2 += warp_buf2[idx1];
+    }
+    warp_buf1[threadIdx.y*row_stride+threadIdx.x] = acc1;
+    warp_buf2[threadIdx.y*row_stride+threadIdx.x] = acc2;
+    __syncthreads();
+    // sum all warps
+    for (int offset = blockDim.y/2;  offset > 1;  offset /= 2) {
+      if (threadIdx.y < offset) {
+        int row1 = threadIdx.y;
+	int row2 = threadIdx.y + offset;
+	int idx1 = row1*row_stride + threadIdx.x;
+	int idx2 = row2*row_stride + threadIdx.x;
+	warp_buf1[idx1] += warp_buf1[idx2];
+	warp_buf2[idx1] += warp_buf2[idx2];
+      }
+      __syncthreads();
+    }
+    int i2 = blockIdx.x * blockDim.x + threadIdx.x;
+    if (threadIdx.y == 0 && i2 < n2) {
+      int row1 = threadIdx.y;
+      int row2 = threadIdx.y + 1;
+      int idx1 = row1*row_stride + threadIdx.x;
+      int idx2 = row2*row_stride + threadIdx.x;
+      part_grad_beta[blockIdx.y*n2+i2] = warp_buf1[idx1] + warp_buf1[idx2];
+      part_grad_gamma[blockIdx.y*n2+i2] = warp_buf2[idx1] + warp_buf2[idx2];
+    }
+}
+
+template<typename U, typename V> __global__
+void cuComputeGradGammaBeta(
+    const U* part_grad_gamma,
+    const U* part_grad_beta,
+    const int part_size,
+    const int n1,
+    const int n2,
+    V* grad_gamma,
+    V* grad_beta)
+{
+    // sum partial gradients for gamma and beta
+    SharedMemory<U> shared;
+    U* buf = shared.getPointer(); 
+    int i2 = blockIdx.x * blockDim.x + threadIdx.x;
+    if (i2 < n2) {
+      // each warp does sequential reductions until reduced part_size is num_warps
+      int num_warp_reductions = part_size / blockDim.y;
+      U sum_gamma = U(0);
+      U sum_beta = U(0);
+      const U* part_grad_gamma_ptr = part_grad_gamma + threadIdx.y * num_warp_reductions * n2 + i2;
+      const U* part_grad_beta_ptr = part_grad_beta + threadIdx.y * num_warp_reductions * n2 + i2;
+      for (int warp_offset = 0;  warp_offset < num_warp_reductions;  ++warp_offset) {
+        sum_gamma += part_grad_gamma_ptr[warp_offset*n2];
+        sum_beta += part_grad_beta_ptr[warp_offset*n2];
+      }
+      // inter-warp reductions
+      const int nbsize3 = blockDim.x * blockDim.y / 2;
+      for (int offset = blockDim.y/2;  offset >= 1;  offset /= 2) {
+        // top half write to shared memory
+        if (threadIdx.y >= offset && threadIdx.y < 2*offset) {
+          const int write_idx = (threadIdx.y - offset) * blockDim.x + threadIdx.x;
+          buf[write_idx] = sum_gamma;
+          buf[write_idx+nbsize3] = sum_beta;
+        }
+        __syncthreads();
+        // bottom half sums
+        if (threadIdx.y < offset) {
+          const int read_idx = threadIdx.y * blockDim.x + threadIdx.x;
+          sum_gamma += buf[read_idx];
+          sum_beta += buf[read_idx+nbsize3];
+        }
+        __syncthreads();
+      }
+      // write out fully summed gradients
+      if (threadIdx.y == 0) {
+        grad_gamma[i2] = sum_gamma;
+        grad_beta[i2] = sum_beta;
+      }
+    }
+}
+
+template<typename T, typename U, typename V> __global__
+void cuComputeGradInput(
+    const V* __restrict__ dout,
+    const T* __restrict__ input,
+    const int n1,
+    const int n2,
+    const U* __restrict__ mean,
+    const U* __restrict__ invvar,
+    U epsilon,
+    const V* gamma,
+    T* grad_input)
+{
+  for (auto i1=blockIdx.y; i1 < n1; i1 += gridDim.y) {
+    U sum_loss1 = U(0);
+    U sum_loss2 = U(0);
+    const U c_mean = mean[i1];
+    const U c_invvar = invvar[i1];
+    const T* k_input = input + i1*n2;
+    const V* k_dout = dout + i1*n2;
+    const int numx = blockDim.x * blockDim.y;
+    const int thrx = threadIdx.x + threadIdx.y * blockDim.x;
+    if (gamma != NULL) {
+      int l = 4*thrx;
+      for (;  l+3 < n2;  l+=4*numx) {
+        for (int k = 0;  k < 4;  ++k) {
+          const U c_h = static_cast<U>(k_input[l+k]);
+          const U c_loss = static_cast<U>(k_dout[l+k]);
+          sum_loss1 += c_loss * gamma[l+k];
+          sum_loss2 += c_loss * gamma[l+k] * (c_h - c_mean) * c_invvar;
+        }
+      }
+      for (;  l < n2;  ++l) {
+        const U c_h = static_cast<U>(k_input[l]);
+        const U c_loss = static_cast<U>(k_dout[l]);
+        sum_loss1 += c_loss * gamma[l];
+        sum_loss2 += c_loss * gamma[l] * (c_h - c_mean) * c_invvar;
+      }
+    } else {
+      int l = 4*thrx;
+      for (;  l+3 < n2;  l+=4*numx) {
+        for (int k = 0;  k < 4;  ++k) {
+          const U c_h = static_cast<U>(k_input[l+k]);
+          const U c_loss = static_cast<U>(k_dout[l+k]);
+          sum_loss1 += c_loss;
+          sum_loss2 += c_loss * (c_h - c_mean) * c_invvar;
+        }
+      }
+      for (;  l < n2;  ++l) {
+        const U c_h = static_cast<U>(k_input[l]);
+        const U c_loss = static_cast<U>(k_dout[l]);
+        sum_loss1 += c_loss;
+        sum_loss2 += c_loss * (c_h - c_mean) * c_invvar;
+      }
+    }
+    // intra-warp reductions
+    for (int mask = blockDim.x/2;  mask > 0;  mask /= 2) {
+      sum_loss1 += WARP_SHFL_XOR(sum_loss1, mask);
+      sum_loss2 += WARP_SHFL_XOR(sum_loss2, mask);
+    }
+    // inter-warp reductions
+    if (blockDim.y > 1) {
+      SharedMemory<U> shared;
+      U* buf = shared.getPointer(); 
+      for (int offset = blockDim.y/2;  offset > 0;  offset /= 2) {
+        // upper half of warps write to shared
+        if (threadIdx.y >= offset && threadIdx.y < 2*offset) {
+          const int wrt_i = (threadIdx.y - offset) * blockDim.x + threadIdx.x;
+          buf[2*wrt_i] = sum_loss1;
+          buf[2*wrt_i+1] = sum_loss2;
+        }
+        __syncthreads();
+        // lower half merges
+        if (threadIdx.y < offset) {
+          const int read_i = threadIdx.y * blockDim.x + threadIdx.x;
+          sum_loss1 += buf[2*read_i];
+          sum_loss2 += buf[2*read_i+1];
+        }
+        __syncthreads();
+      }
+      if (threadIdx.y == 0) {
+        buf[2*threadIdx.x] = sum_loss1;
+        buf[2*threadIdx.x+1] = sum_loss2;
+      }
+      __syncthreads();
+      if (threadIdx.y !=0) {
+        sum_loss1 = buf[2*threadIdx.x];
+        sum_loss2 = buf[2*threadIdx.x+1];
+      } 
+    }
+    // all threads now have the two sums over l
+    U fH = (U)n2;
+    U term1 = (U(1) / fH) * c_invvar;
+    T* k_grad_input = grad_input + i1*n2;
+    if (gamma != NULL) {
+      for (int l = thrx;  l < n2;  l+=numx) {
+        const U c_h = static_cast<U>(k_input[l]);
+        const U c_loss = static_cast<U>(k_dout[l]);
+        U f_grad_input = fH * c_loss * gamma[l];
+        f_grad_input -= sum_loss1;
+        f_grad_input -= (c_h - c_mean) * c_invvar * sum_loss2;
+        f_grad_input *= term1;
+        k_grad_input[l] = static_cast<T>(f_grad_input);
+      }
+    } else {
+      for (int l = thrx;  l < n2;  l+=numx) {
+        const U c_h = static_cast<U>(k_input[l]);
+        const U c_loss = static_cast<U>(k_dout[l]);
+        U f_grad_input = fH * c_loss;
+        f_grad_input -= sum_loss1;
+        f_grad_input -= (c_h - c_mean) * c_invvar * sum_loss2;
+        f_grad_input *= term1;
+        k_grad_input[l] = static_cast<T>(f_grad_input);
+      }
+    }
+  }
+}
+
+template<typename T, typename U, typename V> 
+void HostApplyLayerNorm(
+    V* output,
+    U* mean,
+    U* invvar,
+    const T* input,
+    int n1,
+    int n2,
+    double epsilon,
+    const V* gamma,
+    const V* beta
+    )
+{
+    auto stream = at::cuda::getCurrentCUDAStream().stream();
+    const dim3 threads(32,4,1);
+    const uint64_t maxGridY = at::cuda::getCurrentDeviceProperties()->maxGridSize[1];
+    const dim3 blocks(1, std::min((uint64_t)n1, maxGridY), 1);
+    int nshared = 
+        threads.y > 1 ? 
+	    threads.y*sizeof(U)+(threads.y/2)*sizeof(U) : 
+	    0;
+    cuApplyLayerNorm<<<blocks, threads, nshared, stream>>>(
+		    output,
+		    mean,
+		    invvar,
+		    input,
+		    n1,n2,
+		    U(epsilon),
+            gamma,beta);
+}
+
+void cuda_layer_norm(
+    at::Tensor* output,
+    at::Tensor* mean,
+    at::Tensor* invvar,
+    at::Tensor* input,
+    int n1,
+    int n2,
+    #ifdef VERSION_GE_1_1
+    at::IntArrayRef normalized_shape,
+    #else
+    at::IntList normalized_shape,
+    #endif
+    at::Tensor* gamma,
+    at::Tensor* beta,
+    double epsilon)
+{
+    using namespace at;
+    DISPATCH_DOUBLE_FLOAT_AND_HALF(input->scalar_type(), 0, "layer_norm_cuda_kernel",
+        using accscalar_t = at::acc_type<scalar_t_0, true>;
+        using output_t = at::Half;
+        HostApplyLayerNorm(
+        output->DATA_PTR<output_t>(),
+	    mean->DATA_PTR<accscalar_t>(),
+	    invvar->DATA_PTR<accscalar_t>(),
+	    input->DATA_PTR<scalar_t_0>(),
+	    n1,n2,
+	    epsilon,
+	    gamma != NULL ? gamma->DATA_PTR<output_t>() : NULL,
+	    beta != NULL ? beta->DATA_PTR<output_t>() : NULL);
+      )
+}
+
+template<typename T, typename U, typename V>
+void HostLayerNormGradient(
+    const V* dout,
+    const U* mean,
+    const U* invvar,
+    at::Tensor* input,
+    int n1,
+    int n2,
+    const V* gamma,
+    const V* beta,
+    double epsilon,
+    T* grad_input,
+    V* grad_gamma,
+    V* grad_beta
+    )
+{
+    auto stream = at::cuda::getCurrentCUDAStream().stream();
+
+    if (gamma != NULL && beta != NULL) {
+      // compute grad_gamma(j) and grad_beta(j)
+      const int part_size = 16;
+      const dim3 threads2(32,4,1);
+      const dim3 blocks2((n2+threads2.x-1)/threads2.x,part_size,1);
+      const int nshared2_a = 2 * sizeof(U) * threads2.y * threads2.y * (threads2.x + 1);
+      const int nshared2_b = threads2.x * threads2.y * sizeof(U);
+      const int nshared2 = nshared2_a > nshared2_b ? nshared2_a : nshared2_b;
+      at::Tensor part_grad_gamma = at::empty({part_size,n2}, input->options().dtype(input->scalar_type()==at::ScalarType::Half ? at::ScalarType::Float : input->scalar_type()));
+      at::Tensor part_grad_beta = at::empty_like(part_grad_gamma);
+      cuComputePartGradGammaBeta<<<blocks2, threads2, nshared2, stream>>>(
+		      dout,
+		      input->DATA_PTR<T>(),
+		      n1,n2,
+		      mean,
+		      invvar,
+		      U(epsilon),
+		      part_grad_gamma.DATA_PTR<U>(),
+		      part_grad_beta.DATA_PTR<U>());
+
+      const dim3 threads3(32,8,1);
+      const dim3 blocks3((n2+threads2.x-1)/threads2.x,1,1);
+      const int nshared3 = threads3.x * threads3.y * sizeof(U);
+      cuComputeGradGammaBeta<<<blocks3, threads3, nshared3, stream>>>(
+		      part_grad_gamma.DATA_PTR<U>(),
+		      part_grad_beta.DATA_PTR<U>(),
+		      part_size,
+		      n1,n2,
+		      grad_gamma,
+		      grad_beta);
+    }
+
+    // compute grad_input
+    const uint64_t maxGridY = at::cuda::getCurrentDeviceProperties()->maxGridSize[1];
+    const dim3 blocks1(1, std::min((uint64_t)n1, maxGridY), 1);
+    const dim3 threads1(32,4,1);
+    int nshared =
+	    threads1.y > 1 ?
+	    threads1.y*threads1.x*sizeof(U) :
+	    0;
+    cuComputeGradInput<<<blocks1, threads1, nshared, stream>>>(
+            dout,
+            input->DATA_PTR<T>(),
+            n1,n2,
+            mean,
+            invvar,
+            U(epsilon),
+            gamma,
+            grad_input);
+}
+
+void cuda_layer_norm_gradient(
+    at::Tensor* dout,
+    at::Tensor* mean,
+    at::Tensor* invvar,
+    at::Tensor* input,
+    int n1,
+    int n2,
+    #ifdef VERSION_GE_1_1
+    at::IntArrayRef normalized_shape,
+    #else
+    at::IntList normalized_shape,
+    #endif
+    at::Tensor* gamma,
+    at::Tensor* beta,
+    double epsilon,
+    at::Tensor* grad_input,
+    at::Tensor* grad_gamma,
+    at::Tensor* grad_beta)
+{
+    using namespace at;
+    DISPATCH_FLOAT_AND_HALF(input->scalar_type(), 0, "cuComputeGradInput",
+        using accscalar_t = at::acc_type<scalar_t_0, true>;
+        using output_t = at::Half;
+        HostLayerNormGradient(
+	    dout->DATA_PTR<output_t>(),
+	    mean->DATA_PTR<accscalar_t>(),
+	    invvar->DATA_PTR<accscalar_t>(),
+	    input,
+	    n1,n2,
+            // TMJ pass NULL argument for gamma, beta, grad_gamma and grad_beta
+            // if gamma Tensor is NULL on input.
+	    gamma != NULL ? gamma->DATA_PTR<output_t>() : NULL,
+	    gamma != NULL ? beta->DATA_PTR<output_t>() : NULL,
+	    epsilon,
+	    grad_input->DATA_PTR<scalar_t_0>(),
+	    gamma != NULL ? grad_gamma->DATA_PTR<output_t>() : NULL,
+	    gamma != NULL ? grad_beta->DATA_PTR<output_t>() : NULL);
+      )
+}

megatron/fused_kernels/type_shim.h

+/* coding=utf-8
+ * Copyright (c) 2020, NVIDIA CORPORATION.  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.
+ */
+
+/*This code is copied fron NVIDIA apex:
+ *     https://github.com/NVIDIA/apex
+ *     with minor changes. */
+
+#include <ATen/ATen.h>
+#include "compat.h"
+
+// Forward/backward compatiblity hack around
+// https://github.com/pytorch/pytorch/commit/3aeb78079bcd68282fe9117088e138b77318e288
+// pending more future-proof guidance from upstream.
+// struct TypeShim
+// {
+//   const at::Type& payload;
+//   TypeShim(const at::Type& type) : payload(type) {}
+//   // Enable trivial conversion to a const at::Type& for pre-3aeb78
+//   operator const at::Type&(){ return payload; };
+//   // Enable dispatch switch statements to take *this directly for  post-3aeb78
+//   //operator at::ScalarType(){ return payload.; };
+// };
+
+#define DISPATCH_FLOAT_AND_HALF(TYPE, LEVEL, NAME, ...) \
+  switch(TYPE) \
+  { \
+    case at::ScalarType::Float: \
+    { \
+      using scalar_t_##LEVEL = float; \
+      __VA_ARGS__; \
+      break; \
+    } \
+    case at::ScalarType::Half: \
+    { \
+      using scalar_t_##LEVEL = at::Half; \
+      __VA_ARGS__; \
+      break; \
+    } \
+    default: \
+      AT_ERROR(#NAME, " not implemented for '", toString(TYPE), "'");  \
+  }
+
+
+#define DISPATCH_FLOAT_HALF_AND_BYTE(TYPE, LEVEL, NAME, ...) \
+  switch(TYPE) \
+  { \
+    case at::ScalarType::Float: \
+    { \
+      using scalar_t_##LEVEL = float; \
+      __VA_ARGS__; \
+      break; \
+    } \
+    case at::ScalarType::Half: \
+    { \
+      using scalar_t_##LEVEL = at::Half; \
+      __VA_ARGS__; \
+      break; \
+    } \
+    case at::ScalarType::Byte: \
+    { \
+      using scalar_t_##LEVEL = uint8_t; \
+      __VA_ARGS__; \
+      break; \
+    } \
+    default: \
+      AT_ERROR(#NAME, " not implemented for '", toString(TYPE), "'");  \
+  }
+
+
+#define DISPATCH_DOUBLE_FLOAT_AND_HALF(TYPE, LEVEL, NAME, ...) \
+  switch(TYPE) \
+  { \
+    case at::ScalarType::Double: \
+    { \
+      using scalar_t_##LEVEL = double; \
+      __VA_ARGS__; \
+      break; \
+    } \
+    case at::ScalarType::Float: \
+    { \
+      using scalar_t_##LEVEL = float; \
+      __VA_ARGS__; \
+      break; \
+    } \
+    case at::ScalarType::Half: \
+    { \
+      using scalar_t_##LEVEL = at::Half; \
+      __VA_ARGS__; \
+      break; \
+    } \
+    default: \
+      AT_ERROR(#NAME, " not implemented for '", toString(TYPE), "'");  \
+  }
+
+
+  #define DISPATCH_DOUBLE_AND_FLOAT(TYPE, LEVEL, NAME, ...) \
+  switch(TYPE) \
+  { \
+    case at::ScalarType::Double: \
+    { \
+      using scalar_t_##LEVEL = double; \
+      __VA_ARGS__; \
+      break; \
+    } \
+    case at::ScalarType::Float: \
+    { \
+      using scalar_t_##LEVEL = float; \
+      __VA_ARGS__; \
+      break; \
+    } \
+    default: \
+      AT_ERROR(#NAME, " not implemented for '", toString(TYPE), "'");  \
+  }
+
+
+template<typename T>
+__device__ __forceinline__ T reduce_block_into_lanes
+  (T *x,
+   T val,
+   int lanes=1,
+   bool share_result=false) // lanes is intended to be <= 32.
+{
+  int tid = threadIdx.x + threadIdx.y*blockDim.x;
+  int blockSize = blockDim.x*blockDim.y; // blockSize is intended to be a multiple of 32.
+
+  if(blockSize >= 64)
+  {
+    x[tid] = val;
+    __syncthreads();
+  }
+
+  #pragma unroll
+  for(int i = (blockSize >> 1); i >= 64; i >>= 1)
+  {
+    if(tid < i)
+      x[tid] = x[tid] + x[tid+i];
+    __syncthreads();
+  }
+
+  T final;
+
+  if(tid < 32)
+  {
+    if(blockSize >= 64)
+      final = x[tid] + x[tid+32];
+    else
+      final = val;
+    // __SYNCWARP();
+
+    #pragma unroll
+    for(int i = 16; i >= lanes; i >>= 1)
+      final = final + __shfl_down_sync(0xffffffff, final, i);
+  }
+
+  if(share_result)
+  {
+    if(tid < lanes)
+      x[tid] = final; // EpilogueOp
+    // Make sure the smem result is visible to all warps.
+    __syncthreads();
+  }
+
+  return final;
+}
+
+template<typename T>
+__device__ __forceinline__ T reduce_block_into_lanes_max_op
+  (T *x,
+   T val,
+   int lanes=1,
+   bool share_result=false) // lanes is intended to be <= 32.
+{
+  int tid = threadIdx.x + threadIdx.y*blockDim.x;
+  int blockSize = blockDim.x*blockDim.y; // blockSize is intended to be a multiple of 32.
+
+  if(blockSize >= 64)
+  {
+    x[tid] = val;
+    __syncthreads();
+  }
+
+  #pragma unroll
+  for(int i = (blockSize >> 1); i >= 64; i >>= 1)
+  {
+    if(tid < i)
+      x[tid] = fmaxf(fabsf(x[tid]), fabsf(x[tid+i]));
+    __syncthreads();
+  }
+
+  T final;
+
+  if(tid < 32)
+  {
+    if(blockSize >= 64)
+      final = fmaxf(fabsf(x[tid]), fabsf(x[tid+32]));
+    else
+      final = val;
+    // __SYNCWARP();
+
+    #pragma unroll
+    for(int i = 16; i >= lanes; i >>= 1)
+      final = fmaxf(fabsf(final), fabsf(__shfl_down_sync(0xffffffff, final, i)));
+  }
+
+  if(share_result)
+  {
+    if(tid < lanes)
+      x[tid] = final; // EpilogueOp
+    // Make sure the smem result is visible to all warps.
+    __syncthreads();
+  }
+
+  return final;
+}

megatron/model/__init__.py

@@ -13,9 +13,27 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+_LAYER_NORM = None
+
+
+def import_layernorm(fp32_residual_connection):
+
+    global _LAYER_NORM
+    if not _LAYER_NORM:
+        if fp32_residual_connection:
+            from .fused_layer_norm import MixedFusedLayerNorm as LayerNorm
+        else:
+            from apex.normalization.fused_layer_norm import FusedLayerNorm as LayerNorm
+        _LAYER_NORM = LayerNorm
+            
+    return _LAYER_NORM
+
+
 from .distributed import *
 from .bert_model import BertModel, BertModelFirstStage, BertModelIntermediateStage, BertModelLastStage
 from .realm_model import ICTBertModel
 from .gpt2_model import GPT2Model, GPT2ModelFirstStage, GPT2ModelIntermediateStage, GPT2ModelLastStage
 from .utils import get_params_for_weight_decay_optimization
 from .language_model import get_language_model
+
+

megatron/model/bert_model.py

@@ -21,7 +21,7 @@ from megatron import get_args
 from megatron import mpu
 from megatron.model.language_model import parallel_lm_logits
 from megatron.model.language_model import get_language_model
-from megatron.model.transformer import LayerNorm
+from megatron.model import import_layernorm
 from megatron.model.utils import openai_gelu, erf_gelu
 from megatron.model.utils import get_linear_layer
 from megatron.model.utils import init_method_normal
@@ -83,6 +83,7 @@ class BertLMHead(MegatronModule):
         self.parallel_output = parallel_output
 
         self.dense = get_linear_layer(hidden_size, hidden_size, init_method)
+        LayerNorm = import_layernorm(args.fp32_residual_connection)
         self.layernorm = LayerNorm(hidden_size, eps=layernorm_epsilon)
         self.gelu = torch.nn.functional.gelu
         if args.openai_gelu:

megatron/model/fused_layer_norm.py

+# coding=utf-8
+# Copyright (c) 2020, NVIDIA CORPORATION.  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.
+
+"""This code is copied fron NVIDIA apex:
+      https://github.com/NVIDIA/apex
+   with minor changes. """
+
+
+import math
+import torch
+import numbers
+from torch.nn.parameter import Parameter
+from torch.nn import init
+from torch.nn import functional as F
+import importlib
+
+global fused_layer_norm_cuda
+fused_layer_norm_cuda = None
+global fused_mix_prec_layer_norm_cuda
+fused_mix_prec_layer_norm_cuda = None
+
+class FusedLayerNormAffineFunction(torch.autograd.Function):
+
+  @staticmethod
+  def forward(ctx, input, weight, bias, normalized_shape, eps):
+    global fused_mix_prec_layer_norm_cuda
+    if fused_mix_prec_layer_norm_cuda is None:
+        fused_mix_prec_layer_norm_cuda = importlib.import_module("fused_mix_prec_layer_norm_cuda")
+    ctx.normalized_shape = normalized_shape
+    ctx.eps = eps
+    input_ = input.contiguous()
+    weight_ = weight.contiguous()
+    bias_ = bias.contiguous()
+    output, mean, invvar = fused_mix_prec_layer_norm_cuda.forward_affine(
+        input_, ctx.normalized_shape, weight_, bias_, ctx.eps)
+    ctx.save_for_backward(input_, weight_, bias_, mean, invvar)
+    return output
+
+  @staticmethod
+  def backward(ctx, grad_output):
+    input_, weight_, bias_, mean, invvar = ctx.saved_tensors
+    grad_input = grad_weight = grad_bias = None
+    grad_input, grad_weight, grad_bias = fused_mix_prec_layer_norm_cuda.backward_affine(
+        grad_output.contiguous(), mean, invvar,
+        input_, ctx.normalized_shape,
+        weight_, bias_, ctx.eps)
+    return grad_input, grad_weight, grad_bias, None, None
+
+class FusedLayerNormFunction(torch.autograd.Function):
+
+  @staticmethod
+  def forward(ctx, input, normalized_shape, eps):
+    global fused_layer_norm_cuda
+    if fused_layer_norm_cuda is None:
+        fused_layer_norm_cuda = importlib.import_module("fused_layer_norm_cuda")
+    ctx.normalized_shape = normalized_shape
+    ctx.eps = eps
+    input_ = input.contiguous()
+    output, mean, invvar = fused_layer_norm_cuda.forward(
+        input_, ctx.normalized_shape, ctx.eps)
+    ctx.save_for_backward(input_, mean, invvar)
+    return output
+
+  @staticmethod
+  def backward(ctx, grad_output):
+    input_, mean, invvar = ctx.saved_tensors
+    grad_input = None
+    grad_input = fused_layer_norm_cuda.backward(
+        grad_output.contiguous(), mean, invvar,
+        input_, ctx.normalized_shape,
+        ctx.eps)
+    return grad_input, None, None
+
+def fused_layer_norm_affine(input, normalized_shape, weight, bias, eps=1e-6):
+    return FusedLayerNormAffineFunction.apply(input, weight, bias, normalized_shape, eps)
+
+def fused_layer_norm(input, normalized_shape, eps=1e-6):
+    return FusedLayerNormFunction.apply(input, normalized_shape, eps)
+
+class MixedFusedLayerNorm(torch.nn.Module):
+    r"""Applies Layer Normalization over a mini-batch of inputs as described in
+    the paper `Layer Normalization`_ .
+    Currently only runs on cuda() tensors.
+    .. math::
+        y = \frac{x - \mathrm{E}[x]}{ \sqrt{\mathrm{Var}[x] + \epsilon}} * \gamma + \beta
+    The mean and standard-deviation are calculated separately over the last
+    certain number dimensions which have to be of the shape specified by
+    :attr:`normalized_shape`.
+    :math:`\gamma` and :math:`\beta` are learnable affine transform parameters of
+    :attr:`normalized_shape` if :attr:`elementwise_affine` is ``True``.
+    .. note::
+        Unlike Batch Normalization and Instance Normalization, which applies
+        scalar scale and bias for each entire channel/plane with the
+        :attr:`affine` option, Layer Normalization applies per-element scale and
+        bias with :attr:`elementwise_affine`.
+    This layer uses statistics computed from input data in both training and
+    evaluation modes.
+    Args:
+        normalized_shape (int or list or torch.Size): input shape from an expected input
+            of size
+            .. math::
+                [* \times \text{normalized}\_\text{shape}[0] \times \text{normalized}\_\text{shape}[1]
+                    \times \ldots \times \text{normalized}\_\text{shape}[-1]]
+            If a single integer is used, it is treated as a singleton list, and this module will
+            normalize over the last dimension which is expected to be of that specific size.
+        eps: a value added to the denominator for numerical stability. Default: 1e-5
+        elementwise_affine: a boolean value that when set to ``True``, this module
+            has learnable per-element affine parameters initialized to ones (for weights)
+            and zeros (for biases). Default: ``True``.
+    Shape:
+        - Input: :math:`(N, *)`
+        - Output: :math:`(N, *)` (same shape as input)
+    Examples::
+        >>> input = torch.randn(20, 5, 10, 10)
+        >>> # With Learnable Parameters
+        >>> m = apex.normalization.FusedLayerNorm(input.size()[1:])
+        >>> # Without Learnable Parameters
+        >>> m = apex.normalization.FusedLayerNorm(input.size()[1:], elementwise_affine=False)
+        >>> # Normalize over last two dimensions
+        >>> m = apex.normalization.FusedLayerNorm([10, 10])
+        >>> # Normalize over last dimension of size 10
+        >>> m = apex.normalization.FusedLayerNorm(10)
+        >>> # Activating the module
+        >>> output = m(input)
+    .. _`Layer Normalization`: https://arxiv.org/abs/1607.06450
+    """
+    def __init__(self, normalized_shape, eps=1e-5, elementwise_affine=True):
+        super(MixedFusedLayerNorm, self).__init__()
+
+        global fused_layer_norm_cuda
+        fused_layer_norm_cuda = importlib.import_module("fused_layer_norm_cuda")
+        global fused_mix_prec_layer_norm_cuda
+        fused_mix_prec_layer_norm_cuda = importlib.import_module("fused_mix_prec_layer_norm_cuda")
+
+
+        if isinstance(normalized_shape, numbers.Integral):
+            normalized_shape = (normalized_shape,)
+        self.normalized_shape = torch.Size(normalized_shape)
+        self.eps = eps
+        self.elementwise_affine = elementwise_affine
+        if self.elementwise_affine:
+            self.weight = Parameter(torch.Tensor(*normalized_shape))
+            self.bias = Parameter(torch.Tensor(*normalized_shape))
+        else:
+            self.register_parameter('weight', None)
+            self.register_parameter('bias', None)
+        self.reset_parameters()
+
+    def reset_parameters(self):
+        if self.elementwise_affine:
+            init.ones_(self.weight)
+            init.zeros_(self.bias)
+
+    def forward(self, input):
+        if not input.is_cuda:
+            return  F.layer_norm(
+                input, self.normalized_shape, self.weight, self.bias, self.eps)
+        
+        if self.elementwise_affine:
+          return FusedLayerNormAffineFunction.apply(
+              input, self.weight, self.bias, self.normalized_shape,self.eps)
+        else:
+          return FusedLayerNormFunction.apply(input, self.normalized_shape, self.eps) 
+
+    def extra_repr(self):
+        return '{normalized_shape}, eps={eps}, ' \
+            'elementwise_affine={elementwise_affine}'.format(**self.__dict__)

megatron/model/transformer.py

@@ -21,9 +21,9 @@ import torch.nn.functional as F
 
 from megatron import get_args
 from megatron import mpu
-from megatron.mpu import LayerNorm
 from megatron.module import MegatronModule
 from megatron.checkpointing import get_checkpoint_version
+from megatron.model import import_layernorm
 from megatron.model.fused_softmax import FusedScaleMaskSoftmax
 from megatron.model.fused_bias_gelu import bias_gelu_impl
 from megatron.model.utils import openai_gelu, erf_gelu
@@ -404,6 +404,7 @@ class ParallelTransformerLayer(MegatronModule):
             = args.apply_residual_connection_post_layernorm
 
         # Layernorm on the input data.
+        LayerNorm = import_layernorm(args.fp32_residual_connection)
         self.input_layernorm = LayerNorm(
             args.hidden_size,
             eps=args.layernorm_epsilon)
@@ -500,6 +501,8 @@ class ParallelTransformer(MegatronModule):
         super(ParallelTransformer, self).__init__()
         args = get_args()
 
+        self.fp32_residual_connection = args.fp32_residual_connection
+
         # Store activation checkpoiting flag.
         self.checkpoint_activations = args.checkpoint_activations
         self.checkpoint_num_layers = args.checkpoint_num_layers
@@ -520,6 +523,7 @@ class ParallelTransformer(MegatronModule):
 
         if mpu.is_pipeline_last_stage():
             # Final layer norm before output.
+            LayerNorm = import_layernorm(args.fp32_residual_connection)
             self.final_layernorm = LayerNorm(
                 args.hidden_size,
                 eps=args.layernorm_epsilon)
@@ -564,7 +568,12 @@ class ParallelTransformer(MegatronModule):
 
         if mpu.is_pipeline_first_stage():
             # Data format change to avoid explicit tranposes : [b s h] --> [s b h].
-            hidden_states = hidden_states.transpose(0, 1).contiguous()
+            # If the input flag for fp32 residual connection is set, convert for float.
+            if self.fp32_residual_connection:
+                hidden_states = hidden_states.transpose(0, 1).contiguous().float()
+            # Otherwise, leave it as is.
+            else:
+                hidden_states = hidden_states.transpose(0, 1).contiguous()
 
         if self.checkpoint_activations:
             hidden_states = self._checkpointed_forward(hidden_states,

megatron/model/utils.py

@@ -19,8 +19,8 @@ import math
 
 import torch
 
-from .transformer import LayerNorm
-
+from megatron import get_args
+from megatron.model import import_layernorm
 
 def init_method_normal(sigma):
     """Init method based on N(0, sigma)."""
@@ -65,6 +65,10 @@ def get_params_for_weight_decay_optimization(module):
     """Divide params into with-weight-decay and without-weight-decay groups.
     Layernorms and baises will have no weight decay but the rest will.
     """
+
+    args = get_args()
+    LayerNorm = import_layernorm(args.fp32_residual_connection)
+    
     weight_decay_params = {'params': []}
     no_weight_decay_params = {'params': [], 'weight_decay': 0.0}
     for module_ in module.modules():

megatron/mpu/__init__.py

@@ -41,7 +41,6 @@ from .initialize import get_pipeline_model_parallel_world_size, set_pipeline_mod
 from .initialize import initialize_model_parallel
 from .initialize import model_parallel_is_initialized
 
-from .layers import LayerNorm
 from .layers import ColumnParallelLinear
 from .layers import RowParallelLinear
 from .layers import VocabParallelEmbedding

megatron/mpu/layers.py

@@ -25,16 +25,6 @@ import torch.nn.functional as F
 import torch.nn.init as init
 from torch.nn.parameter import Parameter
 
-try:
-    from apex.normalization.fused_layer_norm import FusedLayerNorm as LayerNorm
-    # Try to use FusedLayerNorm from Apex - this will trigger an error.
-    _ = LayerNorm(8, eps=1e-5)
-
-except Exception as e:
-    print('WARNING: APEX is not installed, using torch.nn.LayerNorm '
-          'instead of apex.normalization.FusedLayerNorm!')
-    from torch.nn import LayerNorm
-
 from .initialize import get_tensor_model_parallel_rank
 from .initialize import get_tensor_model_parallel_world_size
 from .mappings import copy_to_tensor_model_parallel_region

megatron/training.py

@@ -333,16 +333,19 @@ def communicate(tensor_send_next, tensor_send_prev, recv_forward, recv_backward)
     tensor_recv_prev = None
     tensor_recv_next = None
     tensor_shape = (args.seq_length, args.micro_batch_size, args.hidden_size)
+    dtype = args.params_dtype
+    if args.fp32_residual_connection:
+        dtype = torch.float
     if recv_forward:
         tensor_recv_prev = torch.empty(tensor_shape,
                                        requires_grad=True,
                                        device=torch.cuda.current_device(),
-                                       dtype=args.params_dtype)
+                                       dtype=dtype)
     if recv_backward:
         tensor_recv_next = torch.empty(tensor_shape,
                                        requires_grad=True,
                                        device=torch.cuda.current_device(),
-                                       dtype=args.params_dtype)
+                                       dtype=dtype)
 
     # Send tensors in both the forward and backward directions as appropriate.
     torch.distributed.ring_exchange(tensor_send_prev=tensor_send_prev,