File size: 10,395 Bytes
e45d058
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
// Adapted from https://github.com/NVIDIA/apex/blob/master/csrc/fused_dense.cpp
// We make it work for bfloat16
#include <torch/extension.h>
#include <torch/torch.h>
#include <ATen/cuda/CUDAContext.h>
#include <c10/cuda/CUDAGuard.h>
#include <vector>

#include <stdio.h>

#define CHECK_SHAPE(x, ...) TORCH_CHECK(x.sizes() == torch::IntArrayRef({__VA_ARGS__}), #x " must have shape (" #__VA_ARGS__ ")")

// https://github.com/NVIDIA/apex/blob/master/csrc/type_shim.h
// #if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800
#define DISPATCH_HALF_AND_BF16(TYPE, NAME, ...)                                \
  switch (TYPE) {                                                              \
  case at::ScalarType::Half: {                                                 \
    using scalar_t = at::Half;                                                 \
    __VA_ARGS__();                                                             \
    break;                                                                     \
  }                                                                            \
  case at::ScalarType::BFloat16: {                                             \
    using scalar_t = at::BFloat16;                                             \
    __VA_ARGS__();                                                             \
    break;                                                                     \
  }                                                                            \
  default:                                                                     \
    AT_ERROR(#NAME, " not implemented for '", toString(TYPE), "'");            \
  }

template <typename T>
int linear_bias_wgrad_cuda(const T *input, const T *d_output, int64_t in_features, int64_t batch_size, int64_t out_features, T *d_weight, T *d_bias, void *lt_workspace, size_t workspaceSize);

template <typename T>
int linear_act_forward_cuda(const T *input, const T *weight, const T *bias, int64_t in_features, int64_t batch_size, int64_t out_features, bool is_gelu, int heuristic, T *output, void *pre_act, void *lt_workspace, size_t workspaceSize);

template <typename T>
int bias_act_linear_dgrad_bgrad_cuda(const T *weight, const T *d_output, const void *pre_act, int64_t in_features, int64_t batch_size, int64_t out_features, bool is_gelu, int heuristic, T *d_input, T *d_bias, void *lt_workspace, size_t workspaceSize);

std::vector<at::Tensor> linear_bias_wgrad(at::Tensor input, at::Tensor d_output, bool has_d_bias) {

  int64_t batch_size = input.size(0);
  int64_t in_features = input.size(1);
  int64_t out_features = d_output.size(1);

  TORCH_CHECK(input.dtype() == torch::kFloat16 || input.dtype() == torch::kBFloat16);
  TORCH_CHECK(input.dtype() == d_output.dtype());
  TORCH_CHECK(input.is_cuda());
  TORCH_CHECK(d_output.is_cuda());
  TORCH_CHECK(input.is_contiguous());
  TORCH_CHECK(d_output.is_contiguous());
  CHECK_SHAPE(input, batch_size, in_features);
  CHECK_SHAPE(d_output, batch_size, out_features);

  // Otherwise the kernel will be launched from cuda:0 device
  // Cast to char to avoid compiler warning about narrowing
  at::cuda::CUDAGuard device_guard{(char)input.get_device()};

  // create output/workspace tensor
  auto opts = input.options();
  auto d_weight = at::empty({out_features, in_features}, opts);
  at::Tensor d_bias;
  if (has_d_bias) {
#if defined(CUBLAS_VERSION) && CUBLAS_VERSION < 11600
    d_bias = d_output.view({-1, out_features}).sum(0, false);
#else
    d_bias = at::empty({out_features}, opts);
#endif
  }
  // See https://github.com/pytorch/pytorch/issues/73328 for reasoning behind setting this to 1M.
  // However, Apex sets it to 4M and TransformerEngine sets to 32M for Hopper and 4M for other GPUs
  // https://github.com/NVIDIA/TransformerEngine/blob/a0f0065498bbcfc1da78cf9e8b166f5381613fbc/transformer_engine/pytorch/module.py#L91
  size_t workspaceSize = 1024 * 1024 * (at::cuda::getCurrentDeviceProperties()->major >= 9 ? 32 : 4);
  auto lt_workspace = at::empty({static_cast<int64_t>(workspaceSize)}, opts.dtype(torch::kUInt8));

  DISPATCH_HALF_AND_BF16(input.scalar_type(), "linear_bias_wgrad", [&] {
    auto result = linear_bias_wgrad_cuda<scalar_t>(
        input.data_ptr<scalar_t>(),
        d_output.data_ptr<scalar_t>(),
        in_features,
        batch_size,
        out_features,
        d_weight.data_ptr<scalar_t>(),
        has_d_bias ? d_bias.data_ptr<scalar_t>() : nullptr,
        (void*) (lt_workspace.data_ptr()),
        workspaceSize);
    TORCH_CHECK(result == 0, "linear_bias_wgrad failed.");
  });

  return {d_weight, d_bias};
}

std::vector<at::Tensor> linear_act_forward(at::Tensor input, at::Tensor weight,

                                           c10::optional<at::Tensor> bias_,

                                           bool is_gelu, bool save_pre_act, int heuristic) {

  int64_t batch_size = input.size(0);
  int64_t in_features = input.size(1);
  int64_t out_features = weight.size(0);

  TORCH_CHECK(input.dtype() == torch::kFloat16 || input.dtype() == torch::kBFloat16);
  TORCH_CHECK(input.dtype() == weight.dtype());
  TORCH_CHECK(input.is_cuda());
  TORCH_CHECK(weight.is_cuda());
  TORCH_CHECK(input.is_contiguous());
  TORCH_CHECK(weight.is_contiguous());
  CHECK_SHAPE(input, batch_size, in_features);
  CHECK_SHAPE(weight, out_features, in_features);
  if (bias_.has_value()) {
    auto bias = bias_.value();
    TORCH_CHECK(bias.dtype() == input.dtype());
    TORCH_CHECK(bias.is_cuda());
    TORCH_CHECK(bias.is_contiguous());
    CHECK_SHAPE(bias, out_features);
  }

  // Otherwise the kernel will be launched from cuda:0 device
  // Cast to char to avoid compiler warning about narrowing
  at::cuda::CUDAGuard device_guard{(char)input.get_device()};

  // create output/workspace tensor
  auto opts = input.options();
  auto output = at::empty({batch_size, out_features}, opts);
  at::Tensor pre_act;
  // If ReLU, cuBlasLT stores a bit-mask (1 bit per element)
  if (save_pre_act) { pre_act = at::empty({batch_size, is_gelu ? out_features : out_features / 8},
                                          is_gelu ? opts : opts.dtype(torch::kUInt8)); }
  // See https://github.com/pytorch/pytorch/issues/73328 for reasoning behind setting this to 1M.
  // However, Apex sets it to 4M and TransformerEngine sets to 32M for Hopper and 4M for other GPUs
  // https://github.com/NVIDIA/TransformerEngine/blob/a0f0065498bbcfc1da78cf9e8b166f5381613fbc/transformer_engine/pytorch/module.py#L91
  size_t workspaceSize = 1024 * 1024 * (at::cuda::getCurrentDeviceProperties()->major >= 9 ? 32 : 4);
  auto lt_workspace = at::empty({static_cast<int64_t>(workspaceSize)}, opts.dtype(torch::kUInt8));

  DISPATCH_HALF_AND_BF16(input.scalar_type(), "linear_act_forward", [&] {
    auto result = linear_act_forward_cuda<scalar_t>(
        input.data_ptr<scalar_t>(),
        weight.data_ptr<scalar_t>(),
        bias_.has_value()? bias_.value().data_ptr<scalar_t>() : nullptr,
        in_features,
        batch_size,
        out_features,
        is_gelu,
        heuristic,
        output.data_ptr<scalar_t>(),
        save_pre_act ? pre_act.data_ptr() : nullptr,
        (void*) (lt_workspace.data_ptr()),
        workspaceSize);
    TORCH_CHECK(result == 0, "linear_act_forward failed.");
  });

  std::vector<at::Tensor> result = {output};
  if (save_pre_act) { result.push_back(pre_act); };
  return result;
}

std::vector<at::Tensor> bias_act_linear_dgrad_bgrad(

  at::Tensor weight, at::Tensor d_output, at::Tensor pre_act, bool is_gelu, int heuristic

) {

  int64_t batch_size = d_output.size(0);
  int64_t out_features = d_output.size(1);
  int64_t in_features = weight.size(1);

  TORCH_CHECK(weight.dtype() == torch::kFloat16 || weight.dtype() == torch::kBFloat16);
  TORCH_CHECK(weight.dtype() == d_output.dtype());
  TORCH_CHECK(is_gelu ? (pre_act.dtype() == weight.dtype()) : (pre_act.dtype() == torch::kUInt8));
  TORCH_CHECK(weight.is_cuda());
  TORCH_CHECK(d_output.is_cuda());
  TORCH_CHECK(pre_act.is_cuda());
  TORCH_CHECK(weight.is_contiguous());
  TORCH_CHECK(d_output.is_contiguous());
  TORCH_CHECK(pre_act.is_contiguous());
  CHECK_SHAPE(weight, out_features, in_features);
  CHECK_SHAPE(d_output, batch_size, out_features);
  // If ReLU, cuBlasLT stores a bit-mask (1 bit per element)
  CHECK_SHAPE(pre_act, batch_size, is_gelu ? in_features : in_features / 8);

  // Otherwise the kernel will be launched from cuda:0 device
  // Cast to char to avoid compiler warning about narrowing
  at::cuda::CUDAGuard device_guard{(char)weight.get_device()};

  // create output/workspace tensor
  auto opts = weight.options();
  auto d_bias = at::empty({in_features}, opts);
  auto d_input = at::empty({batch_size, in_features}, opts);
  // See https://github.com/pytorch/pytorch/issues/73328 for reasoning behind setting this to 1M.
  // However, Apex sets it to 4M and TransformerEngine sets to 32M for Hopper and 4M for other GPUs
  // https://github.com/NVIDIA/TransformerEngine/blob/a0f0065498bbcfc1da78cf9e8b166f5381613fbc/transformer_engine/pytorch/module.py#L91
  size_t workspaceSize = 1024 * 1024 * (at::cuda::getCurrentDeviceProperties()->major >= 9 ? 32 : 4);
  auto lt_workspace = at::empty({static_cast<int64_t>(workspaceSize)}, opts.dtype(torch::kUInt8));

  DISPATCH_HALF_AND_BF16(weight.scalar_type(), "bias_act_linear_dgrad_bgrad", [&] {
    auto result = bias_act_linear_dgrad_bgrad_cuda<scalar_t>(
        weight.data_ptr<scalar_t>(),
        d_output.data_ptr<scalar_t>(),
        pre_act.data_ptr(),
        in_features,
        batch_size,
        out_features,
        is_gelu,
        heuristic,
        d_input.data_ptr<scalar_t>(),
        d_bias.data_ptr<scalar_t>(),
        (void*) (lt_workspace.data_ptr()),
        workspaceSize);
    TORCH_CHECK(result == 0, "bias_act_linear_dgrad_bgrad failed.");
  });

  return {d_input, d_bias};
}

PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
  m.def("linear_bias_wgrad", &linear_bias_wgrad, "linear bias wgrad");
  m.def("linear_act_forward", &linear_act_forward, "linear gelu/relu forward");
  m.def("bias_act_linear_dgrad_bgrad", &bias_act_linear_dgrad_bgrad, "bias gelu/relu linear dgrad bgrad");
}