Spaces:
Runtime error
Runtime error
File size: 8,188 Bytes
8a42f8f |
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 |
#include <torch/extension.h>
#include <torch/torch.h>
#include <vector>
#include <stdio.h>
template <typename T>
int linear_bias_forward_cuda(at::Tensor input, T *weight, at::Tensor bias, int in_features, int batch_size, int out_features, at::Tensor output, void *lt_workspace);
template <typename T>
int linear_bias_backward_cuda(T *input, T *weight, T *d_output, int in_features, int batch_size, int out_features, T *d_weight, T *d_bias, T *d_input, void *lt_workspace);
template <typename T>
int linear_gelu_linear_forward_cuda(T *input, T *weight1, T *bias1, T *weight2, T *bias2, int in_features, int hidden_features, int batch_size, int out_features, T *output1, T *output2, T *gelu_in, void *lt_workspace) ;
template <typename T>
int linear_gelu_linear_backward_cuda(T *input, T *gelu_in, T *output1, T *weight1, T *weight2, T *d_output1, T *d_output2, int in_features, int batch_size, int hidden_features, int out_features, T *d_weight1, T *d_weight2, T *d_bias1, T *d_bias2, T *d_input, void *lt_workspace);
at::Tensor linear_bias_forward(at::Tensor input, at::Tensor weight, at::Tensor bias) {
auto batch_size = input.size(0);
auto in_features = input.size(1);
int out_features = weight.size(0);
//auto reserved_size = get_mlp_reserved_space(batch_size, num_layers, output_features.data());
// create output/workspace tensor
auto out = at::empty({batch_size, out_features}, input.type());
//auto reserved_space = at::empty({reserved_size}, inputs[0].type());
// allocate fixed 4MB workspace for cublaslt for now, and this gets at least 4 MB
auto lt_workspace = at::empty({1 << 22}, input.type());
AT_DISPATCH_FLOATING_TYPES_AND2(at::ScalarType::Half, at::ScalarType::BFloat16, input.scalar_type(), "linear_bias_forward", [&] {
scalar_t* w_ptr = weight.data_ptr<scalar_t>();
scalar_t* b_ptr = bias.data_ptr<scalar_t>();
auto result = linear_bias_forward_cuda<scalar_t>(
input,
w_ptr,
bias,
in_features,
batch_size,
out_features,
out,
//out.data_ptr<scalar_t>(),
// reserved_space.data_ptr<scalar_t>(),
(void*) (lt_workspace.data_ptr<scalar_t>()));
});
return {out};
}
std::vector<at::Tensor> linear_bias_backward(at::Tensor input, at::Tensor weight, at::Tensor d_output) {
auto batch_size = input.size(0);
auto in_features = input.size(1);
int out_features = weight.size(0);
//auto reserved_size = get_mlp_reserved_space(batch_size, num_layers, output_features.data());
// create output/workspace tensor
auto d_weight = at::empty({out_features, in_features}, input.type());
#if defined(CUBLAS_VERSION) && CUBLAS_VERSION < 11600
auto d_bias = d_output.view({-1, out_features}).sum(0, false);
#else
auto d_bias = at::empty({out_features}, input.type());
#endif
auto d_input = at::empty({batch_size, in_features}, input.type());
//auto reserved_space = at::empty({reserved_size}, inputs[0].type());
// allocate fixed 4MB workspace for cublaslt for now, and this gets at least 4 MB
auto lt_workspace = at::empty({1 << 22}, input.type());
AT_DISPATCH_FLOATING_TYPES_AND2(at::ScalarType::Half, at::ScalarType::BFloat16, input.scalar_type(), "linear_bias_backward", [&] {
scalar_t* w_ptr = weight.data_ptr<scalar_t>();
scalar_t* d_b_ptr = d_bias.data_ptr<scalar_t>();
auto result = linear_bias_backward_cuda<scalar_t>(
input.data_ptr<scalar_t>(),
w_ptr,
d_output.data_ptr<scalar_t>(),
in_features,
batch_size,
out_features,
d_weight.data_ptr<scalar_t>(),
d_bias.data_ptr<scalar_t>(),
d_input.data_ptr<scalar_t>(),
// reserved_space.data_ptr<scalar_t>(),
(void*) (lt_workspace.data_ptr<scalar_t>()));
});
return {d_input, d_weight, d_bias};
}
std::vector<at::Tensor> linear_gelu_linear_forward(at::Tensor input, at::Tensor weight1, at::Tensor bias1, at::Tensor weight2, at::Tensor bias2) {
auto batch_size = input.size(0);
auto in_features = input.size(1);
int hidden_features = weight1.size(0);
int out_features = weight2.size(0);
//auto reserved_size = get_mlp_reserved_space(batch_size, num_layers, output_features.data());
// create output/workspace tensor
auto output1 = at::empty({batch_size, hidden_features}, input.type());
auto gelu_in = at::empty({batch_size, hidden_features}, input.type());
auto output2 = at::empty({batch_size, out_features}, input.type());
//auto reserved_space = at::empty({reserved_size}, inputs[0].type());
// allocate fixed 4MB workspace for cublaslt for now, and this gets at least 4 MB
auto lt_workspace = at::empty({1 << 22}, input.type());
AT_DISPATCH_FLOATING_TYPES_AND2(at::ScalarType::Half, at::ScalarType::BFloat16, input.scalar_type(), "linear_gelu_linear_forward", [&] {
scalar_t* w1_ptr = weight1.data_ptr<scalar_t>();
scalar_t* b1_ptr = bias1.data_ptr<scalar_t>();
scalar_t* w2_ptr = weight2.data_ptr<scalar_t>();
scalar_t* b2_ptr = bias2.data_ptr<scalar_t>();
auto result = linear_gelu_linear_forward_cuda<scalar_t>(
input.data_ptr<scalar_t>(),
w1_ptr,
b1_ptr,
w2_ptr,
b2_ptr,
in_features,
hidden_features,
batch_size,
out_features,
output1.data_ptr<scalar_t>(),
output2.data_ptr<scalar_t>(),
gelu_in.data_ptr<scalar_t>(),
// reserved_space.data_ptr<scalar_t>(),
(void*) (lt_workspace.data_ptr<scalar_t>()));
});
return {output1, output2, gelu_in};
}
std::vector<at::Tensor> linear_gelu_linear_backward(at::Tensor input, at::Tensor gelu_in, at::Tensor output1, at::Tensor weight1, at::Tensor weight2, at::Tensor d_output2) {
auto batch_size = input.size(0);
auto in_features = input.size(1);
int hidden_features = weight1.size(0);
int out_features = weight2.size(0);
//auto reserved_size = get_mlp_reserved_space(batch_size, num_layers, output_features.data());
// create output/workspace tensor
auto d_weight1 = at::empty({hidden_features, in_features}, input.type());
auto d_weight2 = at::empty({out_features, hidden_features}, input.type());
auto d_bias1 = at::empty({hidden_features}, input.type());
auto d_bias2 = at::empty({out_features}, input.type());
auto d_input = at::empty({batch_size, in_features}, input.type());
auto d_output1 = at::empty({batch_size, hidden_features}, input.type());
//auto reserved_space = at::empty({reserved_size}, inputs[0].type());
// allocate fixed 4MB workspace for cublaslt for now, and this gets at least 4 MB
auto lt_workspace = at::empty({1 << 22}, input.type());
AT_DISPATCH_FLOATING_TYPES_AND2(at::ScalarType::Half, at::ScalarType::BFloat16, input.scalar_type(), "linear_bias_backward", [&] {
//scalar_t* w_ptr = weight.data_ptr<scalar_t>();
//scalar_t* d_b_ptr = d_bias.data_ptr<scalar_t>();
auto result = linear_gelu_linear_backward_cuda<scalar_t>(
input.data_ptr<scalar_t>(),
gelu_in.data_ptr<scalar_t>(),
output1.data_ptr<scalar_t>(),
weight1.data_ptr<scalar_t>(),
weight2.data_ptr<scalar_t>(),
d_output1.data_ptr<scalar_t>(),
d_output2.data_ptr<scalar_t>(),
in_features,
batch_size,
hidden_features,
out_features,
d_weight1.data_ptr<scalar_t>(),
d_weight2.data_ptr<scalar_t>(),
d_bias1.data_ptr<scalar_t>(),
d_bias2.data_ptr<scalar_t>(),
d_input.data_ptr<scalar_t>(),
// reserved_space.data_ptr<scalar_t>(),
(void*) (lt_workspace.data_ptr<scalar_t>()));
});
return {d_input, d_weight1, d_bias1, d_weight2, d_bias2};
}
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
m.def("linear_bias_forward", &linear_bias_forward, "linear bias forward");
m.def("linear_bias_backward", &linear_bias_backward, "linear bias backward");
m.def("linear_gelu_linear_forward", &linear_gelu_linear_forward, "linear gelu linear forward");
m.def("linear_gelu_linear_backward", &linear_gelu_linear_backward, "linear gelu linear backward");
}
|