Thanks to visit codestin.com
Credit goes to github.com

Skip to content

Commit 1d6a48c

Browse files
crcrparcrcrpar
authored
Fix pr1951: use .options for .type, not .scalar_type (#1953)
* Revert "use `.scalar_type` not `.type` (#1951)" This reverts commit 419adeb. * use Tensor::options to replace Tensor::type, not Tensor::scalar_type Signed-off-by: Masaki Kozuki <[email protected]> --------- Signed-off-by: Masaki Kozuki <[email protected]>
1 parent 419adeb commit 1d6a48c

1 file changed

Lines changed: 23 additions & 23 deletions

File tree

‎csrc/fused_dense.cpp‎

Lines changed: 23 additions & 23 deletions
Original file line numberDiff line numberDiff line change
@@ -27,10 +27,10 @@ at::Tensor linear_bias_forward(at::Tensor input, at::Tensor weight, at::Tensor b
2727
//auto reserved_size = get_mlp_reserved_space(batch_size, num_layers, output_features.data());
2828

2929
// create output/workspace tensor
30-
auto out = at::empty({batch_size, out_features}, input.scalar_type());
31-
//auto reserved_space = at::empty({reserved_size}, inputs[0].scalar_type());
30+
auto out = at::empty({batch_size, out_features}, input.options());
31+
//auto reserved_space = at::empty({reserved_size}, inputs[0].options());
3232
// allocate fixed 4MB workspace for cublaslt for now, and this gets at least 4 MB
33-
auto lt_workspace = at::empty({1 << 22}, input.scalar_type());
33+
auto lt_workspace = at::empty({1 << 22}, input.options());
3434

3535
AT_DISPATCH_FLOATING_TYPES_AND2(at::ScalarType::Half, at::ScalarType::BFloat16, input.scalar_type(), "linear_bias_forward", [&] {
3636
scalar_t* w_ptr = weight.data_ptr<scalar_t>();
@@ -61,16 +61,16 @@ std::vector<at::Tensor> linear_bias_backward(at::Tensor input, at::Tensor weight
6161
//auto reserved_size = get_mlp_reserved_space(batch_size, num_layers, output_features.data());
6262

6363
// create output/workspace tensor
64-
auto d_weight = at::empty({out_features, in_features}, input.scalar_type());
64+
auto d_weight = at::empty({out_features, in_features}, input.options());
6565
#if defined(CUBLAS_VERSION) && CUBLAS_VERSION < 11600
6666
auto d_bias = d_output.view({-1, out_features}).sum(0, false);
67-
#else
68-
auto d_bias = at::empty({out_features}, input.scalar_type());
69-
#endif
70-
auto d_input = at::empty({batch_size, in_features}, input.scalar_type());
71-
//auto reserved_space = at::empty({reserved_size}, inputs[0].scalar_type());
67+
#else
68+
auto d_bias = at::empty({out_features}, input.options());
69+
#endif
70+
auto d_input = at::empty({batch_size, in_features}, input.options());
71+
//auto reserved_space = at::empty({reserved_size}, inputs[0].options());
7272
// allocate fixed 4MB workspace for cublaslt for now, and this gets at least 4 MB
73-
auto lt_workspace = at::empty({1 << 22}, input.scalar_type());
73+
auto lt_workspace = at::empty({1 << 22}, input.options());
7474

7575
AT_DISPATCH_FLOATING_TYPES_AND2(at::ScalarType::Half, at::ScalarType::BFloat16, input.scalar_type(), "linear_bias_backward", [&] {
7676
scalar_t* w_ptr = weight.data_ptr<scalar_t>();
@@ -103,12 +103,12 @@ std::vector<at::Tensor> linear_gelu_linear_forward(at::Tensor input, at::Tensor
103103
//auto reserved_size = get_mlp_reserved_space(batch_size, num_layers, output_features.data());
104104

105105
// create output/workspace tensor
106-
auto output1 = at::empty({batch_size, hidden_features}, input.scalar_type());
107-
auto gelu_in = at::empty({batch_size, hidden_features}, input.scalar_type());
108-
auto output2 = at::empty({batch_size, out_features}, input.scalar_type());
109-
//auto reserved_space = at::empty({reserved_size}, inputs[0].scalar_type());
106+
auto output1 = at::empty({batch_size, hidden_features}, input.options());
107+
auto gelu_in = at::empty({batch_size, hidden_features}, input.options());
108+
auto output2 = at::empty({batch_size, out_features}, input.options());
109+
//auto reserved_space = at::empty({reserved_size}, inputs[0].options());
110110
// allocate fixed 4MB workspace for cublaslt for now, and this gets at least 4 MB
111-
auto lt_workspace = at::empty({1 << 22}, input.scalar_type());
111+
auto lt_workspace = at::empty({1 << 22}, input.options());
112112

113113
AT_DISPATCH_FLOATING_TYPES_AND2(at::ScalarType::Half, at::ScalarType::BFloat16, input.scalar_type(), "linear_gelu_linear_forward", [&] {
114114
scalar_t* w1_ptr = weight1.data_ptr<scalar_t>();
@@ -146,15 +146,15 @@ std::vector<at::Tensor> linear_gelu_linear_backward(at::Tensor input, at::Tensor
146146
//auto reserved_size = get_mlp_reserved_space(batch_size, num_layers, output_features.data());
147147

148148
// create output/workspace tensor
149-
auto d_weight1 = at::empty({hidden_features, in_features}, input.scalar_type());
150-
auto d_weight2 = at::empty({out_features, hidden_features}, input.scalar_type());
151-
auto d_bias1 = at::empty({hidden_features}, input.scalar_type());
152-
auto d_bias2 = at::empty({out_features}, input.scalar_type());
153-
auto d_input = at::empty({batch_size, in_features}, input.scalar_type());
154-
auto d_output1 = at::empty({batch_size, hidden_features}, input.scalar_type());
155-
//auto reserved_space = at::empty({reserved_size}, inputs[0].scalar_type());
149+
auto d_weight1 = at::empty({hidden_features, in_features}, input.options());
150+
auto d_weight2 = at::empty({out_features, hidden_features}, input.options());
151+
auto d_bias1 = at::empty({hidden_features}, input.options());
152+
auto d_bias2 = at::empty({out_features}, input.options());
153+
auto d_input = at::empty({batch_size, in_features}, input.options());
154+
auto d_output1 = at::empty({batch_size, hidden_features}, input.options());
155+
//auto reserved_space = at::empty({reserved_size}, inputs[0].options());
156156
// allocate fixed 4MB workspace for cublaslt for now, and this gets at least 4 MB
157-
auto lt_workspace = at::empty({1 << 22}, input.scalar_type());
157+
auto lt_workspace = at::empty({1 << 22}, input.options());
158158

159159
AT_DISPATCH_FLOATING_TYPES_AND2(at::ScalarType::Half, at::ScalarType::BFloat16, input.scalar_type(), "linear_bias_backward", [&] {
160160

0 commit comments

Comments
 (0)