diff --git a/onnxruntime/contrib_ops/cuda/bert/fast_gelu.cc b/onnxruntime/contrib_ops/cuda/bert/fast_gelu.cc index 864c614b99..a547089232 100644 --- a/onnxruntime/contrib_ops/cuda/bert/fast_gelu.cc +++ b/onnxruntime/contrib_ops/cuda/bert/fast_gelu.cc @@ -34,7 +34,7 @@ using namespace ONNX_NAMESPACE; template FastGelu::FastGelu(const OpKernelInfo& op_kernel_info) : CudaKernel(op_kernel_info) { const TransformerOptions* options = TransformerOptions::GetInstance(); - use_half2_ = !options->DisableHalf2(); + use_half2_ = options->EnableHalf2(); } template diff --git a/onnxruntime/contrib_ops/cuda/bert/transformer_common.cc b/onnxruntime/contrib_ops/cuda/bert/transformer_common.cc index 847ad11cbf..bac94463aa 100644 --- a/onnxruntime/contrib_ops/cuda/bert/transformer_common.cc +++ b/onnxruntime/contrib_ops/cuda/bert/transformer_common.cc @@ -25,7 +25,7 @@ const TransformerOptions* TransformerOptions::GetInstance() { if (value > 0) std::cout << "ORT_TRANSFORMER_OPTIONS: IsPrecisionMode=" << instance.IsPrecisionMode() << ",DisablePersistentSoftmax=" << instance.DisablePersistentSoftmax() - << ",DisableHalf2=" << instance.DisableHalf2() + << ",EnableHalf2=" << instance.EnableHalf2() << std::endl; } diff --git a/onnxruntime/contrib_ops/cuda/bert/transformer_common.h b/onnxruntime/contrib_ops/cuda/bert/transformer_common.h index 74c1b2522d..1ef75639c1 100644 --- a/onnxruntime/contrib_ops/cuda/bert/transformer_common.h +++ b/onnxruntime/contrib_ops/cuda/bert/transformer_common.h @@ -15,12 +15,12 @@ class TransformerOptions { bool DisablePersistentSoftmax() const { return disable_persistent_softmax_; } - bool DisableHalf2() const { return disable_half2_; } + bool EnableHalf2() const { return enable_half2_; } void Initialize(int value) { is_precision_mode_ = (value & 0x01) > 0; disable_persistent_softmax_ = (value & 0x02) > 0; - disable_half2_ = (value & 0x04) > 0; + enable_half2_ = (value & 0x04) > 0; initialized_ = true; } @@ -31,8 +31,8 @@ class TransformerOptions { // Disable persistent softmax. bool disable_persistent_softmax_{false}; - // Disable half2 kernel. - bool disable_half2_{false}; + // Enable half2 kernel. + bool enable_half2_{false}; bool initialized_{false};