From c3b4170a43ee941f7d8357eb9e21b6046fcbb256 Mon Sep 17 00:00:00 2001
From: yujiaoliang <1229775764@qq.com>
Date: Mon, 26 May 2025 22:45:03 +0800
Subject: [PATCH] cuda dtype transfer

---
 excuter/op-mem-cuda/src/deepx/dtype_cuda.hpp | 22 ++++++++++++++++++++
 1 file changed, 22 insertions(+)

diff --git a/excuter/op-mem-cuda/src/deepx/dtype_cuda.hpp b/excuter/op-mem-cuda/src/deepx/dtype_cuda.hpp
index 57ef5ebd..98f212c6 100644
--- a/excuter/op-mem-cuda/src/deepx/dtype_cuda.hpp
+++ b/excuter/op-mem-cuda/src/deepx/dtype_cuda.hpp
@@ -3,6 +3,7 @@
 
 #include <cuda_fp16.h>
 #include <cuda_bf16.h>
+#include <cuda_fp8.h>
 
 #include "deepx/dtype.hpp"
 
@@ -34,6 +35,27 @@ namespace deepx
         else
             return Precision::Any;
     }   
+
+
+    template <>
+    struct to_tensor_type<PrecisionWrapper<Precision::BFloat16>> {
+        using type = nv_bfloat16;
+    };
+
+    template <>
+    struct to_tensor_type<PrecisionWrapper<Precision::Float16>> {
+        using type = half;
+    };
+
+    template <>
+    struct to_tensor_type<PrecisionWrapper<Precision::Float8E5M2>> {
+        using type =  __nv_fp8_e5m2;
+    };
+
+    template <>
+    struct to_tensor_type<PrecisionWrapper<Precision::Float8E4M3>> {
+        using type = __nv_fp8_e4m3;
+    };
 }
 
 #endif // DEEPX_DTYPE_CUDA_HPP