issue/620 处理在python代码中tensor从gpu to到cpu上出错的问题

src/infinicore/tensor/tensor.cc

@@ -173,8 +173,6 @@ std::shared_ptr<TensorImpl> TensorImpl::empty(const Shape &shape,
     auto t = std::shared_ptr<TensorImpl>(new TensorImpl(shape, dtype));
     t->data_.offset = 0;
 
-    context::setDevice(device);
-
     if (device == Device::Type::CPU) {
         if (pin_memory) {
             if (context::getDevice() == Device::Type::CPU) {
@@ -187,6 +185,7 @@ std::shared_ptr<TensorImpl> TensorImpl::empty(const Shape &shape,
             t->data_.memory = context::allocateHostMemory(t->numel() * dsize(dtype));
         }
     } else {
+        context::setDevice(device);
         t->data_.memory = context::allocateMemory(t->numel() * dsize(dtype));
     }
 
@@ -203,8 +202,6 @@ std::shared_ptr<TensorImpl> TensorImpl::strided_empty(
     auto impl = std::shared_ptr<TensorImpl>(new TensorImpl(shape, strides, dtype));
     impl->data_.offset = 0;
 
-    context::setDevice(device);
-
     size_t max_offset = 0;
 
     for (size_t i = 0; i < shape.size(); ++i) {
@@ -228,6 +225,7 @@ std::shared_ptr<TensorImpl> TensorImpl::strided_empty(
             impl->data_.memory = context::allocateHostMemory(required_bytes);
         }
     } else {
+        context::setDevice(device);
         impl->data_.memory = context::allocateMemory(required_bytes);
     }
 

test/infinicore/test.py

@@ -107,7 +107,45 @@ def test3():
     print("abs error: ", torch.abs(ans_torch_ref - torch_ans_result).max())
 
 
+def test4_to():
+    """
+    解决在python代码中 tensor从gpu to到cpu上出错的问题.
+    """
+    if True:
+        x = torch.rand((2, 3), dtype=torch.float32, device="cpu")
+        x_infini = infinicore.from_torch(x.clone())
+        print(" ---------------> test 1")
+        x_infini.debug()
+        x_gpu = x_infini.to(infinicore.device("cuda", 0))
+        x_gpu = x_gpu.to(infinicore.device("cuda", 0))
+
+        x_gpu.debug()
+        x_cpu = x_infini.to(infinicore.device("cpu", 0))
+        x_cpu = x_cpu.to(infinicore.device("cpu", 0))
+
+        x_cpu.debug()
+
+    if True:
+        x = infinicore.empty(
+            (2, 3), dtype=infinicore.float32, device=infinicore.device("cuda", 0)
+        )
+        x.debug()
+        x.to(infinicore.device("cuda", 0))
+
+        x_cpu = x.to(infinicore.device("cpu", 0))
+        x_cpu.debug()
+
+        x_cpu_gpu = x_cpu.to(infinicore.device("cuda", 0))
+        x_cpu_gpu.debug()
+
+        x_gpu = x.to(infinicore.device("cuda", 0))
+        x_gpu.debug()
+
+    print(" 简单的测试用例，通过!!")
+
+
 if __name__ == "__main__":
-    # test()
+    test()
     test2()
     test3()
+    test4_to()