diff --git a/src/transformers/models/swin2sr/modeling_swin2sr.py b/src/transformers/models/swin2sr/modeling_swin2sr.py
index fb694ad7efea00..b0a773c8af3472 100644
--- a/src/transformers/models/swin2sr/modeling_swin2sr.py
+++ b/src/transformers/models/swin2sr/modeling_swin2sr.py
@@ -301,6 +301,8 @@ def __init__(self, config, dim, num_heads, window_size, pretrained_window_size=[
         relative_coords_table = (
             torch.sign(relative_coords_table) * torch.log2(torch.abs(relative_coords_table) + 1.0) / math.log2(8)
         )
+        # set to same dtype as mlp weight
+        relative_coords_table = relative_coords_table.to(next(self.continuous_position_bias_mlp.parameters()).dtype)
         self.register_buffer("relative_coords_table", relative_coords_table, persistent=False)
 
         # get pair-wise relative position index for each token inside the window
diff --git a/src/transformers/models/swinv2/modeling_swinv2.py b/src/transformers/models/swinv2/modeling_swinv2.py
index b0682eacaade3d..b3e037c60b9d9b 100644
--- a/src/transformers/models/swinv2/modeling_swinv2.py
+++ b/src/transformers/models/swinv2/modeling_swinv2.py
@@ -492,6 +492,8 @@ def __init__(self, config, dim, num_heads, window_size, pretrained_window_size=[
         relative_coords_table = (
             torch.sign(relative_coords_table) * torch.log2(torch.abs(relative_coords_table) + 1.0) / math.log2(8)
         )
+        # set to same dtype as mlp weight
+        relative_coords_table = relative_coords_table.to(next(self.continuous_position_bias_mlp.parameters()).dtype)
         self.register_buffer("relative_coords_table", relative_coords_table, persistent=False)
 
         # get pair-wise relative position index for each token inside the window
diff --git a/tests/models/swin2sr/test_modeling_swin2sr.py b/tests/models/swin2sr/test_modeling_swin2sr.py
index ffcbca4c881960..a1358c9de0bfec 100644
--- a/tests/models/swin2sr/test_modeling_swin2sr.py
+++ b/tests/models/swin2sr/test_modeling_swin2sr.py
@@ -333,3 +333,24 @@ def test_inference_image_super_resolution_head(self):
             [[0.5458, 0.5546, 0.5638], [0.5526, 0.5565, 0.5651], [0.5396, 0.5426, 0.5621]]
         ).to(torch_device)
         self.assertTrue(torch.allclose(outputs.reconstruction[0, 0, :3, :3], expected_slice, atol=1e-4))
+
+    def test_inference_fp16(self):
+        processor = Swin2SRImageProcessor()
+        model = Swin2SRForImageSuperResolution.from_pretrained(
+            "caidas/swin2SR-classical-sr-x2-64", torch_dtype=torch.float16
+        ).to(torch_device)
+
+        image = Image.open("./tests/fixtures/tests_samples/COCO/000000039769.png")
+        inputs = processor(images=image, return_tensors="pt").to(model.dtype).to(torch_device)
+
+        # forward pass
+        with torch.no_grad():
+            outputs = model(**inputs)
+
+        # verify the logits
+        expected_shape = torch.Size([1, 3, 976, 1296])
+        self.assertEqual(outputs.reconstruction.shape, expected_shape)
+        expected_slice = torch.tensor(
+            [[0.5454, 0.5542, 0.5640], [0.5518, 0.5562, 0.5649], [0.5391, 0.5425, 0.5620]], dtype=model.dtype
+        ).to(torch_device)
+        self.assertTrue(torch.allclose(outputs.reconstruction[0, 0, :3, :3], expected_slice, atol=1e-4))
diff --git a/tests/models/swinv2/test_modeling_swinv2.py b/tests/models/swinv2/test_modeling_swinv2.py
index cef6c34debaca0..5ef9a4b92e1388 100644
--- a/tests/models/swinv2/test_modeling_swinv2.py
+++ b/tests/models/swinv2/test_modeling_swinv2.py
@@ -487,6 +487,26 @@ def test_inference_image_classification_head(self):
         expected_slice = torch.tensor([-0.3947, -0.4306, 0.0026]).to(torch_device)
         self.assertTrue(torch.allclose(outputs.logits[0, :3], expected_slice, atol=1e-4))
 
+    @slow
+    def test_inference_fp16(self):
+        model = Swinv2ForImageClassification.from_pretrained(
+            "microsoft/swinv2-tiny-patch4-window8-256", torch_dtype=torch.float16
+        ).to(torch_device)
+        image_processor = self.default_image_processor
+
+        image = Image.open("./tests/fixtures/tests_samples/COCO/000000039769.png")
+        inputs = image_processor(images=image, return_tensors="pt").to(model.dtype).to(torch_device)
+
+        # forward pass
+        with torch.no_grad():
+            outputs = model(**inputs)
+
+        # verify the logits
+        expected_shape = torch.Size((1, 1000))
+        self.assertEqual(outputs.logits.shape, expected_shape)
+        expected_slice = torch.tensor([-0.3938, -0.4290, 0.0020], dtype=model.dtype).to(torch_device)
+        self.assertTrue(torch.allclose(outputs.logits[0, :3], expected_slice, atol=1e-4))
+
     @slow
     def test_inference_interpolate_pos_encoding(self):
         # Swinv2 models have an `interpolate_pos_encoding` argument in their forward method,