diff --git a/modules/esrgan_model.py b/modules/esrgan_model.py
index 3dcef5a6e..2ed1d2739 100644
--- a/modules/esrgan_model.py
+++ b/modules/esrgan_model.py
@@ -14,8 +14,11 @@ import modules.images
 
 def load_model(filename):
     # this code is adapted from https://github.com/xinntao/ESRGAN
-
-    pretrained_net = torch.load(filename)
+    if torch.has_mps:
+        map_l = 'cpu'
+    else:
+        map_l = None
+    pretrained_net = torch.load(filename, map_location=map_l)
     crt_model = arch.RRDBNet(3, 3, 64, 23, gc=32)
 
     if 'conv_first.weight' in pretrained_net:
diff --git a/modules/lowvram.py b/modules/lowvram.py
index 4b78deab7..bd1174915 100644
--- a/modules/lowvram.py
+++ b/modules/lowvram.py
@@ -2,9 +2,12 @@ import torch
 
 module_in_gpu = None
 cpu = torch.device("cpu")
-gpu = torch.device("cuda")
-device = gpu if torch.cuda.is_available() else cpu
-
+if torch.has_cuda:
+    device = gpu = torch.device("cuda")
+elif torch.has_mps:
+    device = gpu = torch.device("mps")
+else:
+    device = gpu = torch.device("cpu")
 
 def setup_for_low_vram(sd_model, use_medvram):
     parents = {}
diff --git a/modules/sd_hijack.py b/modules/sd_hijack.py
index 2d26b5f71..1084e2484 100644
--- a/modules/sd_hijack.py
+++ b/modules/sd_hijack.py
@@ -232,7 +232,7 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module):
         z = outputs.last_hidden_state
 
         # restoring original mean is likely not correct, but it seems to work well to prevent artifacts that happen otherwise
-        batch_multipliers = torch.asarray(np.array(batch_multipliers)).to(device)
+        batch_multipliers = torch.asarray(batch_multipliers).to(device)
         original_mean = z.mean()
         z *= batch_multipliers.reshape(batch_multipliers.shape + (1,)).expand(z.shape)
         new_mean = z.mean()
diff --git a/modules/shared.py b/modules/shared.py
index beb6f9bb0..e529ec27a 100644
--- a/modules/shared.py
+++ b/modules/shared.py
@@ -36,9 +36,12 @@ parser.add_argument("--opt-split-attention", action='store_true', help="enable o
 parser.add_argument("--listen", action='store_true', help="launch gradio with 0.0.0.0 as server name, allowing to respond to network requests")
 cmd_opts = parser.parse_args()
 
-cpu = torch.device("cpu")
-gpu = torch.device("cuda")
-device = gpu if torch.cuda.is_available() else cpu
+if torch.has_cuda:
+    device = torch.device("cuda")
+elif torch.has_mps:
+    device = torch.device("mps")
+else:
+    device = torch.device("cpu")
 batch_cond_uncond = cmd_opts.always_batch_cond_uncond or not (cmd_opts.lowvram or cmd_opts.medvram)
 parallel_processing_allowed = not cmd_opts.lowvram and not cmd_opts.medvram