diff --git a/elk/utils/gpu_utils.py b/elk/utils/gpu_utils.py
index a4294298..0492b1ec 100644
--- a/elk/utils/gpu_utils.py
+++ b/elk/utils/gpu_utils.py
@@ -55,7 +55,10 @@ def select_usable_devices(
     # Trivial case: no GPUs requested or available
     num_visible = torch.cuda.device_count()
     if num_gpus == 0 or num_visible == 0:
-        return ["cpu"]
+        if torch.backends.mps.is_available():
+            return ["mps"]
+        else:
+            return ["cpu"]
 
     # Sanity checks
     if num_gpus > num_visible:
diff --git a/elk/utils/hf_utils.py b/elk/utils/hf_utils.py
index 9f429921..670e99ca 100644
--- a/elk/utils/hf_utils.py
+++ b/elk/utils/hf_utils.py
@@ -47,7 +47,7 @@ def instantiate_model(
             kwargs["torch_dtype"] = torch.float16
 
         # CPUs generally don't support anything other than fp32.
-        elif device.type == "cpu":
+        elif device.type in ("cpu", "mps"):
             kwargs["torch_dtype"] = torch.float32
 
         # If the model is fp32 but bf16 is available, convert to bf16.