diff --git a/gptqmodel/utils/linalg_warmup.py b/gptqmodel/utils/linalg_warmup.py
index cace622f8..10bfacd00 100644
--- a/gptqmodel/utils/linalg_warmup.py
+++ b/gptqmodel/utils/linalg_warmup.py
@@ -44,7 +44,11 @@ def run_torch_linalg_warmup(device: torch.device) -> None:
     still runs once per physical device so backend-specific handles are initialized where needed.
     """
     with _GLOBAL_WARMUP_LOCK:
-        dtypes = (torch.float32, torch.float64)
+        if device.type == "mps":
+            dtypes = (torch.float32,)  # MPS backend does not implement float64.
+        else:
+            dtypes = (torch.float32, torch.float64)
+
         for dtype in dtypes:
             _run_cholesky_and_eigh(device, dtype)
             _run_svd(device, dtype)