diff --git a/src/mobius/__main__.py b/src/mobius/__main__.py
index b8dc10c16..706d448b5 100644
--- a/src/mobius/__main__.py
+++ b/src/mobius/__main__.py
@@ -214,6 +214,14 @@ def _save_package(
         _apply_optimize(model, optimize)
 
     max_shard_size_bytes = _parse_size(args.max_shard_size) if args.max_shard_size else None
+
+    if args.external_data == "safetensors" and getattr(args, "ep", "default") == "cuda":
+        logging.getLogger(__name__).warning(
+            "Safetensors external data does not guarantee 256-byte offset "
+            "alignment, which can cause CUBLAS misaligned address errors on "
+            "CUDA. Consider using --external-data onnx for CUDA builds."
+        )
+
     pkg.save(
         output_dir,
         external_data=args.external_data,
diff --git a/src/mobius/_model_package.py b/src/mobius/_model_package.py
index 85274eb97..7c099296f 100644
--- a/src/mobius/_model_package.py
+++ b/src/mobius/_model_package.py
@@ -78,6 +78,14 @@ def save(
             external_data: External data format. ``"onnx"`` (default) saves
                 weights to ``model.onnx.data``. ``"safetensors"`` saves
                 weights in safetensors format.
+
+                .. warning::
+                    The safetensors format does not guarantee 256-byte offset
+                    alignment for tensor data within the file.  This can cause
+                    ``CUBLAS_STATUS_INVALID_VALUE`` ("misaligned address")
+                    errors on some CUDA/cuBLAS versions when loading weights
+                    via memory-mapped I/O.  Use ``"onnx"`` (the default) for
+                    models targeting CUDA execution.
             max_shard_size_bytes: Maximum shard size in bytes for safetensors
                 format.  Only used when *external_data* is ``"safetensors"``.
             components: Optional predicate ``(name) -> bool`` that selects