diff --git a/aiu_fms_testing_utils/scripts/inference.py b/aiu_fms_testing_utils/scripts/inference.py
index 45d52cc3..502cc07f 100644
--- a/aiu_fms_testing_utils/scripts/inference.py
+++ b/aiu_fms_testing_utils/scripts/inference.py
@@ -138,6 +138,13 @@
     default="inductor",
     choices=["inductor", "eager", "aot_eager"],
 )
+parser.add_argument(
+    "--compile_backend_aiu",
+    type=str,
+    help="Backend for AIU compilation",
+    default="sendnn",
+    choices=["sendnn", "sendnn_compile_only"],
+)
 parser.add_argument(
     "--compile_dynamic",
     action="store_true",
@@ -387,7 +394,10 @@
             print("must set AIU_WORLD_RANK_0")
             exit()
         os.environ.setdefault("FLEX_COMPUTE", "SENTIENT")
-        os.environ.setdefault("FLEX_DEVICE", "PF")
+        if args.compile_backend_aiu == "sendnn_compile_only":
+            os.environ.setdefault("FLEX_DEVICE", "COMPILE")
+        else:
+            os.environ.setdefault("FLEX_DEVICE", "PF")
 
     device = torch.device("cpu")
 else:
@@ -597,7 +607,8 @@ def select_int8_module(
     fx_config.backed_size_oblivious = "paged" in attn_name
     if is_aiu_backend:
         model.compile(
-            backend="sendnn", options={"sendnn.dynamic": args.compile_dynamic_sendnn}
+            backend=args.compile_backend_aiu,
+            options={"sendnn.dynamic": args.compile_dynamic_sendnn},
         )
     else:
         # compiling can make first inference pass slow