Add smoke test

build/packaging/smoke_test.py

@@ -15,6 +15,14 @@
 # will fail and the process will exit.
 from executorch.extension.pybindings import portable_lib  # usort: skip
 
+# Import custom ops. This requires portable_lib to be loaded first.
+from executorch.extension.llm.custom_ops import (
+    sdpa_with_kv_cache,
+)  # usort: skip # noqa: F401, F403
+
+# Import quantized ops. This requires portable_lib to be loaded first.
+from executorch.kernels import quantized  # usort: skip # noqa: F401, F403
+
 # Import this after importing the ExecuTorch pybindings. If the pybindings
 # links against a different torch.so than this uses, there will be a set of
 # symbol comflicts; the process will either exit now, or there will be issues
@@ -75,6 +83,15 @@ def main():
     assert len(ops) > 0, "Empty operator list"
     print(f"Found {len(ops)} operators; first element '{ops[0]}'")
 
+    # Make sure custom ops are registered.
+    assert (
+        "llama::sdpa_with_kv_cache" in ops
+    ), f"sdpa_with_kv_cache not registered, Got ops: {ops}"
+
+    # Make sure quantized ops are registered.
+    assert (
+        "quantized_decomposed::add.out" in ops
+    ), f"quantized_decomposed::add.out not registered, Got ops: {ops}"
     # Export LinearModel to .pte data.
     pte_data: bytes = export_linear_model()
 

kernels/quantized/CMakeLists.txt

@@ -127,7 +127,7 @@ if(NOT CMAKE_GENERATOR STREQUAL "Xcode"
       # values, run `otool -l libquantized_ops_lib.dylib`.
       if(APPLE)
         set_target_properties(
-          quantized_ops_lib
+          quantized_ops_aot_lib
           PROPERTIES # Assume this library will be installed in
                      # <site-packages>/executorch/kernels/quantized/, and the
                      # _portable_lib.so is installed in