[Runtime][AOTExecutor] set_input does not show error if input name does not exist in the model

[mehrdadh]

If we pass input dictionary with incorrect input name, aot executor would execute set_input function without actually setting the correct input or showing error that input name doesn't exist in the model. This will result in silently outputting the incorrect output from the model. I think it should show an error when the input name doesn't exist.

I found this error when I working on a model which had input name of input:0 which is not acceptable in Relay. However, when I tested it I realized it's not just this name, if you pass any name for the input the AOTExecutor would accept it.

cc @areusch @alanmacd

[alanmacd]

@mehrdadh which flavor of AoT executor is this, also can you post some example code that shows the issue?

[mehrdadh]

@alanmacd I have tested this on AOT with C++ and it doesn't show any error. It might also be a problem with CRT.
To reproduce the error, you can use this test:

tvm/tests/python/relay/aot/test_cpp_aot.py

Line 54 in 0b034d7

def test_conv2d(enable_usmp, target_kind):

If you apply this modification, then the test fails because the output is incorrect but you don't see any error regarding passing data to the wrong input name.

diff --git a/tests/python/relay/aot/test_cpp_aot.py b/tests/python/relay/aot/test_cpp_aot.py
index b67bc90d3..44d336d88 100644
--- a/tests/python/relay/aot/test_cpp_aot.py
+++ b/tests/python/relay/aot/test_cpp_aot.py
@@ -90,7 +90,7 @@ def test_conv2d(enable_usmp, target_kind):
     weight_data = np.random.randint(1, 255, shape_dict["weight"]).astype(type_dict["weight"])
     input_data = np.ones(shape_dict["data"]).astype(type_dict["data"])
     params = {"weight": weight_data}
-    inputs = {"data": input_data}
+    inputs = {"data1": input_data}
     ref_outputs = generate_ref_data(ir_mod, inputs, params)
 
     with tvm.transform.PassContext(