Merge pull request #1158 from fengyuentau/matmul_bias

* make model and data for biased matmul using onnxruntime for inference

* add note about onnxscript producing incorrect results in eager mode

* update script to put random numpy arrays outside of function

Author: fengyuentau

testdata/dnn/onnx/data/input_biased_matmul.npy

@@ -6,6 +6,13 @@
 from onnxscript import opset11
 from onnxscript import opset13
 
+###############
+### CAUTION!!!
+### Be sure to put constant numpy arrays out of @ost.script() decorated fucntion.
+### Otherwise random values change each time eager mode is enter.
+### See discussions in https://github.com/microsoft/onnxscript/issues/1313
+###############
+
 np.random.seed(0)
 
 def make_model_and_data(model, *args, **kwargs):
@@ -339,3 +346,22 @@ def layer_norm_no_fusion(x: ost.FLOAT[n, c, h, w]) -> ost.FLOAT[n, c, h, w]:
 
     return add
 make_model_and_data(layer_norm_no_fusion, np.random.rand(n, c, h, w).astype(np.float32))
+
+
+''' Subgraph: [Input] -> MatMul<B> -> Add<A> -> [Output]
+'''
+
+b = 2
+m = 32
+n = 64
+k = 16
+weight_data = np.random.rand(k, n).astype(np.float32)
+bias_data = np.random.rand(n).astype(np.float32)
+
+@ost.script()
+def biased_matmul(x: ost.FLOAT[b, m, k]) -> ost.FLOAT[b, m, n]:
+    weight = op.Constant(value=onnx.helper.make_tensor("", onnx.TensorProto.FLOAT, [k, n], weight_data))
+    matmul = op.MatMul(x, weight)
+    bias = op.Constant(value=onnx.helper.make_tensor("", onnx.TensorProto.FLOAT, [n], bias_data))
+    return op.Add(bias, matmul)
+make_model_and_data(biased_matmul, np.random.rand(b, m, k).astype(np.float32), use_ort=True, ort_input_keys=["x"])