diff --git a/ethosu/vela/graph_optimiser_util.py b/ethosu/vela/graph_optimiser_util.py
index 2822feb..24a5583 100644
--- a/ethosu/vela/graph_optimiser_util.py
+++ b/ethosu/vela/graph_optimiser_util.py
@@ -417,7 +417,8 @@
 
 def convert_to_lut(op, lut_values, lut_name):
     # Rewrite the operation by Add with scalar 0 + LUT activation
-    ifm = op.inputs[0]
+    ifm = op.ifm
+    ofm = op.ofm
     if ifm is None:
         return op
     assert ifm.dtype.size_in_bytes() == 1
@@ -429,7 +430,7 @@
     quantization = QuantizationParameters(0.0, 255.0)
     quantization.scale_f32 = ifm.quantization.scale_f32
     quantization.zero_point = 0
-    tens = create_const_tensor(op.inputs[0].name + "_scalar0", [], ifm.dtype, [0], quantization=quantization)
+    tens = create_const_tensor(ifm.name + "_scalar0", [], ifm.dtype, [0], quantization=quantization)
     op.add_input_tensor(tens)
     op.ifm_shapes.append(Shape4D(tens.shape))  # TODO no shape?
 
@@ -437,7 +438,13 @@
     # so even if the OFM has a different scale than the IFM, the generated OFM scale instructions
     # should be the same as the IFM
     op.forced_output_quantization = ifm.quantization
-    lut_tensor = lut.create_lut_tensor(op.name + "_values", lut_values, DataType.int8)
+
+    # the lut tensor datatype needs to match both; the ofm datatype, because these are the values output; and the
+    # datatype used to generate the lut values (which is probably the ifm datatype), because we want to avoid any
+    # potential overflow errors in create_lut_tensor() caused by converting Python int (which could represent a uint)
+    # to NumPy int. this can be guaranteed by checking that the ifm and ofm datatypes are the same
+    assert ifm.dtype == ofm.dtype
+    lut_tensor = lut.create_lut_tensor(op.name + "_values", lut_values, ofm.dtype)
     op.set_activation_lut(lut_tensor)
     op.set_ifm_ofm_shapes()
     DebugDatabase.add_optimised(op, op)
diff --git a/ethosu/vela/softmax.py b/ethosu/vela/softmax.py
index 575e1e6..5a06c1b 100644
--- a/ethosu/vela/softmax.py
+++ b/ethosu/vela/softmax.py
@@ -270,7 +270,7 @@
             ifm2_shape=ifm_max_shape,
         )
         sub_op.set_activation_lut(
-            create_const_tensor(f"{sub_op.name}_exp_lut", [1, 1, 1, 256], DataType.int32, exp_lut, TensorPurpose.LUT)
+            create_const_tensor(f"{sub_op.name}_exp_lut", [1, 1, 1, 256], DataType.uint32, exp_lut, TensorPurpose.LUT)
         )
         ifm_exp = add_op_get_ofm(sub_op)
         # Note: activation.min/max are non-quantized values
@@ -505,8 +505,10 @@
             f"{name}_const", [1, 1, 1, 1], DataType.int32, [32767], quantization=no_scale_quant
         )
         add_op = create_add(name, mul2_ofm, const_add, mul2_ofm.quantization.clone(), dtype=DataType.int16)
+        # lut activation values are int32 type however they are defined as Python ints. If these are converted to
+        # numpy.int32 it could result in an overflow error. Therefore, they are forced to uint32 to avoid this
         add_op.set_activation_lut(
-            create_const_tensor(f"{name}_exp_lut", [1, 1, 1, 512], DataType.int32, self.EXP_LUT, TensorPurpose.LUT)
+            create_const_tensor(f"{name}_exp_lut", [1, 1, 1, 512], DataType.uint32, self.EXP_LUT, TensorPurpose.LUT)
         )
         ifm_exp = add_op_get_ofm(add_op)
 
@@ -550,11 +552,13 @@
             f"{name}_const", [1, 1, 1, 1], DataType.int32, [32768], quantization=no_scale_quant
         )
         sub11_op = create_sub(name, shifted_sum_minus_one_16, sub11_const, no_scale_quant, dtype=DataType.int16)
+        # lut activation values are int32 type however they are defined as Python ints. If these are converted to
+        # numpy.int32 it could result in an overflow error. Therefore, they are forced to uint32 to avoid this
         sub11_op.set_activation_lut(
             create_const_tensor(
                 f"{name}_one_over_one_plus_x_lut",
                 [1, 1, 1, 512],
-                DataType.int32,
+                DataType.uint32,
                 self.ONE_OVER_ONE_PLUS_X_LUT,
                 TensorPurpose.LUT,
             )
diff --git a/ethosu/vela/test/test_lut.py b/ethosu/vela/test/test_lut.py
index 712be7a..58e72bb 100644
--- a/ethosu/vela/test/test_lut.py
+++ b/ethosu/vela/test/test_lut.py
@@ -35,7 +35,7 @@
 def set_256_lut(op, key, arch):
     random.seed(key)
     values = random.choices(range(256), k=256)
-    lut_tensor = create_const_tensor(op.name + "_lut", [1, 1, 1, 256], DataType.int8, values, TensorPurpose.LUT)
+    lut_tensor = create_const_tensor(op.name + "_lut", [1, 1, 1, 256], DataType.uint8, values, TensorPurpose.LUT)
     scratch_lut_tensor = lut_tensor.clone_into_fast_storage(arch)
     op.set_activation_lut(scratch_lut_tensor)