Update SoftmaxSampling to support dynamic dtypes (#304)

The output types match the input dtypes passed. And the output schema
is computed from the input types passed.

merlin/systems/dag/ops/softmax_sampling.py

@@ -101,8 +101,14 @@ def compute_output_schema(
         """Describe the operator's outputs"""
         return Schema(
             [
-                ColumnSchema("ordered_ids", dtype=np.int32, dims=(None, 1)),
-                ColumnSchema("ordered_scores", dtype=np.float32, dims=(None, 1)),
+                ColumnSchema(
+                    "ordered_ids", dtype=input_schema[self._input_col_name].dtype, dims=(None, 1)
+                ),
+                ColumnSchema(
+                    "ordered_scores",
+                    dtype=input_schema[self._relevance_col_name].dtype,
+                    dims=(None, 1),
+                ),
             ]
         )
 

tests/unit/systems/dag/ops/test_ops.py

@@ -45,8 +45,47 @@ def test_compute_dims(column_schema, expected_dims):
     assert compute_dims(column_schema) == expected_dims
 
 
+@pytest.mark.parametrize(
+    ["id_dtype", "score_dtype"],
+    [
+        ("int32", "float32"),
+        ("int64", "float64"),
+    ],
+)
+def test_softmax_sampling(id_dtype, score_dtype):
+    input_schema = Schema(
+        [
+            ColumnSchema("movie_ids", dtype=id_dtype, dims=(None, 100)),
+            ColumnSchema("relevance_score", dtype=score_dtype, dims=(None, 100)),
+        ]
+    )
+
+    movie_ids = np.array(random.sample(range(10000), 100), dtype=id_dtype)
+    relevance_score = np.random.random(100).astype(score_dtype)
+
+    combined_features = {
+        "movie_ids": np.expand_dims(movie_ids, axis=0),
+        "relevance_score": np.expand_dims(relevance_score, axis=0),
+    }
+
+    input_table = TensorTable(combined_features)
+
+    ordering = ["movie_ids"] >> SoftmaxSampling(
+        relevance_col="relevance_score", topk=10, temperature=20.0
+    )
+
+    ensemble = Ensemble(ordering, input_schema)
+    output_table = ensemble.transform(input_table)
+
+    assert output_table["ordered_ids"].dtype.to_numpy == input_schema["movie_ids"].dtype.to_numpy
+    assert (
+        output_table["ordered_scores"].dtype.to_numpy
+        == input_schema["relevance_score"].dtype.to_numpy
+    )
+
+
 @pytest.mark.skipif(not TRITON_SERVER_PATH, reason="triton server not found")
-def test_softmax_sampling(tmpdir):
+def test_softmax_sampling_with_triton(tmpdir):
     request_schema = Schema(
         [
             ColumnSchema("movie_ids", dtype=np.int32, dims=(None, 100)),