diff --git a/einx/backend/tracer.py b/einx/backend/tracer.py
index 66bbf36..d7fe70a 100644
--- a/einx/backend/tracer.py
+++ b/einx/backend/tracer.py
@@ -352,10 +352,10 @@ def reduce(tensor, axis, *, op=None, **kwargs):
     else:
         for a in reversed(sorted(axes)):
             del shape[a]
-    return Op(op, args=[tensor], kwargs=kwargs | {"axis": axis}, output_shapes=np.asarray(shape)).output_tracers
+    return Op(op, args=[tensor], kwargs={**kwargs, **{"axis": axis}}, output_shapes=np.asarray(shape)).output_tracers
 
 def map(tensor, axis, op, *args, **kwargs):
-    return Op(op, args=[tensor], kwargs=kwargs | {"axis": axis}, output_shapes=np.asarray(tensor.shape)).output_tracers
+    return Op(op, args=[tensor], kwargs={**kwargs, **{"axis": axis}}, output_shapes=np.asarray(tensor.shape)).output_tracers
 
 def index(tensor, coordinates, update=None, op=None):
     return Op(op, args=[tensor, coordinates, update], output_shapes=np.asarray(coordinates[0].shape)).output_tracers
diff --git a/einx/lru_cache.py b/einx/lru_cache.py
index 74b3f24..a460524 100644
--- a/einx/lru_cache.py
+++ b/einx/lru_cache.py
@@ -34,7 +34,7 @@ def lru_cache(func=None, trace=None):
         if max_cache_size == 0:
             inner = func
         elif max_cache_size < 0:
-            inner = freeze(functools.cache(func)) # No cache limit
+            inner = freeze(functools.lru_cache(maxsize=None)(func)) # No cache limit
         else:
             inner = freeze(functools.lru_cache(maxsize=max_cache_size)(func))
     else:
diff --git a/einx/param.py b/einx/param.py
index 1b01565..415c397 100644
--- a/einx/param.py
+++ b/einx/param.py
@@ -30,7 +30,7 @@ def instantiate(x, shape, backend, **kwargs):
         raise TypeError("instantiate cannot be called on None")
     if backend == einx.backend.tracer:
         if is_tensor_factory(x):
-            return einx.backend.tracer.Op(instantiate, [x], {"shape": shape} | kwargs, output_shapes=np.asarray(shape), pass_backend=True).output_tracers
+            return einx.backend.tracer.Op(instantiate, [x], {**{"shape": shape}, **kwargs}, output_shapes=np.asarray(shape), pass_backend=True).output_tracers
         else:
             return einx.backend.tracer.Op("to_tensor", [x], output_shapes=np.asarray(shape)).output_tracers
     else: