Fix JAX dispatch for multi-output Composite

pytensor/link/jax/dispatch/scalar.py

@@ -63,11 +63,18 @@ def jax_funcify_Clip(op, **kwargs):
 
 
 @jax_funcify.register(Composite)
-def jax_funcify_Composite(op, vectorize=True, **kwargs):
+def jax_funcify_Composite(op, node, vectorize=True, **kwargs):
     jax_impl = jax_funcify(op.fgraph)
 
-    def composite(*args):
-        return jax_impl(*args)[0]
+    if len(node.outputs) == 1:
+
+        def composite(*args):
+            return jax_impl(*args)[0]
+
+    else:
+
+        def composite(*args):
+            return jax_impl(*args)
 
     return jnp.vectorize(composite)
 

tests/link/jax/test_scalar.py

@@ -63,7 +63,7 @@ def test_identity():
         ),
     ],
 )
-def test_jax_Composite(x, y, x_val, y_val):
+def test_jax_Composite_singe_output(x, y, x_val, y_val):
     x_s = aes.float64("x")
     y_s = aes.float64("y")
 
@@ -80,6 +80,16 @@ def test_jax_Composite(x, y, x_val, y_val):
     _ = compare_jax_and_py(out_fg, test_input_vals)
 
 
+def test_jax_Composite_multi_output():
+    x = vector("x")
+
+    x_s = aes.float64("xs")
+    outs = Elemwise(Composite(inputs=[x_s], outputs=[x_s + 1, x_s - 1]))(x)
+
+    fgraph = FunctionGraph([x], outs)
+    compare_jax_and_py(fgraph, [np.arange(10, dtype=config.floatX)])
+
+
 def test_erf():
     x = scalar("x")
     out = erf(x)