diff --git a/jax/lib/xla_bridge.py b/jax/lib/xla_bridge.py
index 8ba20225f16d..ead4e7be0bb6 100644
--- a/jax/lib/xla_bridge.py
+++ b/jax/lib/xla_bridge.py
@@ -63,7 +63,10 @@
     'jax_disable_most_optimizations', False,
     'Try not to do much optimization work. This can be useful if the cost of '
     'optimization is greater than that of running a less-optimized program.')
-
+flags.DEFINE_string(
+    'jax_cpu_backend_variant', 'tfrt',
+    'jax_cpu_backend_variant selects cpu backend variant: stream_executor or '
+    'tfrt')
 
 def get_compile_options(
     num_replicas: int,
@@ -148,28 +151,39 @@ def register_backend_factory(name, factory, *, priority=0):
 
 
 if jax.lib._xla_extension_version >= 23:
-   register_backend_factory('interpreter', xla_client.make_interpreter_client,
-                            priority=-100)
-   register_backend_factory('cpu', xla_client.make_cpu_client,
-                            priority=0)
-   register_backend_factory('tpu_driver', _make_tpu_driver_client,
-                            priority=100)
-   register_backend_factory('gpu', xla_client.make_gpu_client,
-                            priority=200)
-   register_backend_factory('tpu', xla_client.make_tpu_client,
-                            priority=300)
+  register_backend_factory('interpreter', xla_client.make_interpreter_client,
+                           priority=-100)
+  if jax.lib._xla_extension_version >= 24:
+    if FLAGS.jax_cpu_backend_variant == 'stream_executor':
+      register_backend_factory('cpu',
+                               partial(xla_client.make_cpu_client, use_tfrt=False),
+                               priority=0)
+    else:
+      assert FLAGS.jax_cpu_backend_variant == 'tfrt'
+      register_backend_factory('cpu',
+                               partial(xla_client.make_cpu_client, use_tfrt=True),
+                               priority=0)
+  else:
+    register_backend_factory('cpu',
+                              partial(xla_client.make_cpu_client, use_tfrt=False),
+                              priority=0)
+  register_backend_factory('tpu_driver', _make_tpu_driver_client,
+                           priority=100)
+  register_backend_factory('gpu', xla_client.make_gpu_client,
+                           priority=200)
+  register_backend_factory('tpu', xla_client.make_tpu_client,
+                           priority=300)
 else:
-   register_backend_factory('interpreter',
-                            xla_client._interpreter_backend_factory,
-                            priority=-100)
-   register_backend_factory('cpu', xla_client._cpu_backend_factory, priority=0)
-   register_backend_factory('tpu_driver', _make_tpu_driver_client,
-                            priority=100)
-   register_backend_factory('gpu', xla_client._gpu_backend_factory,
-                            priority=200)
-   register_backend_factory('tpu', xla_client._tpu_backend_factory,
-                            priority=300)
-
+  register_backend_factory('interpreter',
+                           xla_client._interpreter_backend_factory,
+                           priority=-100)
+  register_backend_factory('cpu', xla_client._cpu_backend_factory, priority=0)
+  register_backend_factory('tpu_driver', _make_tpu_driver_client,
+                           priority=100)
+  register_backend_factory('gpu', xla_client._gpu_backend_factory,
+                           priority=200)
+  register_backend_factory('tpu', xla_client._tpu_backend_factory,
+                           priority=300)
 
 _default_backend = None
 _backends = None