[JAX] Keep CPU host callbacks alive via IFRT, rather than by attachin…

…g them to the Python object.

We need to keep callback objects alive as long as any running executables are alive. It is possible to discard the Python data structures for an executable before the runtime has finished running that executable, which can lead to a use after free. Instead, make the runtime keep host callbacks alive.

PiperOrigin-RevId: 571141106

third_party/xla/xla/python/pjrt_ifrt/pjrt_executable.cc

@@ -440,11 +440,6 @@ PjRtLoadedExecutable::CreateInternal(
       host_send_and_recv_callbacks.push_back(host_send_and_recv_callback);
     }
   }
-  if (!loaded_host_callbacks.empty() &&
-      !client->pjrt_client()->SupportsSendRecvCallbacks()) {
-    return InternalError("Host callback not supported for runtime type: %s",
-                         client->runtime_type());
-  }
 
   return std::unique_ptr<LoadedExecutable>(new PjRtLoadedExecutable(
       client, std::move(pjrt_loaded_executable), std::move(devices),
@@ -473,11 +468,7 @@ PjRtLoadedExecutable::PjRtLoadedExecutable(
       output_shapes_(std::move(output_shapes)),
       output_shardings_(std::move(output_shardings)) {}
 
-PjRtLoadedExecutable::~PjRtLoadedExecutable() {
-  // Reset the PjRt executable before host callbacks.
-  pjrt_loaded_executable_ = nullptr;
-  all_loaded_host_callbacks_->clear();
-}
+PjRtLoadedExecutable::~PjRtLoadedExecutable() = default;
 
 StatusOr<PjRtLoadedExecutable::ExecuteResult> PjRtLoadedExecutable::Execute(
     absl::Span<tsl::RCReference<Array>> args, const ExecuteOptions& options,

third_party/xla/xla/python/xla_client.py

@@ -45,7 +45,7 @@
 
 # Just an internal arbitrary increasing number to help with backward-compatible
 # changes. In JAX, reference this via jax._src.lib.xla_extension_version.
-_version = 201
+_version = 202
 
 # Version number for MLIR:Python components.
 mlir_api_version = 54