HPUAccelerator: remove support in set_visible_devices_envs

The way deepspeed sets it is not correct with all HPU instances
and may lead to incorrect behavior.

accelerator/hpu_accelerator.py

@@ -298,7 +298,10 @@ def export_envs(self):
         return []
 
     def visible_devices_envs(self):
-        return ['HABANA_VISIBLE_MODULES']
+        # Current way deepspeed set this env var is not applicable with all HPU instances
+        # User has to follow instructions in:
+        # https://docs.habana.ai/en/latest/PyTorch/Reference/PT_Multiple_Tenants_on_HPU/Multiple_Workloads_Single_Docker.html
+        return []  #['HABANA_VISIBLE_MODULES']
 
     def set_visible_devices_envs(self, current_env, local_accelerator_ids):
         for env in self.visible_devices_envs():