deepspeed sharding (#840)

* debugging

* deepspeed: pin ze memory

---------

Co-authored-by: bghira <bghira@users.github.com>

train.py

@@ -873,7 +873,7 @@ def main():
         ]["offload_param"]
         accelerator.state.deepspeed_plugin.deepspeed_config["zero_optimization"][
             "offload_param"
-        ]["pin_memory"] = False
+        ]["pin_memory"] = True
         if offload_param["device"] == "nvme":
             if offload_param["nvme_path"] == "none":
                 if args.offload_param_path is None: