use low_cpu_mem_usage when using deepspeed

src/axolotl/utils/models.py

@@ -355,8 +355,9 @@ def load_model(
     #         else:
     #             model_kwargs["device_map"] = "cuda:" + str(torch.cuda.current_device())
 
-    if is_deepspeed_zero3_enabled():
+    if is_deepspeed_zero3_enabled() or cfg.deepspeed:
         del model_kwargs["device_map"]
+        model_kwargs["low_cpu_mem_usage"] = True
 
     if cfg.model_revision:
         model_kwargs["revision"] = cfg.model_revision