diff --git a/modules/model_quant_sdnq.py b/modules/model_quant_sdnq.py
index a40faee58..f7c867e90 100644
--- a/modules/model_quant_sdnq.py
+++ b/modules/model_quant_sdnq.py
@@ -6,6 +6,7 @@ from enum import Enum
 import os
 import sys
 import torch
+import diffusers
 from diffusers.quantizers.base import DiffusersQuantizer
 from diffusers.quantizers.quantization_config import QuantizationConfigMixin
 from diffusers.utils import get_module_from_name
@@ -681,6 +682,8 @@ class SDNQQuantizer(DiffusersQuantizer):
         state_dict: Dict[str, Any],
         **kwargs,
     ):
+        if hasattr(diffusers, model.__class__.__name__):
+            param_value.data = param_value.clone() # safetensors is unable to release the cpu memory without this
         if param_name.endswith(".weight"):
             split_param_name = param_name.split(".")
             if param_name not in self.modules_to_not_convert and not any(param in split_param_name for param in self.modules_to_not_convert):
diff --git a/modules/teacache/teacache_hidream.py b/modules/teacache/teacache_hidream.py
index b3bd84bb8..8f7f4b859 100644
--- a/modules/teacache/teacache_hidream.py
+++ b/modules/teacache/teacache_hidream.py
@@ -1,7 +1,4 @@
 from typing import Any, Dict, List, Optional, Tuple
-from transformers import PreTrainedTokenizerFast, LlamaForCausalLM
-from diffusers import HiDreamImagePipeline
-from diffusers.models import HiDreamImageTransformer2DModel
 from diffusers.models.modeling_outputs import Transformer2DModelOutput
 from diffusers.utils import logging, deprecate, USE_PEFT_BACKEND, logging, scale_lora_layers, unscale_lora_layers