diff --git a/src/diffusers/pipeline_utils.py b/src/diffusers/pipeline_utils.py
index 35996d65072c..81118967aade 100644
--- a/src/diffusers/pipeline_utils.py
+++ b/src/diffusers/pipeline_utils.py
@@ -32,7 +32,19 @@
 from .configuration_utils import ConfigMixin
 from .dynamic_modules_utils import get_class_from_dynamic_module
 from .schedulers.scheduling_utils import SCHEDULER_CONFIG_NAME
-from .utils import CONFIG_NAME, DIFFUSERS_CACHE, ONNX_WEIGHTS_NAME, WEIGHTS_NAME, BaseOutput, logging
+from .utils import (
+    CONFIG_NAME,
+    DIFFUSERS_CACHE,
+    ONNX_WEIGHTS_NAME,
+    WEIGHTS_NAME,
+    BaseOutput,
+    is_transformers_available,
+    logging,
+)
+
+
+if is_transformers_available():
+    from transformers import PreTrainedModel
 
 
 INDEX_FILE = "diffusion_pytorch_model.bin"
@@ -338,6 +350,7 @@ def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.P
         custom_pipeline = kwargs.pop("custom_pipeline", None)
         provider = kwargs.pop("provider", None)
         sess_options = kwargs.pop("sess_options", None)
+        device_map = kwargs.pop("device_map", None)
 
         # 1. Download the checkpoints and configs
         # use snapshot download here to get it working from from_pretrained
@@ -463,6 +476,13 @@ def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.P
                     loading_kwargs["provider"] = provider
                     loading_kwargs["sess_options"] = sess_options
 
+                if (
+                    issubclass(class_obj, diffusers.ModelMixin)
+                    or is_transformers_available()
+                    and issubclass(class_obj, PreTrainedModel)
+                ):
+                    loading_kwargs["device_map"] = device_map
+
                 # check if the module is in a subdirectory
                 if os.path.isdir(os.path.join(cached_folder, name)):
                     loaded_sub_model = load_method(os.path.join(cached_folder, name), **loading_kwargs)
diff --git a/src/diffusers/pipelines/stable_diffusion/safety_checker.py b/src/diffusers/pipelines/stable_diffusion/safety_checker.py
index 773a7d4b2107..3984171f57db 100644
--- a/src/diffusers/pipelines/stable_diffusion/safety_checker.py
+++ b/src/diffusers/pipelines/stable_diffusion/safety_checker.py
@@ -19,6 +19,8 @@ def cosine_distance(image_embeds, text_embeds):
 class StableDiffusionSafetyChecker(PreTrainedModel):
     config_class = CLIPConfig
 
+    _no_split_modules = ["CLIPEncoderLayer"]
+
     def __init__(self, config: CLIPConfig):
         super().__init__(config)
 
@@ -28,8 +30,8 @@ def __init__(self, config: CLIPConfig):
         self.concept_embeds = nn.Parameter(torch.ones(17, config.projection_dim), requires_grad=False)
         self.special_care_embeds = nn.Parameter(torch.ones(3, config.projection_dim), requires_grad=False)
 
-        self.register_buffer("concept_embeds_weights", torch.ones(17))
-        self.register_buffer("special_care_embeds_weights", torch.ones(3))
+        self.concept_embeds_weights = nn.Parameter(torch.ones(17), requires_grad=False)
+        self.special_care_embeds_weights = nn.Parameter(torch.ones(3), requires_grad=False)
 
     @torch.no_grad()
     def forward(self, clip_input, images):
diff --git a/tests/test_pipelines.py b/tests/test_pipelines.py
index 4bf1e5e47c8e..30beb033fca7 100644
--- a/tests/test_pipelines.py
+++ b/tests/test_pipelines.py
@@ -17,12 +17,15 @@
 import os
 import random
 import tempfile
+import tracemalloc
 import unittest
 
 import numpy as np
 import torch
 
+import accelerate
 import PIL
+import transformers
 from diffusers import (
     AutoencoderKL,
     DDIMPipeline,
@@ -50,6 +53,7 @@
 from diffusers.schedulers.scheduling_utils import SCHEDULER_CONFIG_NAME
 from diffusers.utils import CONFIG_NAME, WEIGHTS_NAME, floats_tensor, load_image, slow, torch_device
 from diffusers.utils.testing_utils import get_tests_dir
+from packaging import version
 from PIL import Image
 from transformers import CLIPFeatureExtractor, CLIPModel, CLIPTextConfig, CLIPTextModel, CLIPTokenizer
 
@@ -2034,3 +2038,53 @@ def test_callback_fn(step: int, timestep: int, latents: np.ndarray) -> None:
         pipe(prompt=prompt, num_inference_steps=5, guidance_scale=7.5, callback=test_callback_fn, callback_steps=1)
         assert test_callback_fn.has_been_called
         assert number_of_steps == 6
+
+    @slow
+    @unittest.skipIf(torch_device == "cpu", "Stable diffusion is supposed to run on GPU")
+    def test_stable_diffusion_accelerate_load_works(self):
+        if version.parse(version.parse(transformers.__version__).base_version) < version.parse("4.23"):
+            return
+
+        if version.parse(version.parse(accelerate.__version__).base_version) < version.parse("0.14"):
+            return
+
+        model_id = "CompVis/stable-diffusion-v1-4"
+        _ = StableDiffusionPipeline.from_pretrained(
+            model_id, revision="fp16", torch_dtype=torch.float16, use_auth_token=True, device_map="auto"
+        ).to(torch_device)
+
+    @slow
+    @unittest.skipIf(torch_device == "cpu", "This test is supposed to run on GPU")
+    def test_stable_diffusion_accelerate_load_reduces_memory_footprint(self):
+        if version.parse(version.parse(transformers.__version__).base_version) < version.parse("4.23"):
+            return
+
+        if version.parse(version.parse(accelerate.__version__).base_version) < version.parse("0.14"):
+            return
+
+        pipeline_id = "CompVis/stable-diffusion-v1-4"
+
+        torch.cuda.empty_cache()
+        gc.collect()
+
+        tracemalloc.start()
+        pipeline_normal_load = StableDiffusionPipeline.from_pretrained(
+            pipeline_id, revision="fp16", torch_dtype=torch.float16, use_auth_token=True
+        )
+        pipeline_normal_load.to(torch_device)
+        _, peak_normal = tracemalloc.get_traced_memory()
+        tracemalloc.stop()
+
+        del pipeline_normal_load
+        torch.cuda.empty_cache()
+        gc.collect()
+
+        tracemalloc.start()
+        _ = StableDiffusionPipeline.from_pretrained(
+            pipeline_id, revision="fp16", torch_dtype=torch.float16, use_auth_token=True, device_map="auto"
+        )
+        _, peak_accelerate = tracemalloc.get_traced_memory()
+
+        tracemalloc.stop()
+
+        assert peak_accelerate < peak_normal