Sequential cpu offload: require accelerate 0.14.0 (#2517)

* Sequential cpu offload: require accelerate 0.14.0.

* Import is_accelerate_version

* Missing copy.

src/diffusers/pipelines/alt_diffusion/pipeline_alt_diffusion.py

@@ -191,10 +191,10 @@ class AltDiffusionPipeline(DiffusionPipeline):
         Note that offloading happens on a submodule basis. Memory savings are higher than with
         `enable_model_cpu_offload`, but performance is lower.
         """
-        if is_accelerate_available():
+        if is_accelerate_available() and is_accelerate_version(">=", "0.14.0"):
             from accelerate import cpu_offload
         else:
-            raise ImportError("Please install accelerate via `pip install accelerate`")
+            raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher")
 
         device = torch.device(f"cuda:{gpu_id}")
 

src/diffusers/pipelines/alt_diffusion/pipeline_alt_diffusion_img2img.py

@@ -213,10 +213,10 @@ class AltDiffusionImg2ImgPipeline(DiffusionPipeline):
         Note that offloading happens on a submodule basis. Memory savings are higher than with
         `enable_model_cpu_offload`, but performance is lower.
         """
-        if is_accelerate_available():
+        if is_accelerate_available() and is_accelerate_version(">=", "0.14.0"):
             from accelerate import cpu_offload
         else:
-            raise ImportError("Please install accelerate via `pip install accelerate`")
+            raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher")
 
         device = torch.device(f"cuda:{gpu_id}")
 

src/diffusers/pipelines/stable_diffusion/pipeline_cycle_diffusion.py

@@ -230,10 +230,10 @@ class CycleDiffusionPipeline(DiffusionPipeline):
         Note that offloading happens on a submodule basis. Memory savings are higher than with
         `enable_model_cpu_offload`, but performance is lower.
         """
-        if is_accelerate_available():
+        if is_accelerate_available() and is_accelerate_version(">=", "0.14.0"):
             from accelerate import cpu_offload
         else:
-            raise ImportError("Please install accelerate via `pip install accelerate`")
+            raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher")
 
         device = torch.device(f"cuda:{gpu_id}")
 

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py

@@ -194,10 +194,10 @@ class StableDiffusionPipeline(DiffusionPipeline):
         Note that offloading happens on a submodule basis. Memory savings are higher than with
         `enable_model_cpu_offload`, but performance is lower.
         """
-        if is_accelerate_available():
+        if is_accelerate_available() and is_accelerate_version(">=", "0.14.0"):
             from accelerate import cpu_offload
         else:
-            raise ImportError("Please install accelerate via `pip install accelerate`")
+            raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher")
 
         device = torch.device(f"cuda:{gpu_id}")
 

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_attend_and_excite.py

@@ -24,7 +24,7 @@ from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
 from ...models import AutoencoderKL, UNet2DConditionModel
 from ...models.cross_attention import CrossAttention
 from ...schedulers import KarrasDiffusionSchedulers
-from ...utils import is_accelerate_available, logging, randn_tensor, replace_example_docstring
+from ...utils import is_accelerate_available, is_accelerate_version, logging, randn_tensor, replace_example_docstring
 from ..pipeline_utils import DiffusionPipeline
 from . import StableDiffusionPipelineOutput
 from .safety_checker import StableDiffusionSafetyChecker
@@ -256,10 +256,10 @@ class StableDiffusionAttendAndExcitePipeline(DiffusionPipeline):
         Note that offloading happens on a submodule basis. Memory savings are higher than with
         `enable_model_cpu_offload`, but performance is lower.
         """
-        if is_accelerate_available():
+        if is_accelerate_available() and is_accelerate_version(">=", "0.14.0"):
             from accelerate import cpu_offload
         else:
-            raise ImportError("Please install accelerate via `pip install accelerate`")
+            raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher")
 
         device = torch.device(f"cuda:{gpu_id}")
 

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py

@@ -218,10 +218,10 @@ class StableDiffusionImg2ImgPipeline(DiffusionPipeline):
         Note that offloading happens on a submodule basis. Memory savings are higher than with
         `enable_model_cpu_offload`, but performance is lower.
         """
-        if is_accelerate_available():
+        if is_accelerate_available() and is_accelerate_version(">=", "0.14.0"):
             from accelerate import cpu_offload
         else:
-            raise ImportError("Please install accelerate via `pip install accelerate`")
+            raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher")
 
         device = torch.device(f"cuda:{gpu_id}")
 

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py

@@ -265,10 +265,10 @@ class StableDiffusionInpaintPipeline(DiffusionPipeline):
         Note that offloading happens on a submodule basis. Memory savings are higher than with
         `enable_model_cpu_offload`, but performance is lower.
         """
-        if is_accelerate_available():
+        if is_accelerate_available() and is_accelerate_version(">=", "0.14.0"):
             from accelerate import cpu_offload
         else:
-            raise ImportError("Please install accelerate via `pip install accelerate`")
+            raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher")
 
         device = torch.device(f"cuda:{gpu_id}")
 

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint_legacy.py

@@ -209,10 +209,10 @@ class StableDiffusionInpaintPipelineLegacy(DiffusionPipeline):
         Note that offloading happens on a submodule basis. Memory savings are higher than with
         `enable_model_cpu_offload`, but performance is lower.
         """
-        if is_accelerate_available():
+        if is_accelerate_available() and is_accelerate_version(">=", "0.14.0"):
             from accelerate import cpu_offload
         else:
-            raise ImportError("Please install accelerate via `pip install accelerate`")
+            raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher")
 
         device = torch.device(f"cuda:{gpu_id}")
 

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_instruct_pix2pix.py

@@ -398,10 +398,10 @@ class StableDiffusionInstructPix2PixPipeline(DiffusionPipeline):
         Note that offloading happens on a submodule basis. Memory savings are higher than with
         `enable_model_cpu_offload`, but performance is lower.
         """
-        if is_accelerate_available():
+        if is_accelerate_available() and is_accelerate_version(">=", "0.14.0"):
             from accelerate import cpu_offload
         else:
-            raise ImportError("Please install accelerate via `pip install accelerate`")
+            raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher")
 
         device = torch.device(f"cuda:{gpu_id}")
 

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_k_diffusion.py

@@ -130,10 +130,10 @@ class StableDiffusionKDiffusionPipeline(DiffusionPipeline):
         Note that offloading happens on a submodule basis. Memory savings are higher than with
         `enable_model_cpu_offload`, but performance is lower.
         """
-        if is_accelerate_available():
+        if is_accelerate_available() and is_accelerate_version(">=", "0.14.0"):
             from accelerate import cpu_offload
         else:
-            raise ImportError("Please install accelerate via `pip install accelerate`")
+            raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher")
 
         device = torch.device(f"cuda:{gpu_id}")
 

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_panorama.py

@@ -19,7 +19,7 @@ from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
 
 from ...models import AutoencoderKL, UNet2DConditionModel
 from ...schedulers import DDIMScheduler, PNDMScheduler
-from ...utils import is_accelerate_available, logging, randn_tensor, replace_example_docstring
+from ...utils import is_accelerate_available, is_accelerate_version, logging, randn_tensor, replace_example_docstring
 from ..pipeline_utils import DiffusionPipeline
 from . import StableDiffusionPipelineOutput
 from .safety_checker import StableDiffusionSafetyChecker
@@ -151,10 +151,10 @@ class StableDiffusionPanoramaPipeline(DiffusionPipeline):
         Note that offloading happens on a submodule basis. Memory savings are higher than with
         `enable_model_cpu_offload`, but performance is lower.
         """
-        if is_accelerate_available():
+        if is_accelerate_available() and is_accelerate_version(">=", "0.14.0"):
             from accelerate import cpu_offload
         else:
-            raise ImportError("Please install accelerate via `pip install accelerate`")
+            raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher")
 
         device = torch.device(f"cuda:{gpu_id}")
 

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_pix2pix_zero.py

@@ -365,10 +365,10 @@ class StableDiffusionPix2PixZeroPipeline(DiffusionPipeline):
         Note that offloading happens on a submodule basis. Memory savings are higher than with
         `enable_model_cpu_offload`, but performance is lower.
         """
-        if is_accelerate_available():
+        if is_accelerate_available() and is_accelerate_version(">=", "0.14.0"):
             from accelerate import cpu_offload
         else:
-            raise ImportError("Please install accelerate via `pip install accelerate`")
+            raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher")
 
         device = torch.device(f"cuda:{gpu_id}")
 

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_sag.py

@@ -22,7 +22,7 @@ from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
 
 from ...models import AutoencoderKL, UNet2DConditionModel
 from ...schedulers import KarrasDiffusionSchedulers
-from ...utils import is_accelerate_available, logging, randn_tensor, replace_example_docstring
+from ...utils import is_accelerate_available, is_accelerate_version, logging, randn_tensor, replace_example_docstring
 from ..pipeline_utils import DiffusionPipeline
 from . import StableDiffusionPipelineOutput
 from .safety_checker import StableDiffusionSafetyChecker
@@ -169,10 +169,10 @@ class StableDiffusionSAGPipeline(DiffusionPipeline):
         Note that offloading happens on a submodule basis. Memory savings are higher than with
         `enable_model_cpu_offload`, but performance is lower.
         """
-        if is_accelerate_available():
+        if is_accelerate_available() and is_accelerate_version(">=", "0.14.0"):
             from accelerate import cpu_offload
         else:
-            raise ImportError("Please install accelerate via `pip install accelerate`")
+            raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher")
 
         device = torch.device(f"cuda:{gpu_id}")
 

tests/pipelines/stable_diffusion_2/test_stable_diffusion_depth.py

@@ -39,7 +39,16 @@ from diffusers import (
     StableDiffusionDepth2ImgPipeline,
     UNet2DConditionModel,
 )
-from diffusers.utils import floats_tensor, is_accelerate_available, load_image, load_numpy, nightly, slow, torch_device
+from diffusers.utils import (
+    floats_tensor,
+    is_accelerate_available,
+    is_accelerate_version,
+    load_image,
+    load_numpy,
+    nightly,
+    slow,
+    torch_device,
+)
 from diffusers.utils.testing_utils import require_torch_gpu, skip_mps
 
 from ...test_pipelines_common import PipelineTesterMixin
@@ -227,8 +236,8 @@ class StableDiffusionDepth2ImgPipelineFastTests(PipelineTesterMixin, unittest.Te
         self.assertLess(max_diff, 1.3e-2, "The outputs of the fp16 and fp32 pipelines are too different.")
 
     @unittest.skipIf(
-        torch_device != "cuda" or not is_accelerate_available(),
+        torch_device != "cuda" or not is_accelerate_available() or is_accelerate_version("<", "0.14.0"),
-        reason="CPU offload is only available with CUDA and `accelerate` installed",
+        reason="CPU offload is only available with CUDA and `accelerate v0.14.0` or higher",
     )
     def test_cpu_offload_forward_pass(self):
         components = self.get_dummy_components()

tests/test_pipelines_common.py

@@ -13,7 +13,7 @@ import torch
 import diffusers
 from diffusers import DiffusionPipeline
 from diffusers.utils import logging
-from diffusers.utils.import_utils import is_accelerate_available, is_xformers_available
+from diffusers.utils.import_utils import is_accelerate_available, is_accelerate_version, is_xformers_available
 from diffusers.utils.testing_utils import require_torch, torch_device
 
 
@@ -417,8 +417,8 @@ class PipelineTesterMixin:
         assert_mean_pixel_difference(output_with_slicing[0], output_without_slicing[0])
 
     @unittest.skipIf(
-        torch_device != "cuda" or not is_accelerate_available(),
+        torch_device != "cuda" or not is_accelerate_available() or is_accelerate_version("<", "0.14.0"),
-        reason="CPU offload is only available with CUDA and `accelerate` installed",
+        reason="CPU offload is only available with CUDA and `accelerate v0.14.0` or higher",
     )
     def test_cpu_offload_forward_pass(self):
         if not self.test_cpu_offload: