make examples in docs device agnostic

.ai/skills/model-integration/SKILL.md

@@ -41,9 +41,22 @@ A pipeline in Diffusers (be it standard or modular) will have multiple component
   - [ ] Add a LoRA mixin if applicable
   - [ ] Register in the relevant `__init__.py` files (lazy imports)
   - [ ] Pipeline-level tests (see **Testing**)
-- [ ] **Docs** — see **File structure**
+- [ ] **Docs** — see **File structure** and **Docs device examples**
 - [ ] **Style** — `make style` and `make quality`
 
+## Docs device examples
+
+When adding or updating examples in `docs/source/en/`, avoid hardcoding CUDA as the default execution device. Prefer device-agnostic snippets so the docs work on CUDA, XPU, MPS, and other supported accelerators:
+
+```python
+from diffusers.utils.torch_utils import get_device
+
+device = get_device()
+pipe.to(device)
+```
+
+Use `device` for `.to(...)`, `device_map`, `torch.Generator(device=...)`, tensors, and component placement where applicable. Only keep explicit CUDA APIs or strings when the guide is truly CUDA/NVIDIA-specific, for example CUDA kernel documentation, NCCL multi-GPU examples, or benchmarks that explicitly target NVIDIA GPUs. In those cases, the surrounding text should make the CUDA-only requirement clear.
+
 ## File structure
 
 A new model PR roughly lands these files (the contents of `pipelines/<model>/` and `modular_pipelines/<model>/` live in their guides):

.ai/skills/self-review/SKILL.md

@@ -34,6 +34,8 @@ files, scope to your own commits: `git log main..HEAD --oneline`, then
 it and review against it; don't rely on a remembered copy. For the areas you
 touched, also read `.ai/models.md`, `.ai/pipelines.md`, or `.ai/modular.md`.
 
+When reviewing docs in `docs/source/en/`, check that examples are device agnostic by default. Flag newly added `.to("cuda")`, `device="cuda"`, `device_map="cuda"`, `torch.Generator("cuda")`, or similar hardcoded CUDA usage unless the page is explicitly CUDA/NVIDIA-specific, such as CUDA kernels, NCCL/distributed GPU examples, or NVIDIA GPU benchmarks.
+
 ## 3. Report
 
 - **Blocking issues** — numbered. Each: title → explanation → `file.py:line` →

docs/source/en/advanced_inference/outpaint.md

@@ -62,6 +62,9 @@ from diffusers import (
     StableDiffusionXLControlNetPipeline,
     StableDiffusionXLInpaintPipeline,
 )
+from diffusers.utils.torch_utils import get_device
+
+device = get_device()
 
 def scale_and_paste(original_image):
     aspect_ratio = original_image.width / original_image.height
@@ -117,10 +120,10 @@ controlnets = [
         "diffusers/controlnet-zoe-depth-sdxl-1.0", dtype=torch.float16
     ),
 ]
-vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", dtype=torch.float16).to("cuda")
+vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", dtype=torch.float16).to(device)
 pipeline = StableDiffusionXLControlNetPipeline.from_pretrained(
     "SG161222/RealVisXL_V4.0", dtype=torch.float16, variant="fp16", controlnet=controlnets, vae=vae
-).to("cuda")
+).to(device)
 
 def generate_image(prompt, negative_prompt, inpaint_image, zoe_image, seed: int = None):
     if seed is None:
@@ -176,7 +179,7 @@ pipeline = StableDiffusionXLInpaintPipeline.from_pretrained(
     dtype=torch.float16,
     variant="fp16",
     vae=vae,
-).to("cuda")
+).to(device)
 ```
 
 Prepare a mask for the final outpainted image. To create a more natural transition between the original image and the outpainted background, blur the mask to help it blend better.

docs/source/en/api/models/allegro_transformer3d.md

@@ -17,8 +17,11 @@ The model can be loaded with the following code snippet.
 
 ```python
 from diffusers import AllegroTransformer3DModel
+from diffusers.utils.torch_utils import get_device
 
-transformer = AllegroTransformer3DModel.from_pretrained("rhymes-ai/Allegro", subfolder="transformer", dtype=torch.bfloat16).to("cuda")
+
+device = get_device()
+transformer = AllegroTransformer3DModel.from_pretrained("rhymes-ai/Allegro", subfolder="transformer", dtype=torch.bfloat16).to(device)
 ```
 
 ## AllegroTransformer3DModel

docs/source/en/api/models/asymmetricautoencoderkl.md

@@ -29,9 +29,12 @@ Evaluation results can be found in section 4.1 of the original paper.
 
 ```python
 from diffusers import AsymmetricAutoencoderKL, StableDiffusionInpaintPipeline
+from diffusers.utils.torch_utils import get_device
 from diffusers.utils import load_image, make_image_grid
 
 
+
+device = get_device()
 prompt = "a photo of a person with beard"
 img_url = "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/repaint/celeba_hq_256.png"
 mask_url = "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/repaint/mask_256.png"
@@ -41,7 +44,7 @@ mask_image = load_image(mask_url).resize((512, 512))
 
 pipe = StableDiffusionInpaintPipeline.from_pretrained("stable-diffusion-v1-5/stable-diffusion-inpainting")
 pipe.vae = AsymmetricAutoencoderKL.from_pretrained("cross-attention/asymmetric-autoencoder-kl-x-1-5")
-pipe.to("cuda")
+pipe.to(device)
 
 image = pipe(prompt=prompt, image=original_image, mask_image=mask_image).images[0]
 make_image_grid([original_image, mask_image, image], rows=1, cols=3)

docs/source/en/api/models/autoencoder_dc.md

@@ -35,8 +35,11 @@ Load a model in Diffusers format with [`~ModelMixin.from_pretrained`].
 
 ```python
 from diffusers import AutoencoderDC
+from diffusers.utils.torch_utils import get_device
 
-ae = AutoencoderDC.from_pretrained("mit-han-lab/dc-ae-f32c32-sana-1.0-diffusers", dtype=torch.float32).to("cuda")
+
+device = get_device()
+ae = AutoencoderDC.from_pretrained("mit-han-lab/dc-ae-f32c32-sana-1.0-diffusers", dtype=torch.float32).to(device)
 ```
 
 ## Load a model in Diffusers via `from_single_file`

docs/source/en/api/models/autoencoder_rae.md

@@ -31,27 +31,32 @@ The following RAE models are released and supported in Diffusers:
 
 ```python
 from diffusers import AutoencoderRAE
+from diffusers.utils.torch_utils import get_device
 
+device = get_device()
 model = AutoencoderRAE.from_pretrained(
     "nyu-visionx/RAE-dinov2-wReg-base-ViTXL-n08"
-).to("cuda").eval()
+).to(device).eval()
 ```
 
 ## Encoding and decoding a real image
 
 ```python
 import torch
+from diffusers.utils.torch_utils import get_device
 from diffusers import AutoencoderRAE
 from diffusers.utils import load_image
 from torchvision.transforms.functional import to_tensor, to_pil_image
 
+
+device = get_device()
 model = AutoencoderRAE.from_pretrained(
     "nyu-visionx/RAE-dinov2-wReg-base-ViTXL-n08"
-).to("cuda").eval()
+).to(device).eval()
 
 image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/cat.png")
 image = image.convert("RGB").resize((224, 224))
-x = to_tensor(image).unsqueeze(0).to("cuda")  # (1, 3, 224, 224), values in [0, 1]
+x = to_tensor(image).unsqueeze(0).to(device)  # (1, 3, 224, 224), values in [0, 1]
 
 with torch.no_grad():
     latents = model.encode(x).latent        # (1, 768, 16, 16)
@@ -66,9 +71,12 @@ recon_image.save("recon.png")
 Some pretrained checkpoints include per-channel `latents_mean` and `latents_std` statistics for normalizing the latent space. When present, `encode` and `decode` automatically apply the normalization and denormalization, respectively.
 
 ```python
+from diffusers.utils.torch_utils import get_device
+
+device = get_device()
 model = AutoencoderRAE.from_pretrained(
     "nyu-visionx/RAE-dinov2-wReg-base-ViTXL-n08"
-).to("cuda").eval()
+).to(device).eval()
 
 # Latent normalization is handled automatically inside encode/decode
 # when the checkpoint config includes latents_mean/latents_std.

docs/source/en/api/models/autoencoder_tiny.md

@@ -18,13 +18,16 @@ To use with Stable Diffusion v-2.1:
 
 ```python
 import torch
+from diffusers.utils.torch_utils import get_device
 from diffusers import DiffusionPipeline, AutoencoderTiny
 
+
+device = get_device()
 pipe = DiffusionPipeline.from_pretrained(
     "stabilityai/stable-diffusion-2-1-base", dtype=torch.float16
 )
 pipe.vae = AutoencoderTiny.from_pretrained("madebyollin/taesd", dtype=torch.float16)
-pipe = pipe.to("cuda")
+pipe = pipe.to(device)
 
 prompt = "slice of delicious New York-style berry cheesecake"
 image = pipe(prompt, num_inference_steps=25).images[0]
@@ -36,12 +39,14 @@ To use with Stable Diffusion XL 1.0
 ```python
 import torch
 from diffusers import DiffusionPipeline, AutoencoderTiny
+from diffusers.utils.torch_utils import get_device
 
+device = get_device()
 pipe = DiffusionPipeline.from_pretrained(
     "stabilityai/stable-diffusion-xl-base-1.0", dtype=torch.float16
 )
 pipe.vae = AutoencoderTiny.from_pretrained("madebyollin/taesdxl", dtype=torch.float16)
-pipe = pipe.to("cuda")
+pipe = pipe.to(device)
 
 prompt = "slice of delicious New York-style berry cheesecake"
 image = pipe(prompt, num_inference_steps=25).images[0]

docs/source/en/api/models/autoencoderkl_allegro.md

@@ -17,8 +17,11 @@ The model can be loaded with the following code snippet.
 
 ```python
 from diffusers import AutoencoderKLAllegro
+from diffusers.utils.torch_utils import get_device
 
-vae = AutoencoderKLAllegro.from_pretrained("rhymes-ai/Allegro", subfolder="vae", dtype=torch.float32).to("cuda")
+
+device = get_device()
+vae = AutoencoderKLAllegro.from_pretrained("rhymes-ai/Allegro", subfolder="vae", dtype=torch.float32).to(device)
 ```
 
 ## AutoencoderKLAllegro

docs/source/en/api/models/autoencoderkl_audio_ltx_2.md

@@ -17,8 +17,11 @@ The model can be loaded with the following code snippet.
 
 ```python
 from diffusers import AutoencoderKLLTX2Audio
+from diffusers.utils.torch_utils import get_device
 
-vae = AutoencoderKLLTX2Audio.from_pretrained("Lightricks/LTX-2", subfolder="vae", dtype=torch.float32).to("cuda")
+
+device = get_device()
+vae = AutoencoderKLLTX2Audio.from_pretrained("Lightricks/LTX-2", subfolder="vae", dtype=torch.float32).to(device)
 ```
 
 ## AutoencoderKLLTX2Audio

docs/source/en/api/models/autoencoderkl_cogvideox.md

@@ -17,8 +17,11 @@ The model can be loaded with the following code snippet.
 
 ```python
 from diffusers import AutoencoderKLCogVideoX
+from diffusers.utils.torch_utils import get_device
 
-vae = AutoencoderKLCogVideoX.from_pretrained("THUDM/CogVideoX-2b", subfolder="vae", dtype=torch.float16).to("cuda")
+
+device = get_device()
+vae = AutoencoderKLCogVideoX.from_pretrained("THUDM/CogVideoX-2b", subfolder="vae", dtype=torch.float16).to(device)
 ```
 
 ## AutoencoderKLCogVideoX

docs/source/en/api/models/autoencoderkl_ltx_2.md

@@ -17,8 +17,11 @@ The model can be loaded with the following code snippet.
 
 ```python
 from diffusers import AutoencoderKLLTX2Video
+from diffusers.utils.torch_utils import get_device
 
-vae = AutoencoderKLLTX2Video.from_pretrained("Lightricks/LTX-2", subfolder="vae", dtype=torch.float32).to("cuda")
+
+device = get_device()
+vae = AutoencoderKLLTX2Video.from_pretrained("Lightricks/LTX-2", subfolder="vae", dtype=torch.float32).to(device)
 ```
 
 ## AutoencoderKLLTX2Video

docs/source/en/api/models/autoencoderkl_ltx_video.md

@@ -17,8 +17,11 @@ The model can be loaded with the following code snippet.
 
 ```python
 from diffusers import AutoencoderKLLTXVideo
+from diffusers.utils.torch_utils import get_device
 
-vae = AutoencoderKLLTXVideo.from_pretrained("Lightricks/LTX-Video", subfolder="vae", dtype=torch.float32).to("cuda")
+
+device = get_device()
+vae = AutoencoderKLLTXVideo.from_pretrained("Lightricks/LTX-Video", subfolder="vae", dtype=torch.float32).to(device)
 ```
 
 ## AutoencoderKLLTXVideo

docs/source/en/api/models/autoencoderkl_magvit.md

@@ -17,8 +17,11 @@ The model can be loaded with the following code snippet.
 
 ```python
 from diffusers import AutoencoderKLMagvit
+from diffusers.utils.torch_utils import get_device
 
-vae = AutoencoderKLMagvit.from_pretrained("alibaba-pai/EasyAnimateV5.1-12b-zh", subfolder="vae", dtype=torch.float16).to("cuda")
+
+device = get_device()
+vae = AutoencoderKLMagvit.from_pretrained("alibaba-pai/EasyAnimateV5.1-12b-zh", subfolder="vae", dtype=torch.float16).to(device)
 ```
 
 ## AutoencoderKLMagvit

docs/source/en/api/models/autoencoderkl_minimax_h3.md

@@ -23,11 +23,14 @@ The temporal geometry is fixed by `clip_length` (17 pixel frames per encoder chu
 
 ```python
 import torch
+from diffusers.utils.torch_utils import get_device
 from diffusers import AutoencoderKLMiniMaxH3
 
+
+device = get_device()
 vae = AutoencoderKLMiniMaxH3.from_pretrained(
     "MiniMaxAI/MiniMax-H3", subfolder="vae", dtype=torch.float32
-).to("cuda")
+).to(device)
 ```
 
 ## AutoencoderKLMiniMaxH3

docs/source/en/api/models/autoencoderkl_minimax_h3_audio.md

@@ -21,11 +21,14 @@ The autoencoder is **mono**, and it normalizes latents per channel with `latents
 
 ```python
 import torch
+from diffusers.utils.torch_utils import get_device
 from diffusers import AutoencoderKLMiniMaxH3Audio
 
+
+device = get_device()
 audio_vae = AutoencoderKLMiniMaxH3Audio.from_pretrained(
     "MiniMaxAI/MiniMax-H3", subfolder="audio_vae", dtype=torch.float32
-).to("cuda")
+).to(device)
 ```
 
 ## AutoencoderKLMiniMaxH3Audio

docs/source/en/api/models/autoencoderkl_mochi.md

@@ -17,8 +17,11 @@ The model can be loaded with the following code snippet.
 
 ```python
 from diffusers import AutoencoderKLMochi
+from diffusers.utils.torch_utils import get_device
 
-vae = AutoencoderKLMochi.from_pretrained("genmo/mochi-1-preview", subfolder="vae", dtype=torch.float32).to("cuda")
+
+device = get_device()
+vae = AutoencoderKLMochi.from_pretrained("genmo/mochi-1-preview", subfolder="vae", dtype=torch.float32).to(device)
 ```
 
 ## AutoencoderKLMochi

docs/source/en/api/models/cogvideox_transformer3d.md

@@ -17,8 +17,11 @@ The model can be loaded with the following code snippet.
 
 ```python
 from diffusers import CogVideoXTransformer3DModel
+from diffusers.utils.torch_utils import get_device
 
-transformer = CogVideoXTransformer3DModel.from_pretrained("THUDM/CogVideoX-2b", subfolder="transformer", dtype=torch.float16).to("cuda")
+
+device = get_device()
+transformer = CogVideoXTransformer3DModel.from_pretrained("THUDM/CogVideoX-2b", subfolder="transformer", dtype=torch.float16).to(device)
 ```
 
 ## CogVideoXTransformer3DModel

docs/source/en/api/models/cogview3plus_transformer2d.md

@@ -17,8 +17,11 @@ The model can be loaded with the following code snippet.
 
 ```python
 from diffusers import CogView3PlusTransformer2DModel
+from diffusers.utils.torch_utils import get_device
 
-transformer = CogView3PlusTransformer2DModel.from_pretrained("THUDM/CogView3Plus-3b", subfolder="transformer", dtype=torch.bfloat16).to("cuda")
+
+device = get_device()
+transformer = CogView3PlusTransformer2DModel.from_pretrained("THUDM/CogView3Plus-3b", subfolder="transformer", dtype=torch.bfloat16).to(device)
 ```
 
 ## CogView3PlusTransformer2DModel

docs/source/en/api/models/cogview4_transformer2d.md

@@ -17,8 +17,11 @@ The model can be loaded with the following code snippet.
 
 ```python
 from diffusers import CogView4Transformer2DModel
+from diffusers.utils.torch_utils import get_device
 
-transformer = CogView4Transformer2DModel.from_pretrained("THUDM/CogView4-6B", subfolder="transformer", dtype=torch.bfloat16).to("cuda")
+
+device = get_device()
+transformer = CogView4Transformer2DModel.from_pretrained("THUDM/CogView4-6B", subfolder="transformer", dtype=torch.bfloat16).to(device)
 ```
 
 ## CogView4Transformer2DModel

docs/source/en/api/models/consisid_transformer3d.md

@@ -17,8 +17,11 @@ The model can be loaded with the following code snippet.
 
 ```python
 from diffusers import ConsisIDTransformer3DModel
+from diffusers.utils.torch_utils import get_device
 
-transformer = ConsisIDTransformer3DModel.from_pretrained("BestWishYsh/ConsisID-preview", subfolder="transformer", dtype=torch.bfloat16).to("cuda")
+
+device = get_device()
+transformer = ConsisIDTransformer3DModel.from_pretrained("BestWishYsh/ConsisID-preview", subfolder="transformer", dtype=torch.bfloat16).to(device)
 ```
 
 ## ConsisIDTransformer3DModel

docs/source/en/api/models/controlnet.md

@@ -25,7 +25,10 @@ from the original format using [`FromOriginalModelMixin.from_single_file`] as fo
 
 ```py
 from diffusers import StableDiffusionControlNetPipeline, ControlNetModel
+from diffusers.utils.torch_utils import get_device
 
+
+device = get_device()
 url = "https://huggingface.co/lllyasviel/ControlNet-v1-1/blob/main/control_v11p_sd15_canny.pth"  # can also be a local path
 controlnet = ControlNetModel.from_single_file(url)
 
@@ -43,8 +46,8 @@ from diffusers import ControlNetModel, UNet2DConditionModel
 lora_id = "stabilityai/control-lora"
 lora_filename = "control-LoRAs-rank128/control-lora-canny-rank128.safetensors"
 
-unet = UNet2DConditionModel.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", subfolder="unet", dtype=torch.bfloat16).to("cuda")
-controlnet = ControlNetModel.from_unet(unet).to(device="cuda", dtype=torch.bfloat16)
+unet = UNet2DConditionModel.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", subfolder="unet", dtype=torch.bfloat16).to(device)
+controlnet = ControlNetModel.from_unet(unet).to(device=device, dtype=torch.bfloat16)
 controlnet.load_lora_adapter(lora_id, weight_name=lora_filename, prefix=None, controlnet_config=controlnet.config)
 ```