Merge branch 'main' into add-trtquant-backend

Author: sayakpaul

docs/source/en/training/distributed_inference.md

@@ -223,7 +223,7 @@ from diffusers.image_processor import VaeImageProcessor
 import torch 
 
 vae = AutoencoderKL.from_pretrained(ckpt_id, subfolder="vae", torch_dtype=torch.bfloat16).to("cuda")
-vae_scale_factor = 2 ** (len(vae.config.block_out_channels))
+vae_scale_factor = 2 ** (len(vae.config.block_out_channels) - 1)
 image_processor = VaeImageProcessor(vae_scale_factor=vae_scale_factor)
 
 with torch.no_grad():

docs/source/en/tutorials/autopipeline.md

@@ -12,112 +12,56 @@ specific language governing permissions and limitations under the License.
 
 # AutoPipeline
 
-Diffusers provides many pipelines for basic tasks like generating images, videos, audio, and inpainting. On top of these, there are specialized pipelines for adapters and features like upscaling, super-resolution, and more. Different pipeline classes can even use the same checkpoint because they share the same pretrained model! With so many different pipelines, it can be overwhelming to know which pipeline class to use.
+[AutoPipeline](../api/models/auto_model) is a *task-and-model* pipeline that automatically selects the correct pipeline subclass based on the task. It handles the complexity of loading different pipeline subclasses without needing to know the specific pipeline subclass name.
 
-The [AutoPipeline](../api/pipelines/auto_pipeline) class is designed to simplify the variety of pipelines in Diffusers. It is a generic *task-first* pipeline that lets you focus on a task ([`AutoPipelineForText2Image`], [`AutoPipelineForImage2Image`], and [`AutoPipelineForInpainting`]) without needing to know the specific pipeline class. The [AutoPipeline](../api/pipelines/auto_pipeline) automatically detects the correct pipeline class to use.
+This is unlike [`DiffusionPipeline`], a *model-only* pipeline that automatically selects the pipeline subclass based on the model.
 
-For example, let's use the [dreamlike-art/dreamlike-photoreal-2.0](https://hf.co/dreamlike-art/dreamlike-photoreal-2.0) checkpoint.
-
-Under the hood, [AutoPipeline](../api/pipelines/auto_pipeline):
-
-1. Detects a `"stable-diffusion"` class from the [model_index.json](https://hf.co/dreamlike-art/dreamlike-photoreal-2.0/blob/main/model_index.json) file.
-2. Depending on the task you're interested in, it loads the [`StableDiffusionPipeline`], [`StableDiffusionImg2ImgPipeline`], or [`StableDiffusionInpaintPipeline`]. Any parameter (`strength`, `num_inference_steps`, etc.) you would pass to these specific pipelines can also be passed to the [AutoPipeline](../api/pipelines/auto_pipeline).
-
-<hfoptions id="autopipeline">
-<hfoption id="text-to-image">
+[`AutoPipelineForImage2Image`] returns a specific pipeline subclass, (for example, [`StableDiffusionXLImg2ImgPipeline`]), which can only be used for image-to-image tasks.
 
 ```py
-from diffusers import AutoPipelineForText2Image
 import torch
-
-pipe_txt2img = AutoPipelineForText2Image.from_pretrained(
-    "dreamlike-art/dreamlike-photoreal-2.0", torch_dtype=torch.float16, use_safetensors=True
-).to("cuda")
-
-prompt = "cinematic photo of Godzilla eating sushi with a cat in a izakaya, 35mm photograph, film, professional, 4k, highly detailed"
-generator = torch.Generator(device="cpu").manual_seed(37)
-image = pipe_txt2img(prompt, generator=generator).images[0]
-image
-```
-
-<div class="flex justify-center">
-    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/autopipeline-text2img.png"/>
-</div>
-
-</hfoption>
-<hfoption id="image-to-image">
-
-```py
 from diffusers import AutoPipelineForImage2Image
-from diffusers.utils import load_image
-import torch
-
-pipe_img2img = AutoPipelineForImage2Image.from_pretrained(
-    "dreamlike-art/dreamlike-photoreal-2.0", torch_dtype=torch.float16, use_safetensors=True
-).to("cuda")
-
-init_image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/autopipeline-text2img.png")
-
-prompt = "cinematic photo of Godzilla eating burgers with a cat in a fast food restaurant, 35mm photograph, film, professional, 4k, highly detailed"
-generator = torch.Generator(device="cpu").manual_seed(53)
-image = pipe_img2img(prompt, image=init_image, generator=generator).images[0]
-image
-```
-
-Notice how the [dreamlike-art/dreamlike-photoreal-2.0](https://hf.co/dreamlike-art/dreamlike-photoreal-2.0) checkpoint is used for both text-to-image and image-to-image tasks? To save memory and avoid loading the checkpoint twice, use the [`~DiffusionPipeline.from_pipe`] method.
 
-```py
-pipe_img2img = AutoPipelineForImage2Image.from_pipe(pipe_txt2img).to("cuda")
-image = pipeline(prompt, image=init_image, generator=generator).images[0]
-image
+pipeline = AutoPipelineForImage2Image.from_pretrained(
+  "RunDiffusion/Juggernaut-XL-v9", torch_dtype=torch.bfloat16, device_map="cuda",
+)
+print(pipeline)
+"StableDiffusionXLImg2ImgPipeline {
+  "_class_name": "StableDiffusionXLImg2ImgPipeline",
+  ...
+"
 ```
 
-You can learn more about the [`~DiffusionPipeline.from_pipe`] method in the [Reuse a pipeline](../using-diffusers/loading#reuse-a-pipeline) guide.
-
-<div class="flex justify-center">
-    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/autopipeline-img2img.png"/>
-</div>
-
-</hfoption>
-<hfoption id="inpainting">
+Loading the same model with [`DiffusionPipeline`] returns the [`StableDiffusionXLPipeline`] subclass. It can be used for text-to-image, image-to-image, or inpainting tasks depending on the inputs.
 
 ```py
-from diffusers import AutoPipelineForInpainting
-from diffusers.utils import load_image
 import torch
+from diffusers import DiffusionPipeline
 
-pipeline = AutoPipelineForInpainting.from_pretrained(
-    "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16, use_safetensors=True
-).to("cuda")
-
-init_image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/autopipeline-img2img.png")
-mask_image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/autopipeline-mask.png")
-
-prompt = "cinematic photo of a owl, 35mm photograph, film, professional, 4k, highly detailed"
-generator = torch.Generator(device="cpu").manual_seed(38)
-image = pipeline(prompt, image=init_image, mask_image=mask_image, generator=generator, strength=0.4).images[0]
-image
+pipeline = DiffusionPipeline.from_pretrained(
+  "RunDiffusion/Juggernaut-XL-v9", torch_dtype=torch.bfloat16, device_map="cuda",
+)
+print(pipeline)
+"StableDiffusionXLPipeline {
+  "_class_name": "StableDiffusionXLPipeline",
+  ...
+"
 ```
 
-<div class="flex justify-center">
-    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/autopipeline-inpaint.png"/>
-</div>
+Check the [mappings](https://github.com/huggingface/diffusers/blob/130fd8df54f24ffb006d84787b598d8adc899f23/src/diffusers/pipelines/auto_pipeline.py#L114) to see whether a model is supported or not.
 
-</hfoption>
-</hfoptions>
-
-## Unsupported checkpoints
-
-The [AutoPipeline](../api/pipelines/auto_pipeline) supports [Stable Diffusion](../api/pipelines/stable_diffusion/overview), [Stable Diffusion XL](../api/pipelines/stable_diffusion/stable_diffusion_xl), [ControlNet](../api/pipelines/controlnet), [Kandinsky 2.1](../api/pipelines/kandinsky.md), [Kandinsky 2.2](../api/pipelines/kandinsky_v22), and [DeepFloyd IF](../api/pipelines/deepfloyd_if) checkpoints.
-
-If you try to load an unsupported checkpoint, you'll get an error.
+Trying to load an unsupported model returns an error.
 
 ```py
-from diffusers import AutoPipelineForImage2Image
 import torch
+from diffusers import AutoPipelineForImage2Image
 
 pipeline = AutoPipelineForImage2Image.from_pretrained(
-    "openai/shap-e-img2img", torch_dtype=torch.float16, use_safetensors=True
+    "openai/shap-e-img2img", torch_dtype=torch.float16,
 )
 "ValueError: AutoPipeline can't find a pipeline linked to ShapEImg2ImgPipeline for None"
 ```
+
+There are three types of [AutoPipeline](../api/models/auto_model) classes, [`AutoPipelineForText2Image`], [`AutoPipelineForImage2Image`] and [`AutoPipelineForInpainting`]. Each of these classes have a predefined mapping, linking a pipeline to their task-specific subclass.
+
+When [`~AutoPipelineForText2Image.from_pretrained`] is called, it extracts the class name from the `model_index.json` file and selects the appropriate pipeline subclass for the task based on the mapping.

docs/source/zh/training/distributed_inference.md

@@ -223,7 +223,7 @@ from diffusers.image_processor import VaeImageProcessor
 import torch 
 
 vae = AutoencoderKL.from_pretrained(ckpt_id, subfolder="vae", torch_dtype=torch.bfloat16).to("cuda")
-vae_scale_factor = 2 ** (len(vae.config.block_out_channels))
+vae_scale_factor = 2 ** (len(vae.config.block_out_channels) - 1)
 image_processor = VaeImageProcessor(vae_scale_factor=vae_scale_factor)
 
 with torch.no_grad():

examples/advanced_diffusion_training/train_dreambooth_lora_flux_advanced.py

@@ -1399,6 +1399,7 @@ def main(args):
                 torch_dtype = torch.float16
             elif args.prior_generation_precision == "bf16":
                 torch_dtype = torch.bfloat16
+
             pipeline = FluxPipeline.from_pretrained(
                 args.pretrained_model_name_or_path,
                 torch_dtype=torch_dtype,
@@ -1419,7 +1420,8 @@ def main(args):
             for example in tqdm(
                 sample_dataloader, desc="Generating class images", disable=not accelerator.is_local_main_process
             ):
-                images = pipeline(example["prompt"]).images
+                with torch.autocast(device_type=accelerator.device.type, dtype=torch_dtype):
+                    images = pipeline(prompt=example["prompt"]).images
 
                 for i, image in enumerate(images):
                     hash_image = insecure_hashlib.sha1(image.tobytes()).hexdigest()

examples/dreambooth/train_dreambooth_lora_flux.py

@@ -1131,6 +1131,7 @@ def main(args):
                 torch_dtype = torch.float16
             elif args.prior_generation_precision == "bf16":
                 torch_dtype = torch.bfloat16
+
             pipeline = FluxPipeline.from_pretrained(
                 args.pretrained_model_name_or_path,
                 torch_dtype=torch_dtype,
@@ -1151,16 +1152,16 @@ def main(args):
             for example in tqdm(
                 sample_dataloader, desc="Generating class images", disable=not accelerator.is_local_main_process
             ):
-                images = pipeline(example["prompt"]).images
+                with torch.autocast(device_type=accelerator.device.type, dtype=torch_dtype):
+                    images = pipeline(prompt=example["prompt"]).images
 
                 for i, image in enumerate(images):
                     hash_image = insecure_hashlib.sha1(image.tobytes()).hexdigest()
                     image_filename = class_images_dir / f"{example['index'][i] + cur_class_images}-{hash_image}.jpg"
                     image.save(image_filename)
 
             del pipeline
-            if torch.cuda.is_available():
-                torch.cuda.empty_cache()
+            free_memory()
 
     # Handle the repository creation
     if accelerator.is_main_process:
@@ -1728,6 +1729,10 @@ def get_sigmas(timesteps, n_dim=4, dtype=torch.float32):
                             device=accelerator.device,
                             prompt=args.instance_prompt,
                         )
+                    else:
+                        prompt_embeds, pooled_prompt_embeds, text_ids = compute_text_embeddings(
+                            prompts, text_encoders, tokenizers
+                        )
 
                 # Convert images to latent space
                 if args.cache_latents:

src/diffusers/models/attention_dispatch.py

@@ -26,6 +26,7 @@
     is_flash_attn_3_available,
     is_flash_attn_available,
     is_flash_attn_version,
+    is_kernels_available,
     is_sageattention_available,
     is_sageattention_version,
     is_torch_npu_available,
@@ -35,7 +36,7 @@
     is_xformers_available,
     is_xformers_version,
 )
-from ..utils.constants import DIFFUSERS_ATTN_BACKEND, DIFFUSERS_ATTN_CHECKS
+from ..utils.constants import DIFFUSERS_ATTN_BACKEND, DIFFUSERS_ATTN_CHECKS, DIFFUSERS_ENABLE_HUB_KERNELS
 
 
 _REQUIRED_FLASH_VERSION = "2.6.3"
@@ -67,6 +68,17 @@
     flash_attn_3_func = None
     flash_attn_3_varlen_func = None
 
+if DIFFUSERS_ENABLE_HUB_KERNELS:
+    if not is_kernels_available():
+        raise ImportError(
+            "To use FA3 kernel for your hardware from the Hub, the `kernels` library must be installed. Install with `pip install kernels`."
+        )
+    from ..utils.kernels_utils import _get_fa3_from_hub
+
+    flash_attn_interface_hub = _get_fa3_from_hub()
+    flash_attn_3_func_hub = flash_attn_interface_hub.flash_attn_func
+else:
+    flash_attn_3_func_hub = None
 
 if _CAN_USE_SAGE_ATTN:
     from sageattention import (
@@ -153,6 +165,8 @@ class AttentionBackendName(str, Enum):
     FLASH_VARLEN = "flash_varlen"
     _FLASH_3 = "_flash_3"
     _FLASH_VARLEN_3 = "_flash_varlen_3"
+    _FLASH_3_HUB = "_flash_3_hub"
+    # _FLASH_VARLEN_3_HUB = "_flash_varlen_3_hub"  # not supported yet.
 
     # PyTorch native
     FLEX = "flex"
@@ -351,6 +365,17 @@ def _check_attention_backend_requirements(backend: AttentionBackendName) -> None
                 f"Flash Attention 3 backend '{backend.value}' is not usable because of missing package or the version is too old. Please build FA3 beta release from source."
             )
 
+    # TODO: add support Hub variant of FA3 varlen later
+    elif backend in [AttentionBackendName._FLASH_3_HUB]:
+        if not DIFFUSERS_ENABLE_HUB_KERNELS:
+            raise RuntimeError(
+                f"Flash Attention 3 Hub backend '{backend.value}' is not usable because the `DIFFUSERS_ENABLE_HUB_KERNELS` env var isn't set. Please set it like `export DIFFUSERS_ENABLE_HUB_KERNELS=yes`."
+            )
+        if not is_kernels_available():
+            raise RuntimeError(
+                f"Flash Attention 3 Hub backend '{backend.value}' is not usable because the `kernels` package isn't available. Please install it with `pip install kernels`."
+            )
+
     elif backend in [
         AttentionBackendName.SAGE,
         AttentionBackendName.SAGE_VARLEN,
@@ -657,6 +682,44 @@ def _flash_attention_3(
     return (out, lse) if return_attn_probs else out
 
 
+@_AttentionBackendRegistry.register(
+    AttentionBackendName._FLASH_3_HUB,
+    constraints=[_check_device, _check_qkv_dtype_bf16_or_fp16, _check_shape],
+)
+def _flash_attention_3_hub(
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    scale: Optional[float] = None,
+    is_causal: bool = False,
+    window_size: Tuple[int, int] = (-1, -1),
+    softcap: float = 0.0,
+    deterministic: bool = False,
+    return_attn_probs: bool = False,
+) -> torch.Tensor:
+    out = flash_attn_3_func_hub(
+        q=query,
+        k=key,
+        v=value,
+        softmax_scale=scale,
+        causal=is_causal,
+        qv=None,
+        q_descale=None,
+        k_descale=None,
+        v_descale=None,
+        window_size=window_size,
+        softcap=softcap,
+        num_splits=1,
+        pack_gqa=None,
+        deterministic=deterministic,
+        sm_margin=0,
+        return_attn_probs=return_attn_probs,
+    )
+    # When `return_attn_probs` is True, the above returns a tuple of
+    # actual outputs and lse.
+    return (out[0], out[1]) if return_attn_probs else out
+
+
 @_AttentionBackendRegistry.register(
     AttentionBackendName._FLASH_VARLEN_3,
     constraints=[_check_device, _check_qkv_dtype_bf16_or_fp16, _check_shape],

src/diffusers/utils/constants.py

@@ -46,6 +46,7 @@
 DEFAULT_HF_PARALLEL_LOADING_WORKERS = 8
 HF_ENABLE_PARALLEL_LOADING = os.environ.get("HF_ENABLE_PARALLEL_LOADING", "").upper() in ENV_VARS_TRUE_VALUES
 DIFFUSERS_DISABLE_REMOTE_CODE = os.getenv("DIFFUSERS_DISABLE_REMOTE_CODE", "false").lower() in ENV_VARS_TRUE_VALUES
+DIFFUSERS_ENABLE_HUB_KERNELS = os.environ.get("DIFFUSERS_ENABLE_HUB_KERNELS", "").upper() in ENV_VARS_TRUE_VALUES
 
 # Below should be `True` if the current version of `peft` and `transformers` are compatible with
 # PEFT backend. Will automatically fall back to PEFT backend if the correct versions of the libraries are

src/diffusers/utils/kernels_utils.py

@@ -0,0 +1,23 @@
+from ..utils import get_logger
+from .import_utils import is_kernels_available
+
+
+logger = get_logger(__name__)
+
+
+_DEFAULT_HUB_ID_FA3 = "kernels-community/flash-attn3"
+
+
+def _get_fa3_from_hub():
+    if not is_kernels_available():
+        return None
+    else:
+        from kernels import get_kernel
+
+        try:
+            # TODO: temporary revision for now. Remove when merged upstream into `main`.
+            flash_attn_3_hub = get_kernel(_DEFAULT_HUB_ID_FA3, revision="fake-ops-return-probs")
+            return flash_attn_3_hub
+        except Exception as e:
+            logger.error(f"An error occurred while fetching kernel '{_DEFAULT_HUB_ID_FA3}' from the Hub: {e}")
+            raise

tests/quantization/quanto/test_quanto.py

@@ -13,7 +13,7 @@
     nightly,
     numpy_cosine_similarity_distance,
     require_accelerate,
-    require_big_accelerator,
+    require_accelerator,
     require_torch_cuda_compatibility,
     torch_device,
 )
@@ -31,7 +31,7 @@
 
 
 @nightly
-@require_big_accelerator
+@require_accelerator
 @require_accelerate
 class QuantoBaseTesterMixin:
     model_id = None