Merge branch 'main' into deprecate-slicing-tiling-pipe

Author: sayakpaul

docs/source/en/api/pipelines/cogvideox.md

@@ -50,7 +50,7 @@ from diffusers.utils import export_to_video
 pipeline_quant_config = PipelineQuantizationConfig(
   quant_backend="torchao",
   quant_kwargs={"quant_type": "int8wo"},
-  components_to_quantize=["transformer"]
+  components_to_quantize="transformer"
 )
 
 # fp8 layerwise weight-casting

docs/source/en/api/pipelines/hunyuan_video.md

@@ -54,7 +54,7 @@ pipeline_quant_config = PipelineQuantizationConfig(
       "bnb_4bit_quant_type": "nf4",
       "bnb_4bit_compute_dtype": torch.bfloat16
       },
-    components_to_quantize=["transformer"]
+    components_to_quantize="transformer"
 )
 
 pipeline = HunyuanVideoPipeline.from_pretrained(
@@ -91,7 +91,7 @@ pipeline_quant_config = PipelineQuantizationConfig(
       "bnb_4bit_quant_type": "nf4",
       "bnb_4bit_compute_dtype": torch.bfloat16
       },
-    components_to_quantize=["transformer"]
+    components_to_quantize="transformer"
 )
 
 pipeline = HunyuanVideoPipeline.from_pretrained(
@@ -139,7 +139,7 @@ export_to_video(video, "output.mp4", fps=15)
         "bnb_4bit_quant_type": "nf4",
         "bnb_4bit_compute_dtype": torch.bfloat16
         },
-      components_to_quantize=["transformer"]
+      components_to_quantize="transformer"
   )
 
   pipeline = HunyuanVideoPipeline.from_pretrained(

docs/source/en/optimization/memory.md

@@ -291,13 +291,53 @@ Group offloading moves groups of internal layers ([torch.nn.ModuleList](https://
 > [!WARNING]
 > Group offloading may not work with all models if the forward implementation contains weight-dependent device casting of inputs because it may clash with group offloading's device casting mechanism.
 
-Call [`~ModelMixin.enable_group_offload`] to enable it for standard Diffusers model components that inherit from [`ModelMixin`]. For other model components that don't inherit from [`ModelMixin`], such as a generic [torch.nn.Module](https://pytorch.org/docs/stable/generated/torch.nn.Module.html), use [`~hooks.apply_group_offloading`] instead.
-
-The `offload_type` parameter can be set to `block_level` or `leaf_level`.
+Enable group offloading by configuring the `offload_type` parameter to `block_level` or `leaf_level`.
 
 - `block_level` offloads groups of layers based on the `num_blocks_per_group` parameter. For example, if `num_blocks_per_group=2` on a model with 40 layers, 2 layers are onloaded and offloaded at a time (20 total onloads/offloads). This drastically reduces memory requirements.
 - `leaf_level` offloads individual layers at the lowest level and is equivalent to [CPU offloading](#cpu-offloading). But it can be made faster if you use streams without giving up inference speed.
 
+Group offloading is supported for entire pipelines or individual models. Applying group offloading to the entire pipeline is the easiest option while selectively applying it to individual models gives users more flexibility to use different offloading techniques for different models.
+
+<hfoptions id="group-offloading">
+<hfoption id="pipeline">
+
+Call [`~DiffusionPipeline.enable_group_offload`] on a pipeline.
+
+```py
+import torch
+from diffusers import CogVideoXPipeline
+from diffusers.hooks import apply_group_offloading
+from diffusers.utils import export_to_video
+
+onload_device = torch.device("cuda")
+offload_device = torch.device("cpu")
+
+pipeline = CogVideoXPipeline.from_pretrained("THUDM/CogVideoX-5b", torch_dtype=torch.bfloat16)
+pipeline.enable_group_offload(
+    onload_device=onload_device,
+    offload_device=offload_device,
+    offload_type="leaf_level",
+    use_stream=True
+)
+
+prompt = (
+    "A panda, dressed in a small, red jacket and a tiny hat, sits on a wooden stool in a serene bamboo forest. "
+    "The panda's fluffy paws strum a miniature acoustic guitar, producing soft, melodic tunes. Nearby, a few other "
+    "pandas gather, watching curiously and some clapping in rhythm. Sunlight filters through the tall bamboo, "
+    "casting a gentle glow on the scene. The panda's face is expressive, showing concentration and joy as it plays. "
+    "The background includes a small, flowing stream and vibrant green foliage, enhancing the peaceful and magical "
+    "atmosphere of this unique musical performance."
+)
+video = pipeline(prompt=prompt, guidance_scale=6, num_inference_steps=50).frames[0]
+print(f"Max memory reserved: {torch.cuda.max_memory_allocated() / 1024**3:.2f} GB")
+export_to_video(video, "output.mp4", fps=8)
+```
+
+</hfoption>
+<hfoption id="model">
+
+Call [`~ModelMixin.enable_group_offload`] on standard Diffusers model components that inherit from [`ModelMixin`]. For other model components that don't inherit from [`ModelMixin`], such as a generic [torch.nn.Module](https://pytorch.org/docs/stable/generated/torch.nn.Module.html), use [`~hooks.apply_group_offloading`] instead.
+
 ```py
 import torch
 from diffusers import CogVideoXPipeline
@@ -328,6 +368,9 @@ print(f"Max memory reserved: {torch.cuda.max_memory_allocated() / 1024**3:.2f} G
 export_to_video(video, "output.mp4", fps=8)
 ```
 
+</hfoption>
+</hfoptions>
+
 #### CUDA stream
 
 The `use_stream` parameter can be activated for CUDA devices that support asynchronous data transfer streams to reduce overall execution time compared to [CPU offloading](#cpu-offloading). It overlaps data transfer and computation by using layer prefetching. The next layer to be executed is loaded onto the GPU while the current layer is still being executed. It can increase CPU memory significantly so ensure you have 2x the amount of memory as the model size.

docs/source/en/quantization/overview.md

@@ -34,7 +34,9 @@ Initialize [`~quantizers.PipelineQuantizationConfig`] with the following paramet
 > [!TIP]
 > These `quant_kwargs` arguments are different for each backend. Refer to the [Quantization API](../api/quantization) docs to view the arguments for each backend.
 
-- `components_to_quantize` specifies which components of the pipeline to quantize. Typically, you should quantize the most compute intensive components like the transformer. The text encoder is another component to consider quantizing if a pipeline has more than one such as [`FluxPipeline`]. The example below quantizes the T5 text encoder in [`FluxPipeline`] while keeping the CLIP model intact.
+- `components_to_quantize` specifies which component(s) of the pipeline to quantize. Typically, you should quantize the most compute intensive components like the transformer. The text encoder is another component to consider quantizing if a pipeline has more than one such as [`FluxPipeline`]. The example below quantizes the T5 text encoder in [`FluxPipeline`] while keeping the CLIP model intact.
+
+   `components_to_quantize` accepts either a list for multiple models or a string for a single model.
 
 The example below loads the bitsandbytes backend with the following arguments from [`~quantizers.quantization_config.BitsAndBytesConfig`], `load_in_4bit`, `bnb_4bit_quant_type`, and `bnb_4bit_compute_dtype`.
 
@@ -62,6 +64,7 @@ pipe = DiffusionPipeline.from_pretrained(
 image = pipe("photo of a cute dog").images[0]
 ```
 
+
 ### Advanced quantization
 
 The `quant_mapping` argument provides more options for how to quantize each individual component in a pipeline, like combining different quantization backends.

docs/source/en/using-diffusers/text-img2vid.md

@@ -98,7 +98,7 @@ pipeline_quant_config = PipelineQuantizationConfig(
     "bnb_4bit_quant_type": "nf4",
     "bnb_4bit_compute_dtype": torch.bfloat16
     },
-  components_to_quantize=["transformer"]
+  components_to_quantize="transformer"
 )
 
 pipeline = HunyuanVideoPipeline.from_pretrained(

src/diffusers/loaders/single_file_model.py

@@ -22,6 +22,7 @@
 from typing_extensions import Self
 
 from .. import __version__
+from ..models.model_loading_utils import _caching_allocator_warmup, _determine_device_map, _expand_device_map
 from ..quantizers import DiffusersAutoQuantizer
 from ..utils import deprecate, is_accelerate_available, is_torch_version, logging
 from ..utils.torch_utils import empty_device_cache
@@ -297,6 +298,7 @@ def from_single_file(cls, pretrained_model_link_or_path_or_dict: Optional[str] =
         low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT)
         device = kwargs.pop("device", None)
         disable_mmap = kwargs.pop("disable_mmap", False)
+        device_map = kwargs.pop("device_map", None)
 
         user_agent = {"diffusers": __version__, "file_type": "single_file", "framework": "pytorch"}
         # In order to ensure popular quantization methods are supported. Can be disable with `disable_telemetry`
@@ -403,19 +405,8 @@ def from_single_file(cls, pretrained_model_link_or_path_or_dict: Optional[str] =
         with ctx():
             model = cls.from_config(diffusers_model_config)
 
-        checkpoint_mapping_kwargs = _get_mapping_function_kwargs(checkpoint_mapping_fn, **kwargs)
-
-        if _should_convert_state_dict_to_diffusers(model.state_dict(), checkpoint):
-            diffusers_format_checkpoint = checkpoint_mapping_fn(
-                config=diffusers_model_config, checkpoint=checkpoint, **checkpoint_mapping_kwargs
-            )
-        else:
-            diffusers_format_checkpoint = checkpoint
+        model_state_dict = model.state_dict()
 
-        if not diffusers_format_checkpoint:
-            raise SingleFileComponentError(
-                f"Failed to load {mapping_class_name}. Weights for this component appear to be missing in the checkpoint."
-            )
         # Check if `_keep_in_fp32_modules` is not None
         use_keep_in_fp32_modules = (cls._keep_in_fp32_modules is not None) and (
             (torch_dtype == torch.float16) or hasattr(hf_quantizer, "use_keep_in_fp32_modules")
@@ -428,6 +419,26 @@ def from_single_file(cls, pretrained_model_link_or_path_or_dict: Optional[str] =
         else:
             keep_in_fp32_modules = []
 
+        # Now that the model is loaded, we can determine the `device_map`
+        device_map = _determine_device_map(model, device_map, None, torch_dtype, keep_in_fp32_modules, hf_quantizer)
+        if device_map is not None:
+            expanded_device_map = _expand_device_map(device_map, model_state_dict.keys())
+            _caching_allocator_warmup(model, expanded_device_map, torch_dtype, hf_quantizer)
+
+        checkpoint_mapping_kwargs = _get_mapping_function_kwargs(checkpoint_mapping_fn, **kwargs)
+
+        if _should_convert_state_dict_to_diffusers(model_state_dict, checkpoint):
+            diffusers_format_checkpoint = checkpoint_mapping_fn(
+                config=diffusers_model_config, checkpoint=checkpoint, **checkpoint_mapping_kwargs
+            )
+        else:
+            diffusers_format_checkpoint = checkpoint
+
+        if not diffusers_format_checkpoint:
+            raise SingleFileComponentError(
+                f"Failed to load {mapping_class_name}. Weights for this component appear to be missing in the checkpoint."
+            )
+
         if hf_quantizer is not None:
             hf_quantizer.preprocess_model(
                 model=model,

src/diffusers/modular_pipelines/flux/before_denoise.py

@@ -454,6 +454,9 @@ def __call__(self, components: FluxModularPipeline, state: PipelineState) -> Pip
         block_state = self.get_block_state(state)
         block_state.device = components._execution_device
 
+        block_state.height = block_state.height or components.default_height
+        block_state.width = block_state.width or components.default_width
+
         scheduler = components.scheduler
         transformer = components.transformer
         batch_size = block_state.batch_size * block_state.num_images_per_prompt
@@ -659,8 +662,6 @@ def intermediate_outputs(self) -> List[OutputParam]:
     def __call__(self, components: FluxModularPipeline, state: PipelineState) -> PipelineState:
         block_state = self.get_block_state(state)
 
-        block_state.height = block_state.height or components.default_height
-        block_state.width = block_state.width or components.default_width
         block_state.device = components._execution_device
         block_state.dtype = torch.bfloat16  # TODO: okay to hardcode this?
         block_state.num_channels_latents = components.num_channels_latents

src/diffusers/modular_pipelines/flux/modular_blocks.py

@@ -148,8 +148,8 @@ def description(self):
     [
         ("text_encoder", FluxTextEncoderStep),
         ("input", FluxInputStep),
-        ("set_timesteps", FluxSetTimestepsStep),
         ("prepare_latents", FluxPrepareLatentsStep),
+        ("set_timesteps", FluxSetTimestepsStep),
         ("denoise", FluxDenoiseStep),
         ("decode", FluxDecodeStep),
     ]

src/diffusers/pipelines/pipeline_utils.py

@@ -1335,6 +1335,133 @@ def enable_sequential_cpu_offload(self, gpu_id: Optional[int] = None, device: Un
                 offload_buffers = len(model._parameters) > 0
                 cpu_offload(model, device, offload_buffers=offload_buffers)
 
+    def enable_group_offload(
+        self,
+        onload_device: torch.device,
+        offload_device: torch.device = torch.device("cpu"),
+        offload_type: str = "block_level",
+        num_blocks_per_group: Optional[int] = None,
+        non_blocking: bool = False,
+        use_stream: bool = False,
+        record_stream: bool = False,
+        low_cpu_mem_usage=False,
+        offload_to_disk_path: Optional[str] = None,
+        exclude_modules: Optional[Union[str, List[str]]] = None,
+    ) -> None:
+        r"""
+        Applies group offloading to the internal layers of a torch.nn.Module. To understand what group offloading is,
+        and where it is beneficial, we need to first provide some context on how other supported offloading methods
+        work.
+
+        Typically, offloading is done at two levels:
+        - Module-level: In Diffusers, this can be enabled using the `ModelMixin::enable_model_cpu_offload()` method. It
+        works by offloading each component of a pipeline to the CPU for storage, and onloading to the accelerator
+        device when needed for computation. This method is more memory-efficient than keeping all components on the
+        accelerator, but the memory requirements are still quite high. For this method to work, one needs memory
+        equivalent to size of the model in runtime dtype + size of largest intermediate activation tensors to be able
+        to complete the forward pass.
+        - Leaf-level: In Diffusers, this can be enabled using the `ModelMixin::enable_sequential_cpu_offload()` method.
+          It
+        works by offloading the lowest leaf-level parameters of the computation graph to the CPU for storage, and
+        onloading only the leafs to the accelerator device for computation. This uses the lowest amount of accelerator
+        memory, but can be slower due to the excessive number of device synchronizations.
+
+        Group offloading is a middle ground between the two methods. It works by offloading groups of internal layers,
+        (either `torch.nn.ModuleList` or `torch.nn.Sequential`). This method uses lower memory than module-level
+        offloading. It is also faster than leaf-level/sequential offloading, as the number of device synchronizations
+        is reduced.
+
+        Another supported feature (for CUDA devices with support for asynchronous data transfer streams) is the ability
+        to overlap data transfer and computation to reduce the overall execution time compared to sequential
+        offloading. This is enabled using layer prefetching with streams, i.e., the layer that is to be executed next
+        starts onloading to the accelerator device while the current layer is being executed - this increases the
+        memory requirements slightly. Note that this implementation also supports leaf-level offloading but can be made
+        much faster when using streams.
+
+        Args:
+            onload_device (`torch.device`):
+                The device to which the group of modules are onloaded.
+            offload_device (`torch.device`, defaults to `torch.device("cpu")`):
+                The device to which the group of modules are offloaded. This should typically be the CPU. Default is
+                CPU.
+            offload_type (`str` or `GroupOffloadingType`, defaults to "block_level"):
+                The type of offloading to be applied. Can be one of "block_level" or "leaf_level". Default is
+                "block_level".
+            offload_to_disk_path (`str`, *optional*, defaults to `None`):
+                The path to the directory where parameters will be offloaded. Setting this option can be useful in
+                limited RAM environment settings where a reasonable speed-memory trade-off is desired.
+            num_blocks_per_group (`int`, *optional*):
+                The number of blocks per group when using offload_type="block_level". This is required when using
+                offload_type="block_level".
+            non_blocking (`bool`, defaults to `False`):
+                If True, offloading and onloading is done with non-blocking data transfer.
+            use_stream (`bool`, defaults to `False`):
+                If True, offloading and onloading is done asynchronously using a CUDA stream. This can be useful for
+                overlapping computation and data transfer.
+            record_stream (`bool`, defaults to `False`): When enabled with `use_stream`, it marks the current tensor
+                as having been used by this stream. It is faster at the expense of slightly more memory usage. Refer to
+                the [PyTorch official docs](https://pytorch.org/docs/stable/generated/torch.Tensor.record_stream.html)
+                more details.
+            low_cpu_mem_usage (`bool`, defaults to `False`):
+                If True, the CPU memory usage is minimized by pinning tensors on-the-fly instead of pre-pinning them.
+                This option only matters when using streamed CPU offloading (i.e. `use_stream=True`). This can be
+                useful when the CPU memory is a bottleneck but may counteract the benefits of using streams.
+            exclude_modules (`Union[str, List[str]]`, defaults to `None`): List of modules to exclude from offloading.
+
+        Example:
+            ```python
+            >>> from diffusers import DiffusionPipeline
+            >>> import torch
+
+            >>> pipe = DiffusionPipeline.from_pretrained("Qwen/Qwen-Image", torch_dtype=torch.bfloat16)
+
+            >>> pipe.enable_group_offload(
+            ...     onload_device=torch.device("cuda"),
+            ...     offload_device=torch.device("cpu"),
+            ...     offload_type="leaf_level",
+            ...     use_stream=True,
+            ... )
+            >>> image = pipe("a beautiful sunset").images[0]
+            ```
+        """
+        from ..hooks import apply_group_offloading
+
+        if isinstance(exclude_modules, str):
+            exclude_modules = [exclude_modules]
+        elif exclude_modules is None:
+            exclude_modules = []
+
+        unknown = set(exclude_modules) - self.components.keys()
+        if unknown:
+            logger.info(
+                f"The following modules are not present in pipeline: {', '.join(unknown)}. Ignore if this is expected."
+            )
+
+        group_offload_kwargs = {
+            "onload_device": onload_device,
+            "offload_device": offload_device,
+            "offload_type": offload_type,
+            "num_blocks_per_group": num_blocks_per_group,
+            "non_blocking": non_blocking,
+            "use_stream": use_stream,
+            "record_stream": record_stream,
+            "low_cpu_mem_usage": low_cpu_mem_usage,
+            "offload_to_disk_path": offload_to_disk_path,
+        }
+        for name, component in self.components.items():
+            if name not in exclude_modules and isinstance(component, torch.nn.Module):
+                if hasattr(component, "enable_group_offload"):
+                    component.enable_group_offload(**group_offload_kwargs)
+                else:
+                    apply_group_offloading(module=component, **group_offload_kwargs)
+
+        if exclude_modules:
+            for module_name in exclude_modules:
+                module = getattr(self, module_name, None)
+                if module is not None and isinstance(module, torch.nn.Module):
+                    module.to(onload_device)
+                    logger.debug(f"Placed `{module_name}` on {onload_device} device as it was in `exclude_modules`.")
+
     def reset_device_map(self):
         r"""
         Resets the device maps (if any) to None.