[docs] CP (#12331)

* init

* feedback

* feedback

* feedback

* feedback

* feedback

* feedback

docs/source/en/_toctree.yml

@@ -70,8 +70,6 @@
     title: Reduce memory usage
   - local: optimization/speed-memory-optims
     title: Compiling and offloading quantized models
-  - local: api/parallel
-    title: Parallel inference
   - title: Community optimizations
     sections:
     - local: optimization/pruna
@@ -282,6 +280,8 @@
       title: Outputs
     - local: api/quantization
       title: Quantization
+    - local: api/parallel
+      title: Parallel inference
   - title: Modular
     sections:
     - local: api/modular_diffusers/pipeline

docs/source/en/api/parallel.md

@@ -11,7 +11,7 @@ specific language governing permissions and limitations under the License. -->
 
 # Parallelism
 
-Parallelism strategies help speed up diffusion transformers by distributing computations across multiple devices, allowing for faster inference/training times.
+Parallelism strategies help speed up diffusion transformers by distributing computations across multiple devices, allowing for faster inference/training times. Refer to the [Distributed inferece](../training/distributed_inference) guide to learn more.
 
 ## ParallelConfig
 

docs/source/en/training/distributed_inference.md

@@ -226,8 +226,64 @@ with torch.no_grad():
     image[0].save("split_transformer.png")
 ```
 
-## Resources
+By selectively loading and unloading the models you need at a given stage and sharding the largest models across multiple GPUs, it is possible to run inference with large models on consumer GPUs.
 
-- Take a look at this [script](https://gist.github.com/sayakpaul/cfaebd221820d7b43fae638b4dfa01ba) for a minimal example of distributed inference with Accelerate.
-- For more details, check out Accelerate's [Distributed inference](https://huggingface.co/docs/accelerate/en/usage_guides/distributed_inference#distributed-inference-with-accelerate) guide.
-- The `device_map` argument assign models or an entire pipeline to devices. Refer to the [device placement](../using-diffusers/loading#device-placement) docs for more information.
+## Context parallelism
+
+[Context parallelism](https://huggingface.co/spaces/nanotron/ultrascale-playbook?section=context_parallelism) splits input sequences across multiple GPUs to reduce memory usage. Each GPU processes its own slice of the sequence.
\ No newline at end of file
+
+Use [`~ModelMixin.set_attention_backend`] to switch to a more optimized attention backend. Refer to this [table](../optimization/attention_backends#available-backends) for a complete list of available backends.
+
+### Ring Attention
+
+Key (K) and value (V) representations communicate between devices using [Ring Attention](https://huggingface.co/papers/2310.01889). This ensures each split sees every other token's K/V. Each GPU computes attention for its local K/V and passes it to the next GPU in the ring. No single GPU holds the full sequence, which reduces communication latency.
+
+Pass a [`ContextParallelConfig`] to the `parallel_config` argument of the transformer model. The config supports the `ring_degree` argument that determines how many devices to use for Ring Attention.
+
+```py
+import torch
+from diffusers import AutoModel, QwenImagePipeline, ContextParallelConfig
+
+try:
+    torch.distributed.init_process_group("nccl")
+    rank = torch.distributed.get_rank()
+    device = torch.device("cuda", rank % torch.cuda.device_count())
+    torch.cuda.set_device(device)
+    
+    transformer = AutoModel.from_pretrained("Qwen/Qwen-Image", subfolder="transformer", torch_dtype=torch.bfloat16, parallel_config=ContextParallelConfig(ring_degree=2))
+    pipeline = QwenImagePipeline.from_pretrained("Qwen/Qwen-Image", transformer=transformer, torch_dtype=torch.bfloat16, device_map="cuda")
+    pipeline.transformer.set_attention_backend("flash")
+
+    prompt = """
+    cinematic film still of a cat sipping a margarita in a pool in Palm Springs, California
+    highly detailed, high budget hollywood movie, cinemascope, moody, epic, gorgeous, film grain
+    """
+    
+    # Must specify generator so all ranks start with same latents (or pass your own)
+    generator = torch.Generator().manual_seed(42)
+    image = pipeline(prompt, num_inference_steps=50, generator=generator).images[0]
+    
+    if rank == 0:
+        image.save("output.png")
+
+except Exception as e:
+    print(f"An error occurred: {e}")
+    torch.distributed.breakpoint()
+    raise
+
+finally:
+    if torch.distributed.is_initialized():
+        torch.distributed.destroy_process_group()
+```
+
+### Ulysses Attention
+
+[Ulysses Attention](https://huggingface.co/papers/2309.14509) splits a sequence across GPUs and performs an *all-to-all* communication (every device sends/receives data to every other device). Each GPU ends up with all tokens for only a subset of attention heads. Each GPU computes attention locally on all tokens for its head, then performs another all-to-all to regroup results by tokens for the next layer.
+
+[`ContextParallelConfig`] supports Ulysses Attention through the `ulysses_degree` argument. This determines how many devices to use for Ulysses Attention.
+
+Pass the [`ContextParallelConfig`] to [`~ModelMixin.enable_parallelism`].
+
+```py
+pipeline.transformer.enable_parallelism(config=ContextParallelConfig(ulysses_degree=2))
+```
\ No newline at end of file