feat: verfication of multi-gpu support for select examples. (huggingface#3126)

* feat: verfication of multi-gpu support for select examples.

* add: multi-gpu training sections to the relvant doc pages.

Author: sayakpaul

docs/source/en/training/controlnet.mdx

@@ -113,6 +113,29 @@ accelerate launch train_controlnet.py \
  --gradient_accumulation_steps=4
 ```
 
+## Training with multiple GPUs
+
+`accelerate` allows for seamless multi-GPU training. Follow the instructions [here](https://huggingface.co/docs/accelerate/basic_tutorials/launch)
+for running distributed training with `accelerate`. Here is an example command:
+
+```bash 
+export MODEL_DIR="runwayml/stable-diffusion-v1-5"
+export OUTPUT_DIR="path to save model"
+
+accelerate launch --mixed_precision="fp16" --multi_gpu train_controlnet.py \
+ --pretrained_model_name_or_path=$MODEL_DIR \
+ --output_dir=$OUTPUT_DIR \
+ --dataset_name=fusing/fill50k \
+ --resolution=512 \
+ --learning_rate=1e-5 \
+ --validation_image "./conditioning_image_1.png" "./conditioning_image_2.png" \
+ --validation_prompt "red circle with blue background" "cyan circle with brown floral background" \
+ --train_batch_size=4 \
+ --mixed_precision="fp16" \
+ --tracker_project_name="controlnet-demo" \
+ --report_to=wandb
+```
+
 ## Example results
 
 #### After 300 steps with batch size 8

docs/source/en/training/instructpix2pix.mdx

@@ -126,6 +126,27 @@ accelerate launch --mixed_precision="fp16" train_instruct_pix2pix.py \
 
  ***Note: In the original paper, the authors observed that even when the model is trained with an image resolution of 256x256, it generalizes well to bigger resolutions such as 512x512. This is likely because of the larger dataset they used during training.***
 
+ ## Training with multiple GPUs
+
+`accelerate` allows for seamless multi-GPU training. Follow the instructions [here](https://huggingface.co/docs/accelerate/basic_tutorials/launch)
+for running distributed training with `accelerate`. Here is an example command:
+
+```bash 
+accelerate launch --mixed_precision="fp16" --multi_gpu train_instruct_pix2pix.py \
+ --pretrained_model_name_or_path=runwayml/stable-diffusion-v1-5 \
+ --dataset_name=sayakpaul/instructpix2pix-1000-samples \
+ --use_ema \
+ --enable_xformers_memory_efficient_attention \
+ --resolution=512 --random_flip \
+ --train_batch_size=4 --gradient_accumulation_steps=4 --gradient_checkpointing \
+ --max_train_steps=15000 \
+ --checkpointing_steps=5000 --checkpoints_total_limit=1 \
+ --learning_rate=5e-05 --lr_warmup_steps=0 \
+ --conditioning_dropout_prob=0.05 \
+ --mixed_precision=fp16 \
+ --seed=42 
+```
+
  ## Inference
 
  Once training is complete, we can perform inference:

docs/source/en/training/text2image.mdx

@@ -106,6 +106,31 @@ accelerate launch train_text_to_image.py \
   --lr_scheduler="constant" --lr_warmup_steps=0 \
   --output_dir=${OUTPUT_DIR}
 ```
+
+#### Training with multiple GPUs
+
+`accelerate` allows for seamless multi-GPU training. Follow the instructions [here](https://huggingface.co/docs/accelerate/basic_tutorials/launch)
+for running distributed training with `accelerate`. Here is an example command:
+
+```bash
+export MODEL_NAME="CompVis/stable-diffusion-v1-4"
+export dataset_name="lambdalabs/pokemon-blip-captions"
+
+accelerate launch --mixed_precision="fp16" --multi_gpu  train_text_to_image.py \
+  --pretrained_model_name_or_path=$MODEL_NAME \
+  --dataset_name=$dataset_name \
+  --use_ema \
+  --resolution=512 --center_crop --random_flip \
+  --train_batch_size=1 \
+  --gradient_accumulation_steps=4 \
+  --gradient_checkpointing \
+  --max_train_steps=15000 \ 
+  --learning_rate=1e-05 \
+  --max_grad_norm=1 \
+  --lr_scheduler="constant" --lr_warmup_steps=0 \
+  --output_dir="sd-pokemon-model" 
+```
+
 </pt>
 <jax>
 With Flax, it's possible to train a Stable Diffusion model faster on TPUs and GPUs thanks to [@duongna211](https://github.com/duongna21). This is very efficient on TPU hardware but works great on GPUs too. The Flax training script doesn't support features like gradient checkpointing or gradient accumulation yet, so you'll need a GPU with at least 30GB of memory or a TPU v3.

docs/source/en/training/unconditional_training.mdx

@@ -122,6 +122,26 @@ accelerate launch train_unconditional.py \
     <img src="https://user-images.githubusercontent.com/26864830/180248200-928953b4-db38-48db-b0c6-8b740fe6786f.png"/>
 </div>
 
+### Training with multiple GPUs
+
+`accelerate` allows for seamless multi-GPU training. Follow the instructions [here](https://huggingface.co/docs/accelerate/basic_tutorials/launch)
+for running distributed training with `accelerate`. Here is an example command:
+
+```bash
+accelerate launch --mixed_precision="fp16" --multi_gpu train_unconditional.py \
+  --dataset_name="huggan/pokemon" \
+  --resolution=64 --center_crop --random_flip \
+  --output_dir="ddpm-ema-pokemon-64" \
+  --train_batch_size=16 \
+  --num_epochs=100 \
+  --gradient_accumulation_steps=1 \
+  --use_ema \
+  --learning_rate=1e-4 \
+  --lr_warmup_steps=500 \
+  --mixed_precision="fp16" \
+  --logger="wandb"
+```
+
 ## Finetuning with your own data
 
 There are two ways to finetune a model on your own dataset:

examples/controlnet/README.md

@@ -96,6 +96,29 @@ accelerate launch train_controlnet.py \
  --gradient_accumulation_steps=4
 ```
 
+## Training with multiple GPUs
+
+`accelerate` allows for seamless multi-GPU training. Follow the instructions [here](https://huggingface.co/docs/accelerate/basic_tutorials/launch)
+for running distributed training with `accelerate`. Here is an example command:
+
+```bash 
+export MODEL_DIR="runwayml/stable-diffusion-v1-5"
+export OUTPUT_DIR="path to save model"
+
+accelerate launch --mixed_precision="fp16" --multi_gpu train_controlnet.py \
+ --pretrained_model_name_or_path=$MODEL_DIR \
+ --output_dir=$OUTPUT_DIR \
+ --dataset_name=fusing/fill50k \
+ --resolution=512 \
+ --learning_rate=1e-5 \
+ --validation_image "./conditioning_image_1.png" "./conditioning_image_2.png" \
+ --validation_prompt "red circle with blue background" "cyan circle with brown floral background" \
+ --train_batch_size=4 \
+ --mixed_precision="fp16" \
+ --tracker_project_name="controlnet-demo" \
+ --report_to=wandb
+```
+
 ## Example results
 
 #### After 300 steps with batch size 8

examples/instruct_pix2pix/README.md

@@ -113,6 +113,27 @@ accelerate launch --mixed_precision="fp16" train_instruct_pix2pix.py \
 
  ***Note: In the original paper, the authors observed that even when the model is trained with an image resolution of 256x256, it generalizes well to bigger resolutions such as 512x512. This is likely because of the larger dataset they used during training.***
 
+ ## Training with multiple GPUs
+
+`accelerate` allows for seamless multi-GPU training. Follow the instructions [here](https://huggingface.co/docs/accelerate/basic_tutorials/launch)
+for running distributed training with `accelerate`. Here is an example command:
+
+```bash 
+accelerate launch --mixed_precision="fp16" --multi_gpu train_instruct_pix2pix.py \
+ --pretrained_model_name_or_path=runwayml/stable-diffusion-v1-5 \
+ --dataset_name=sayakpaul/instructpix2pix-1000-samples \
+ --use_ema \
+ --enable_xformers_memory_efficient_attention \
+ --resolution=512 --random_flip \
+ --train_batch_size=4 --gradient_accumulation_steps=4 --gradient_checkpointing \
+ --max_train_steps=15000 \
+ --checkpointing_steps=5000 --checkpoints_total_limit=1 \
+ --learning_rate=5e-05 --lr_warmup_steps=0 \
+ --conditioning_dropout_prob=0.05 \
+ --mixed_precision=fp16 \
+ --seed=42 
+```
+
  ## Inference
 
  Once training is complete, we can perform inference:

examples/text_to_image/README.md

@@ -111,6 +111,31 @@ image = pipe(prompt="yoda").images[0]
 image.save("yoda-pokemon.png")
 ```
 
+#### Training with multiple GPUs
+
+`accelerate` allows for seamless multi-GPU training. Follow the instructions [here](https://huggingface.co/docs/accelerate/basic_tutorials/launch)
+for running distributed training with `accelerate`. Here is an example command:
+
+```bash
+export MODEL_NAME="CompVis/stable-diffusion-v1-4"
+export dataset_name="lambdalabs/pokemon-blip-captions"
+
+accelerate launch --mixed_precision="fp16" --multi_gpu  train_text_to_image.py \
+  --pretrained_model_name_or_path=$MODEL_NAME \
+  --dataset_name=$dataset_name \
+  --use_ema \
+  --resolution=512 --center_crop --random_flip \
+  --train_batch_size=1 \
+  --gradient_accumulation_steps=4 \
+  --gradient_checkpointing \
+  --max_train_steps=15000 \ 
+  --learning_rate=1e-05 \
+  --max_grad_norm=1 \
+  --lr_scheduler="constant" --lr_warmup_steps=0 \
+  --output_dir="sd-pokemon-model" 
+```
+
+
 #### Training with Min-SNR weighting
 
 We support training with the Min-SNR weighting strategy proposed in [Efficient Diffusion Training via Min-SNR Weighting Strategy](https://arxiv.org/abs/2303.09556) which helps to achieve faster convergence

examples/text_to_image/train_text_to_image.py

@@ -64,8 +64,8 @@ def log_validation(vae, text_encoder, tokenizer, unet, args, accelerator, weight
 
     pipeline = StableDiffusionPipeline.from_pretrained(
         args.pretrained_model_name_or_path,
-        vae=vae,
-        text_encoder=text_encoder,
+        vae=accelerator.unwrap_model(vae),
+        text_encoder=accelerator.unwrap_model(text_encoder),
         tokenizer=tokenizer,
         unet=accelerator.unwrap_model(unet),
         safety_checker=None,

examples/unconditional_image_generation/README.md

@@ -1,4 +1,4 @@
-## Training examples
+## Training an unconditional diffusion model
 
 Creating a training image set is [described in a different document](https://huggingface.co/docs/datasets/image_process#image-datasets).
 
@@ -76,6 +76,27 @@ A full training run takes 2 hours on 4xV100 GPUs.
 
 <img src="https://user-images.githubusercontent.com/26864830/180248200-928953b4-db38-48db-b0c6-8b740fe6786f.png" width="700" />
 
+### Training with multiple GPUs
+
+`accelerate` allows for seamless multi-GPU training. Follow the instructions [here](https://huggingface.co/docs/accelerate/basic_tutorials/launch)
+for running distributed training with `accelerate`. Here is an example command:
+
+```bash
+accelerate launch --mixed_precision="fp16" --multi_gpu train_unconditional.py \
+  --dataset_name="huggan/pokemon" \
+  --resolution=64 --center_crop --random_flip \
+  --output_dir="ddpm-ema-pokemon-64" \
+  --train_batch_size=16 \
+  --num_epochs=100 \
+  --gradient_accumulation_steps=1 \
+  --use_ema \
+  --learning_rate=1e-4 \
+  --lr_warmup_steps=500 \
+  --mixed_precision="fp16" \
+  --logger="wandb"
+```
+
+To be able to use Weights and Biases (`wandb`) as a logger you need to install the library: `pip install wandb`. 
 
 ### Using your own data