Update README.md

ControlNeXt-SVD-v2-Training/README.md

@@ -2,9 +2,9 @@
 
 # Important
 
-I found that sometimes, when I change the version of dependencies, the training may not converge at all. I haven't identified the reason yet, but I've listed all our dependencies in the [requirements.txt](./requirements.txt) file. It's a bit detailed, but you can focus on the key dependencies like `torch`, `deepspeed`, `diffusers`, `accelerate`... When issues arise, checking these first may help.
+I found that sometimes, when I change the version of dependencies, the training may not converge at all. I haven't identified the reason yet, but I've listed all our dependencies in the [requirements.txt](./requirements.txt) file. It's a bit detailed, but you can focus on the key dependencies like `torch`, `deepspeed`, `diffusers`, `accelerate`... When issues arise, checking these first may help. (We use: `diffusers==0.25.0`)
 
-> The training scripts are still under testing after my reconstruction. Since there has been significant demand, I'm releasing them now. Please keep an eye out for updates.
+If you find the differences for the training and inference scripts, such as the `import path`, please refer to the training script!
 
 ## Main
 
@@ -81,6 +81,16 @@ And then replace the `path` to:
 --validation_image=$PATH_TO_THE_REFERENCE_IMAGE_FILE_FOR_VALIDATION \
 ```
 
+## DeepSpeed
+
+When using `DeepSpeed` for training, you’ll notice the use of a `DeepSpeedWrapperModel` in [training script](train_svd.py#L837). This wrapper is necessary because DeepSpeed supports only a single model for parallel training, allowing us to encapsulate different modules, including ControlNet and UNet.
+
+To perform inference with your trained weights, follow these steps:
+
+1. Convert the generated weights to a `.bin` file using `zero_to_fp32.py` generated by DeepSpeed (Under the generated weight directory). This will create a file named `pytorch_model.bin`.
+2. Utilize the script [unwrap_deepspeed.py](utils/train_svd.py) to separate the modules into distinct dictionaries for ControlNet and UNet.
+3. Provide the paths to these weights in your inference script.
+
 ## Meta info
 
 Please construct the training dataset and provide a list of the data entries in a .json file. We give an example in `meta_info_example/meta_info.json` (the data list) and `meta_info_example/meta_info/1.json`(Detailed meta information for each single video recoarding the position and score):
@@ -115,4 +125,4 @@ It requires substantial memory for training, as we use a high resolution and lon
   journal={arXiv preprint arXiv:2408.06070},
   year={2024}
 }
-```
+```