# IMPORTANT: The Wan2.1 14B model is huge. This config should work on 24GB GPUs. It cannot
# support keeping the text encoder on GPU while training with 24GB, so it is only good
# for training on a single prompt, for example a person with a trigger word.
# to train on captions, you need more vran for now.
---
job: extension
config:
  # this name will be the folder and filename name
  name: "my_first_wan21_14b_lora_v1"
  process:
    - type: 'sd_trainer'
      # root folder to save training sessions/samples/weights
      training_folder: "output"
      # uncomment to see performance stats in the terminal every N steps
#      performance_log_every: 1000
      device: cuda:0
      # if a trigger word is specified, it will be added to captions of training data if it does not already exist
      # alternatively, in your captions you can add [trigger] and it will be replaced with the trigger word
      # this is probably needed for 24GB cards when offloading TE to CPU
      trigger_word: "p3r5on"
      network:
        type: "lora"
        linear: 32
        linear_alpha: 32
      save:
        dtype: float16 # precision to save
        save_every: 250 # save every this many steps
        max_step_saves_to_keep: 4 # how many intermittent saves to keep
        push_to_hub: false #change this to True to push your trained model to Hugging Face.
        # You can either set up a HF_TOKEN env variable or you'll be prompted to log-in         
#       hf_repo_id: your-username/your-model-slug
#       hf_private: true #whether the repo is private or public
      datasets:
        # datasets are a folder of images. captions need to be txt files with the same name as the image
        # for instance image2.jpg and image2.txt. Only jpg, jpeg, and png are supported currently
        # images will automatically be resized and bucketed into the resolution specified
        # on windows, escape back slashes with another backslash so
        # "C:\\path\\to\\images\\folder"
        # AI-Toolkit does not currently support video datasets, we will train on 1 frame at a time
        # it works well for characters, but not as well for "actions"
        - folder_path: "/path/to/images/folder"
          caption_ext: "txt"
          caption_dropout_rate: 0.05  # will drop out the caption 5% of time
          shuffle_tokens: false  # shuffle caption order, split by commas
          cache_latents_to_disk: true  # leave this true unless you know what you're doing
          resolution: [ 632 ]  # will be around 480p
      train:
        batch_size: 1
        steps: 2000  # total number of steps to train 500 - 4000 is a good range
        gradient_accumulation: 1
        train_unet: true
        train_text_encoder: false  # probably won't work with wan
        gradient_checkpointing: true  # need the on unless you have a ton of vram
        noise_scheduler: "flowmatch" # for training only
        timestep_type: 'sigmoid'
        optimizer: "adamw8bit"
        lr: 1e-4
        optimizer_params:
          weight_decay: 1e-4
        # uncomment this to skip the pre training sample
#        skip_first_sample: true
        # uncomment to completely disable sampling
#        disable_sampling: true
        # ema will smooth out learning, but could slow it down. Recommended to leave on.
        ema_config:
          use_ema: true
          ema_decay: 0.99
        dtype: bf16
        # required for 24GB cards
        # this will encode your trigger word and use those embeddings for every image in the dataset
        unload_text_encoder: true
      model:
        # huggingface model name or path
        name_or_path: "Wan-AI/Wan2.1-T2V-14B-Diffusers"
        arch: 'wan21'
        # these settings will save as much vram as possible
        quantize: true
        quantize_te: true
        low_vram: true 
      sample:
        sampler: "flowmatch"
        sample_every: 250 # sample every this many steps
        width: 832
        height: 480
        num_frames: 40
        fps: 15
        # samples take a long time. so use them sparingly
        # samples will be animated webp files, if you don't see them animated, open in a browser.
        prompts:
          # you can add [trigger] to the prompts here and it will be replaced with the trigger word
#          - "[trigger] holding a sign that says 'I LOVE PROMPTS!'"\
          - "woman playing the guitar, on stage, singing a song, laser lights, punk rocker"
        neg: ""  
        seed: 42
        walk_seed: true
        guidance_scale: 5
        sample_steps: 30
# you can add any additional meta info here. [name] is replaced with config name at top
meta:
  name: "[name]"
  version: '1.0'