# Qwen-Image 2.1 LoRA Training Configuration - maltrainer (https://github.com/malcolmamal/maltrainer)
#
# The recommended setting from the Qwen-Image 2.1 optimizer benchmark:
#   qwen/docs/QWEN_IMAGE_21_LORA_OPTIMIZER_BENCHMARK.md
# Automagic3, ~1500 steps, 512px, rank 16 -> about 13-16 minutes on one RTX 5090 (~11 GB VRAM).
#
# Run with:  python train_qwen_image_2_1_lora.py --config qwen_image_2_1_template.yaml
# Sanity-check the stack first:  python scripts/qwen_2_1_stack_probe.py --config qwen_image_2_1_template.yaml

# Model configuration - the ComfyUI single-file checkpoints from Comfy-Org/Qwen-Image-2.1, no diffusers conversion
model:
  transformer_path: "[comfyui_location]/models/diffusion_models/qwen_image_2.1_int8_convrot.safetensors"
  vae_path: "[comfyui_location]/models/vae/qwen_image_2.1_vae_bf16.safetensors"
  text_encoder_path: "[comfyui_location]/models/text_encoders/qwen3vl_8b_int8_convrot.safetensors"

# Dataset configuration
# Caption files (.txt next to each image) are recommended. Write them as "[instance_token], a [class_token] with ..."
# so the trigger binds to the class word. If captions only say "a woman ..." the trigger never learns gender,
# and prompts without the class word (e.g. "[instance_token] as a medieval knight") can come out as a man.
dataset:
  path: "[dataset_location]/[person_to_train]"
  resolution: 512                   # square 512x512 only for now (no cropping or bucketing yet: crop your images to
                                    # square first). More resolutions and aspect-ratio bucketing are planned.
  trigger: "[instance_token]"
  default_caption: "a photo of [instance_token], a [class_token]"
  repeats: 1
  num_workers: 0

# LoRA configuration - all 32 blocks, attention + MLP (192 layers, 39.8M params at rank 16)
lora:
  rank: 16
  alpha: 16
  prefix: "diffusion_model."
  target_modules:
    - "to_q"
    - "to_k"
    - "to_v"
    - "to_out.0"
    - "gate_up"
    - "img_mlp.out"
  dtype: "float32"

# Training configuration
train:
  batch_size: 1
  gradient_accumulation_steps: 1
  max_steps: 1500                   # stops exactly here. Set 0 to train num_epochs instead.
  num_epochs: 70                    # used when max_steps is 0. Aim for 60-80 epochs; Automagic3 tolerates running long.
  save_every: 250
  timestep_distribution: "shift"
  sigmoid_scale: 1.0
  max_grad_norm: 1.0
  optimizer: "automagic3"
  learning_rate: 1.0e-6             # launch point only - Automagic3 adapts it (peaks ~2e-4, then decays to ~1e-5)
  weight_decay: 0.01
  automagic_polarity_history: 8
  automagic_clip_threshold: 1.0
  automagic_beta2: 0.999
  automagic_eps: 1.0e-30
  # Alternative: optimizer "adamw8bit" with learning_rate 1.0e-4, keep checkpoints 1250-1750.
  # Prodigy is not recommended on this model: its LR keeps climbing and it overcooks.

dtype: "bfloat16"
mixed_precision: "bf16"
seed: 42

output:
  path: "[output_location]/[person_to_train]"
  name: "qwen21_[person_to_train]"
  format: "diffusers"               # lora_A / lora_B keys with the diffusion_model. prefix - loads in ComfyUI
  dtype: "bfloat16"
  embed_config: true

logging:
  level: "INFO"