LECO: verify_training_args() crashes on missing sample_every_n_epochs/sample_every_n_steps

[bmaltais]

Summary

train_leco.py and sdxl_train_leco.py's setup_parser() never register --sample_every_n_epochs / --sample_every_n_steps, but verify_training_args() in library/args.py unconditionally accesses args.sample_every_n_epochs and args.sample_every_n_steps:

if args.sample_every_n_epochs is not None and args.sample_every_n_epochs <= 0:
    ...
if args.sample_every_n_steps is not None and args.sample_every_n_steps <= 0:
    ...

Since these two scripts don't call whatever add_*_arguments() helper registers those sample-related flags, argparse.Namespace never gets a sample_every_n_epochs / sample_every_n_steps attribute, and this line raises AttributeError — before training starts.

Steps to reproduce

Run either script with a minimal valid LECO config:

accelerate launch train_leco.py \
  --pretrained_model_name_or_path="model.safetensors" \
  --prompts_file="prompts.toml" \
  --output_dir="output" \
  --output_name="test"

main() calls args_util.verify_training_args(args), which hits:

AttributeError: 'Namespace' object has no attribute 'sample_every_n_epochs'

(Same issue in sdxl_train_leco.py.)

Suggested fix

Either:

1. Register --sample_every_n_epochs and --sample_every_n_steps in LECO's setup_parser() with default=None (even if unused, so verify_training_args()'s generic checks don't fail), or
2. Guard the two checks in verify_training_args() with getattr(args, "sample_every_n_epochs", None) / getattr(args, "sample_every_n_steps", None) instead of direct attribute access, consistent with how getattr(args, "train_inpainting", False) is already handled a few lines above in the same function.

Option 2 is more robust against future scripts that similarly opt out of the sample-image machinery.

Context

Found while adding LECO GUI support in kohya_ss (bmaltais/kohya_ss#3539). Not a GUI-side bug — this is purely in the training script / arg verification, hence filing here rather than downstream.