FLUX.1 Dev

# Runs in YOUR Nebius account, on YOUR data — you own the weights and
# you pay for the GPUs. Forge does not run this job; this just starts it.
# Uses Nebius AI Jobs CLI (`nebius ai job create`).
# Fill in these customer-owned values before running:
#   FORGE_NEBIUS_PROJECT_ID: your Nebius project / parent ID.
#   FORGE_TRAIN_PLATFORM/FORGE_TRAIN_PRESET: pick GPU resources available in your project.
#   FORGE_TRAIN_DATASET_URI: point this at your bucket, e.g. s3://my-bucket/train.jsonl.
#   FORGE_TRAIN_OUTPUT_URI: bucket path where trained weights are written.
# Verify the command starts a user-data fine-tune, not a benchmark/probe.
# After completion: verify output artifacts, create the Serverless Endpoint,
# then run endpoint health and one representative sample request.
# Optional training environment:
#   --env-secret WANDB_API_KEY=... Optional W&B API key. Store it in MysteryBox and pass it with --env-secret.
#   --env WANDB_PROJECT=... Optional W&B project name for training progress and sample tracking.
#   --env WANDB_RUN_NAME=... Optional W&B run name, e.g. flux-klein-subject-lora.
#   --env FORGE_TRAIN_TRIGGER_WORD=... Optional FLUX trigger token override. When unset, the trainer infers it from image filenames such as customtoken_001.jpg.
#   --env-secret HF_TOKEN=... Required for gated BFL checkpoints after you accept the model terms on Hugging Face.
export FORGE_NEBIUS_PROJECT_ID="YOUR_PROJECT_ID"
export FORGE_TRAIN_PLATFORM="YOUR_GPU_PLATFORM"
export FORGE_TRAIN_PRESET="YOUR_GPU_PRESET"
export FORGE_TRAIN_JOB_NAME="forge-fine-tune"
export FORGE_TRAIN_DATASET_URI="s3://my-bucket/train.jsonl"
export FORGE_TRAIN_OUTPUT_URI="s3://my-bucket/outputs/"
FORGE_TRAIN_COMMAND='set -eu
mkdir -p /workspace/config /workspace/dataset /workspace/output
# Dataset source: '"$FORGE_TRAIN_DATASET_URI"' (mounted at /workspace/dataset by the Jobs CLI).
# Output destination: '"$FORGE_TRAIN_OUTPUT_URI"' (mounted at /workspace/output by the Jobs CLI).
DATASET_IMAGE_ROOT="/workspace/dataset"
if [ -d /workspace/dataset/target ]; then DATASET_IMAGE_ROOT="/workspace/dataset/target"; fi
sanitize_trigger_word() {
  printf '\''%s'\'' "$1" | tr '\''[:upper:]'\'' '\''[:lower:]'\'' | sed -E '\''s/[^a-z0-9_-]+/-/g; s/^-+|-+$//g'\''
}
TRIGGER_WORD="$(sanitize_trigger_word "${FORGE_TRAIN_TRIGGER_WORD:-}")"
if [ -z "$TRIGGER_WORD" ]; then
  FIRST_IMAGE="$(find "$DATASET_IMAGE_ROOT" -maxdepth 2 -type f \( -iname '\''*.jpg'\'' -o -iname '\''*.jpeg'\'' -o -iname '\''*.png'\'' \) | sort | head -n 1 || true)"
  if [ -n "$FIRST_IMAGE" ]; then
    STEM="$(basename "$FIRST_IMAGE")"
    STEM="${STEM%.*}"
    STEM="$(printf '\''%s'\'' "$STEM" | sed -E '\''s/([_-]?[0-9]+)$//'\'')"
    TRIGGER_WORD="$(sanitize_trigger_word "$STEM")"
  fi
fi
if [ -z "$TRIGGER_WORD" ]; then TRIGGER_WORD="subject"; fi
export FORGE_TRAIN_TRIGGER_WORD="$TRIGGER_WORD"
echo "Using FLUX trigger token: ${FORGE_TRAIN_TRIGGER_WORD}"
for candidate in /app/ai-toolkit /workspace/ai-toolkit /root/ai-toolkit /ai-toolkit /app /workspace; do
  if [ -f "$candidate/run.py" ]; then cd "$candidate"; break; fi
done
test -f run.py
cat > /workspace/config/forge-flux-lora.yaml <<YAML
---
job: extension
config:
  name: "forge_black_forest_labs_flux_1_dev_lora"
  process:
    - type: '\''sd_trainer'\''
      training_folder: "/workspace/output"
      performance_log_every: 50
      device: cuda:0
      trigger_word: "${FORGE_TRAIN_TRIGGER_WORD}"
      network:
        type: "lora"
        linear: 16
        linear_alpha: 16
      save:
        dtype: float16
        save_every: 250
        max_step_saves_to_keep: 4
        push_to_hub: false
      datasets:
        - folder_path: "/workspace/dataset"
          caption_ext: "txt"
          caption_dropout_rate: 0.05
          shuffle_tokens: false
          cache_latents_to_disk: true
          resolution: [512, 768, 1024]
      train:
        batch_size: 1
        steps: 2000
        gradient_accumulation_steps: 1
        train_unet: true
        train_text_encoder: false
        gradient_checkpointing: true
        noise_scheduler: "flowmatch"
        optimizer: "adamw8bit"
        lr: 1e-4
        ema_config:
          use_ema: true
          ema_decay: 0.99
        dtype: bf16
      model:
        name_or_path: "black-forest-labs/FLUX.1-dev"
        is_flux: true
        quantize: true
      sample:
        sampler: "flowmatch"
        sample_every: 250
        width: 1024
        height: 1024
        prompts:
          - "photo of ${FORGE_TRAIN_TRIGGER_WORD} person, studio portrait, natural skin texture"
          - "${FORGE_TRAIN_TRIGGER_WORD} person in everyday clothing, outdoor natural light"
        neg: ""
        seed: 42
        walk_seed: true
        guidance_scale: 4
        sample_steps: 20
meta:
  name: "FLUX.1 dev LoRA"
  version: '\''1.0'\''
YAML
python run.py /workspace/config/forge-flux-lora.yaml'

nebius ai job create \
  --parent-id "$FORGE_NEBIUS_PROJECT_ID" \
  --name "$FORGE_TRAIN_JOB_NAME" \
  --platform "$FORGE_TRAIN_PLATFORM" \
  --preset "$FORGE_TRAIN_PRESET" \
  --image 'docker.io/ostris/aitoolkit@sha256:220d85e443589c6b52521c594a2d9f052d733afe360966d24bb8a5fe853745f7' \
  --volume "$FORGE_TRAIN_DATASET_URI":/workspace/dataset:ro \
  --volume "$FORGE_TRAIN_OUTPUT_URI":/workspace/output:rw \
  --container-command "/bin/sh" \
  --args "-lc \"$FORGE_TRAIN_COMMAND\""