train.py

import os
import torch
from huggingface_hub import snapshot_download
from diffusers import (
    StableDiffusionPipeline,
    DPMSolverMultistepScheduler,
    AutoencoderKL,
    UNet2DConditionModel
)
from transformers import CLIPTextModel, CLIPTokenizer
from peft import LoraConfig, get_peft_model

MODEL_ID = "black-forest-labs/FLUX.1-dev"

# download
model_path = snapshot_download(
    MODEL_ID,
    local_dir="./fluxdev-model",
    use_auth_token=True
)

# later loading
pipe = StableDiffusionPipeline.from_pretrained(
    model_path,
    torch_dtype=torch.float16,
    use_auth_token=True
).to("cuda")

# 1) grab the model locally
print("📥 Downloading Flux‑Dev model…")
model_path = snapshot_download(MODEL_ID, local_dir="./fluxdev-model")

# 2) load each piece with its correct subfolder
print("🔄 Loading scheduler…")
scheduler = DPMSolverMultistepScheduler.from_pretrained(
    model_path, subfolder="scheduler"
)

print("🔄 Loading VAE…")
vae = AutoencoderKL.from_pretrained(
    model_path, subfolder="vae", torch_dtype=torch.float16
)

print("🔄 Loading text encoder + tokenizer…")
text_encoder = CLIPTextModel.from_pretrained(
    model_path, subfolder="text_encoder", torch_dtype=torch.float16
)
tokenizer = CLIPTokenizer.from_pretrained(
    model_path, subfolder="tokenizer"
)

print("🔄 Loading U‑Net…")
unet = UNet2DConditionModel.from_pretrained(
    model_path, subfolder="unet", torch_dtype=torch.float16
)

# 3) assemble the pipeline
print("🛠  Assembling pipeline…")
pipe = StableDiffusionPipeline(
    vae=vae,
    text_encoder=text_encoder,
    tokenizer=tokenizer,
    unet=unet,
    scheduler=scheduler
).to("cuda")

# 4) apply LoRA
print("🧠 Applying LoRA…")
lora_config = LoraConfig(r=16, lora_alpha=16, bias="none", task_type="CAUSAL_LM")
pipe.unet = get_peft_model(pipe.unet, lora_config)

# 5) your training loop (or dummy loop for illustration)
print("🚀 Starting fine‑tuning…")
for step in range(100):
    print(f"Training step {step+1}/100")
    # …insert your actual data‑loader and loss/backprop here…

os.makedirs(output_dir, exist_ok=True)
pipe.save_pretrained(output_dir)
print("✅ Done. LoRA weights in", output_dir)