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

# ─── CONFIG ───────────────────────────────────────────────────────────────────

DATA_DIR      = os.getenv("DATA_DIR", "./data")
MODEL_CACHE   = os.getenv("MODEL_DIR", "./hidream-model")
OUTPUT_DIR    = os.getenv("OUTPUT_DIR", "./lora-trained")
REPO_ID       = "HiDream-ai/HiDream-I1-Dev"

# ─── STEP 1: ENSURE YOU HAVE A COMPLETE SNAPSHOT WITH CONFIGS ─────────────────

print(f"📥 Downloading full model snapshot to {MODEL_CACHE}")
MODEL_ROOT = snapshot_download(
    repo_id=REPO_ID,
    local_dir=MODEL_CACHE,
    local_dir_use_symlinks=False,    # force a copy so config.json ends up there
)

# ─── STEP 2: LOAD SCHEDULER ────────────────────────────────────────────────────

print("🔄 Loading scheduler")
scheduler = DPMSolverMultistepScheduler.from_pretrained(
    MODEL_ROOT,
    subfolder="scheduler",
)

# ─── STEP 3: LOAD VAE ──────────────────────────────────────────────────────────

print("🔄 Loading VAE")
vae = AutoencoderKL.from_pretrained(
    MODEL_ROOT,
    subfolder="vae",
    torch_dtype=torch.float16,
).to("cuda")

# ─── STEP 4: LOAD TEXT ENCODER + TOKENIZER ─────────────────────────────────────

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

# ─── STEP 5: LOAD U‑NET ───────────────────────────────────────────────────────

print("🔄 Loading U‑Net")
unet = UNet2DConditionModel.from_pretrained(
    MODEL_ROOT,
    subfolder="unet",
    torch_dtype=torch.float16,
).to("cuda")

# ─── STEP 6: BUILD THE PIPELINE ───────────────────────────────────────────────

print("🌟 Building StableDiffusionPipeline")
pipe = StableDiffusionPipeline(
    vae=vae,
    text_encoder=text_encoder,
    tokenizer=tokenizer,
    unet=unet,
    scheduler=scheduler,
).to("cuda")

# ─── STEP 7: APPLY LORA ADAPTER ───────────────────────────────────────────────

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

# ─── STEP 8: YOUR TRAINING LOOP (SIMULATED) ────────────────────────────────────

print(f"📂 Loading dataset from: {DATA_DIR}")
for step in range(100):
    # ←– here’s where you’d load your images, run forward/backward, optimizer, etc.
    print(f"Training step {step+1}/100")

# ─── STEP 9: SAVE THE FINE‑TUNED LO‑RA WEIGHTS ───────────────────────────────

os.makedirs(OUTPUT_DIR, exist_ok=True)
pipe.save_pretrained(OUTPUT_DIR)
print("✅ Training complete. Saved to", OUTPUT_DIR)