train.py

# train.py

import os
import torch
from huggingface_hub import snapshot_download
from diffusers import StableDiffusionPipeline, DPMSolverMultistepScheduler
from peft import LoraConfig, get_peft_model

# ── 1) Configuration ───────────────────────────────────────────────────────────

# Where you put your images + prompts
DATA_DIR   = os.getenv("DATA_DIR", "./data")

# Where your base model lives (downloaded or cached)
MODEL_DIR  = os.getenv("MODEL_DIR", "./hidream-model")

# Where to save your LoRA‑fine‑tuned model
OUTPUT_DIR = os.getenv("OUTPUT_DIR", "./lora-trained")

# ── 2) Prepare the base model snapshot ────────────────────────────────────────

print(f"📂 Loading dataset from: {DATA_DIR}")
print("📥 Fetching or verifying base model: HiDream-ai/HiDream-I1-Dev")

# If you’ve pre‑downloaded into MODEL_DIR, just use it; otherwise pull from HF Hub
if not os.path.isdir(MODEL_DIR):
    MODEL_DIR = snapshot_download(
        repo_id="HiDream-ai/HiDream-I1-Dev",
        local_dir=MODEL_DIR
    )

# ── 3) Load the scheduler manually ─────────────────────────────────────────────

# Diffusers’ scheduler config JSON points at FlowMatchLCMScheduler,
# but your installed version doesn’t have that class. Instead we
# force‐load DPMSolverMultistepScheduler via `from_pretrained`.
print(f"🔄 Loading scheduler from: {MODEL_DIR}/scheduler")
scheduler = DPMSolverMultistepScheduler.from_pretrained(
    pretrained_model_name_or_path=MODEL_DIR,
    subfolder="scheduler"
)

# ── 4) Build the Stable Diffusion pipeline ────────────────────────────────────

print("🔧 Creating StableDiffusionPipeline with custom scheduler")
pipe = StableDiffusionPipeline.from_pretrained(
    pretrained_model_name_or_path=MODEL_DIR,
    scheduler=scheduler,
    torch_dtype=torch.float16,
).to("cuda")

# ── 5) Apply PEFT LoRA adapters ───────────────────────────────────────────────

print("🧠 Configuring LoRA adapter on U‑Net")
lora_config = LoraConfig(
    r=16,
    lora_alpha=16,
    bias="none",
    task_type="CAUSAL_LM"
)
pipe.unet = get_peft_model(pipe.unet, lora_config)

# ── 6) (Placeholder) Simulate your training loop ─────────────────────────────

print("🚀 Starting fine‑tuning loop (simulated)")
for step in range(100):
    # Here you'd load your data, compute loss, do optimizer.step(), etc.
    print(f" Training step {step+1}/100")

# ── 7) Save your LoRA‑tuned model ────────────────────────────────────────────

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