train.py

# train.py

import os
import torch
from huggingface_hub import snapshot_download
from peft import LoraConfig, get_peft_model

# 1️⃣ Pick your scheduler class
from diffusers import (
    StableDiffusionPipeline,
    DPMSolverMultistepScheduler,
    UNet2DConditionModel,
    AutoencoderKL,
)
from transformers import CLIPTextModel, CLIPTokenizer

# ─── 1) CONFIG ────────────────────────────────────────────────────────────────

DATA_DIR   = os.getenv("DATA_DIR", "./data")
MODEL_DIR  = os.getenv("MODEL_DIR", "./hidream-model")
OUTPUT_DIR = os.getenv("OUTPUT_DIR", "./lora-trained")

# ─── 2) DOWNLOAD OR VERIFY BASE MODEL ──────────────────────────────────────────

if not os.path.isdir(MODEL_DIR):
    MODEL_DIR = snapshot_download(
        repo_id="HiDream-ai/HiDream-I1-Dev",
        local_dir=MODEL_DIR
    )

# ─── 3) LOAD EACH PIPELINE COMPONENT ──────────────────────────────────────────

# 3a) Scheduler
scheduler = DPMSolverMultistepScheduler.from_pretrained(
    MODEL_DIR,
    subfolder="scheduler"
)

# 3b) VAE
vae = AutoencoderKL.from_pretrained(
    MODEL_DIR,
    subfolder="vae",
    torch_dtype=torch.float16
).to("cuda")

# 3c) Text encoder + tokenizer
text_encoder = CLIPTextModel.from_pretrained(
    MODEL_DIR,
    subfolder="text_encoder",
    torch_dtype=torch.float16
).to("cuda")
tokenizer    = CLIPTokenizer.from_pretrained(
    MODEL_DIR,
    subfolder="tokenizer"
)

# 3d) U‑Net
unet = UNet2DConditionModel.from_pretrained(
    MODEL_DIR,
    subfolder="unet",
    torch_dtype=torch.float16
).to("cuda")

# ─── 4) BUILD THE PIPELINE ────────────────────────────────────────────────────

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

# ─── 5) APPLY LORA ────────────────────────────────────────────────────────────

lora_config = LoraConfig(
    r=16,
    lora_alpha=16,
    bias="none",
    task_type="CAUSAL_LM",
)
pipe.unet = get_peft_model(pipe.unet, lora_config)

# ─── 6) TRAINING LOOP (SIMULATED) ─────────────────────────────────────────────

print(f"📂 Data at {DATA_DIR}")
for step in range(100):
    # … your real data loading + optimizer here …
    print(f"Training step {step+1}/100")

# ─── 7) SAVE THE FINE‑TUNED LO‑RA ─────────────────────────────────────────────

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