app.py

import os
import shutil
import subprocess
import gradio as gr

# Use /tmp in Spaces, or ./data locally if you set LOCAL_TRAIN=1
LOCAL = os.environ.get("LOCAL_TRAIN", "").lower() in ("1","true")
DATA_DIR = os.path.join(os.getcwd(), "data") if LOCAL else "/tmp/data"
os.makedirs(DATA_DIR, exist_ok=True)

def prepare_dataset(files):
    # wipe and copy uploaded files
    for f in os.listdir(DATA_DIR):
        os.remove(os.path.join(DATA_DIR, f))
    for file in files:
        dst = os.path.join(DATA_DIR, os.path.basename(file.name))
        shutil.copyfile(file.name, dst)
    return f"✅ {len(files)} files uploaded to {DATA_DIR}"

def start_training(base_model, trigger_word, steps, r, alpha):
    # pass args via environment to train.py
    env = os.environ.copy()
    env.update({
        "BASE_MODEL":    base_model,
        "TRIGGER_WORD":  trigger_word,
        "NUM_STEPS":     str(steps),
        "LORA_R":        str(r),
        "LORA_ALPHA":    str(alpha),
        "LOCAL_TRAIN":   os.environ.get("LOCAL_TRAIN","")
    })
    # run training and capture all output
    proc = subprocess.run(
        ["python3","train.py"],
        capture_output=True, text=True, env=env
    )
    return proc.stdout + ("\n" + proc.stderr if proc.stderr else "")

model_choices = [
    "HiDream-ai/HiDream-I1-Dev",
    "runwayml/stable-diffusion-v1-5",
    "stabilityai/stable-diffusion-2-1"
]

with gr.Blocks() as demo:
    gr.Markdown("# 🖌️ HiDream LoRA Trainer")
    gr.Markdown(f"Running in **{'local' if LOCAL else 'Spaces'}** mode; data dir: `{DATA_DIR}`")

    with gr.Row():
        uploader = gr.File(file_types=["image",".txt"], file_count="multiple", label="Upload images + texts")
        up_btn   = gr.Button("📂 Upload")
    up_status = gr.Textbox(label="Upload status")

    mdl = gr.Dropdown(model_choices, value=model_choices[0], label="Base model")
    tw  = gr.Textbox(label="Trigger word", placeholder="e.g. rami-style")
    st  = gr.Slider(10,500,value=100,step=10,label="Training steps")
    r_v = gr.Slider(4,128,value=16,step=4,label="LoRA rank (r)")
    a_v = gr.Slider(4,128,value=16,step=4,label="LoRA alpha")

    tr_btn = gr.Button("🚀 Train")
    log_tb  = gr.Textbox(label="Training log", lines=20)

    up_btn.click(prepare_dataset, inputs=uploader, outputs=up_status)
    tr_btn.click(start_training, inputs=[mdl,tw,st,r_v,a_v], outputs=log_tb)

demo.launch(server_name="0.0.0.0", server_port=7860)