app.py

import os
os.environ["HF_HOME"] = "/data/.cache"
os.environ["CUDA_VISIBLE_DEVICES"] = "" # force CPU

from flask import Flask, request, send_file
from huggingface_hub import snapshot_download, login
from diffusers import DiffusionPipeline
import torch
from diffusers.utils import export_to_video
import subprocess

app = Flask(__name__)

MODEL_ID = "damo-vilab/text-to-video-ms-1.7b"
MODEL_DIR = "/data/models/t2v-1.7b"

if os.getenv("HF_TOKEN"):
    login(token=os.getenv("HF_TOKEN"))

if not os.path.exists(MODEL_DIR):
    print("Downloading to /data...")
    snapshot_download(MODEL_ID, local_dir=MODEL_DIR, local_dir_use_symlinks=False)

# PURE CPU — no offload calls
pipe = DiffusionPipeline.from_pretrained(
    MODEL_DIR,
    torch_dtype=torch.float32,
    local_files_only=True
)
pipe = pipe.to("cpu")
pipe.vae.enable_slicing()
pipe.vae.enable_tiling()

@app.route("/generate", methods=["POST"])
def generate():
    prompt = request.json.get("prompt", "cartoon chaiwala at station")

    # 16 frames -> repeat to 48 for 6 sec
    result = pipe(
        prompt,
        num_inference_steps=20,
        guidance_scale=7.5,
        num_frames=16,
        height=256,
        width=256
    )
    frames = result.frames[0]
    frames = (frames * 3)[:48]

    tmp = "/tmp/low.mp4"
    export_to_video(frames, tmp, fps=8)

    out = "/tmp/out_720p.mp4"
    # upscale to 720x1280 9:16
    subprocess.run([
        "ffmpeg","-y","-i",tmp,
        "-vf","scale=720:1280:force_original_aspect_ratio=increase,crop=720:1280",
        "-c:v","libx264","-preset","ultrafast","-crf","28",
        out
    ], check=True, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)

    return send_file(out, mimetype="video/mp4")

if __name__ == "__main__":
    app.run(host="0.0.0.0", port=7860)