memory-studio / app.py
prasadpatil25's picture
Deploy Memory Studio
c9729ae verified
Raw History Blame Contribute Delete
1.88 kB
"""Hugging Face Spaces entry point for the public Memory Studio demo.
Locally, run ``memory_studio.py`` instead: it binds to ``127.0.0.1``, creates no
sharing link, and keeps inference on your machine. This file exists only so a
Gradio Space can start the same UI on the Space's public port.
Deploying this way is a **public demo**: prompts and image cues are sent to the
Space and processed there, so the on-device privacy promise does not apply.
"""
from __future__ import annotations
import os
# Spaces sets SPACE_ID, but set the flag explicitly so the public wording also
# applies when this entry point is tested in a container.
os.environ.setdefault("MEMORY_PUBLIC_DEMO", "1")
# The Space has a free GPU, so it uses a larger image model than the local
# default. Both are public downloads and need no token.
os.environ.setdefault("MEMORY_IMAGE_MODEL", "segmind/SSD-1B")
# The larger text model follows the story-line format far better than the
# 360M local default, and the Space's free GPU runs it in seconds.
os.environ.setdefault("MEMORY_MODEL", "HuggingFaceTB/SmolLM2-1.7B-Instruct")
# When a visitor's free GPU quota runs out, drafts fall back to the small model
# on CPU, which needs no GPU quota.
os.environ.setdefault("MEMORY_FALLBACK_MODEL", "HuggingFaceTB/SmolLM2-360M-Instruct")
os.environ.setdefault("MEMORY_IMAGE_DEFAULT_SIZE", "768")
os.environ.setdefault("MEMORY_IMAGE_DEFAULT_STEPS", "8")
from memory_studio import ( # noqa: E402 (env first)
APP_CSS,
APP_HEAD,
APP_THEME,
QUEUE_CONCURRENCY,
demo,
)
if __name__ == "__main__":
demo.queue(default_concurrency_limit=QUEUE_CONCURRENCY).launch(
server_name=os.getenv("MEMORY_HOST", "0.0.0.0"),
server_port=int(os.getenv("PORT", "7860")),
share=False,
show_error=False,
theme=APP_THEME,
css=APP_CSS,
head=APP_HEAD,
)