app.py

"""analyst-buddy Gradio app — thin entrypoint shim (F004).

The implementation lives in the single canonical module
``sql_env.server.app_ui`` so the process has exactly ONE ``Step``/class identity
(no dual-import footgun where ``app`` and ``sql_env.app`` would create two
distinct ``Step`` types — an F006 ``isinstance`` hazard). This file only
re-exports ``build_demo`` (the public surface HF Spaces / ``uv run python app.py``
and the test suite construct) and launches it when run directly.

Run locally: ``uv run python app.py``.
"""

from __future__ import annotations

# ZeroGPU: the `spaces` package MUST be imported before ANYTHING initializes CUDA
# (i.e. before preload/torch below), or it raises "CUDA has been initialized before
# importing the `spaces` package". So import it FIRST. Absent off-Space (local/tests),
# where the `@spaces.GPU` decorator is a documented no-op anyway.
try:
    import spaces  # noqa: F401
except ImportError:  # pragma: no cover - not on a ZeroGPU Space (local / tests)
    pass

# Prefer the ABSOLUTE packaged module so that `import app` (flat, via the test
# pythonpath) and `import sql_env.app` BOTH delegate to the exact same
# `sql_env.server.app_ui` — a single `Step`/class identity (R1). The flat
# fallback (`server.app_ui`) is only for a bare layout with no installed package;
# under this repo `sql_env` always resolves, so the canonical path wins.
try:
    from sql_env.server.app_ui import _launch_demo, build_demo
except ImportError:  # pragma: no cover - bare flat-layout / direct-run fallback
    from server.app_ui import _launch_demo, build_demo  # type: ignore[no-redef]

__all__ = ["build_demo"]


if __name__ == "__main__":
    # F006/F008 — the HF ZeroGPU Space entrypoint. Heavy serving imports stay inside
    # __main__ (NOT at module import) so `import app` remains dep-light for the tests.
    #
    # Preload the available models in the PARENT process so each per-ask @spaces.GPU
    # fork inherits the loaded weights (ZeroGPU does not persist state created inside
    # the GPU call). A failed preload is logged and the model loads lazily on first ask.
    try:
        from sql_env.server.serving import preload_available
    except ImportError:  # pragma: no cover - bare flat-layout / direct-run fallback
        from server.serving import preload_available  # type: ignore[no-redef]

    preload_available()
    # _launch_demo applies the theme + CSS that build_demo stashed (gradio 6.x reads
    # theme/css at .launch() time, not on the Blocks constructor).
    _launch_demo(build_demo(enable_model_selector=True))