Deploy Lost & Found Desk

Co-authored-by: Codex <codex@openai.com>

README.md

@@ -213,7 +213,7 @@ This repo was updated to match the current official docs as of 2026-06-14:
 | --- | --- | --- |
 | Gradio | `gradio==6.17.3` | Space metadata pins the same `sdk_version`. |
 | Gradio Server mode | `gradio.Server` + `@app.api()` | Gives FastAPI routes plus Gradio API queue/concurrency/ZeroGPU behavior. |
-| ZeroGPU | `spaces>=0.50.4`, `@spaces.GPU` via `maybe_zerogpu()` | GPU-dependent calls are wrapped only when `ZEROGPU=1`; in ZeroGPU mode the real MiniCPM adapters are eagerly loaded at app startup to match the ZeroGPU model-loading recommendation. |
+| ZeroGPU | `spaces>=0.50.4`, `@spaces.GPU` via `maybe_zerogpu()` | GPU-dependent calls are registered with the official decorator; `spaces` activates the wrapper when HF sets `SPACES_ZERO_GPU=true`, and the real MiniCPM adapters are eagerly loaded at app startup to match the ZeroGPU model-loading recommendation. |
 | MiniCPM-V | `openbmb/MiniCPM-V-4.6` through `AutoProcessor` + `AutoModelForImageTextToText` | Matches the model card's Transformers path. |
 | MiniCPM5 | `openbmb/MiniCPM5-1B` through `AutoTokenizer` + `AutoModelForCausalLM` | Matches the model card's standard `LlamaForCausalLM` path. |
 | Frontend | Svelte 5 + Vite 8 + `@gradio/client` | The compiled Svelte app is committed into `static/` so the Space needs no Node build step. |

src/lost_found_desk/config.py

@@ -10,6 +10,10 @@ def _env_flag(name: str, default: str) -> bool:
     return os.getenv(name, default).lower() not in {"0", "false", "no", ""}
 
 
+def _hf_zero_gpu_enabled() -> bool:
+    return os.getenv("SPACES_ZERO_GPU", "").lower() in {"1", "t", "true"}
+
+
 @dataclass(frozen=True)
 class Settings:
     app_name: str
@@ -66,7 +70,7 @@ def _load_settings() -> Settings:
         retrieval_enabled=_env_flag("LFD_RETRIEVAL", "1"),
         embed_strong_threshold=float(os.getenv("LFD_EMBED_STRONG", "0.32")),
         embed_weak_threshold=float(os.getenv("LFD_EMBED_WEAK", "0.18")),
-        zerogpu=os.getenv("ZEROGPU", "").lower() in {"1", "true", "yes"},
+        zerogpu=_hf_zero_gpu_enabled(),
         zerogpu_caption_duration=int(os.getenv("LFD_ZEROGPU_CAPTION_DURATION", "45")),
         zerogpu_text_duration=int(os.getenv("LFD_ZEROGPU_TEXT_DURATION", "30")),
         seed_event_id=os.getenv("LFD_SEED_EVENT_ID", "demo"),

src/lost_found_desk/runtime/device.py

@@ -19,7 +19,7 @@ def detect_runtime(requested: str = "auto") -> RuntimeInfo:
     """
     requested = (requested or "auto").lower()
 
-    if os.getenv("ZEROGPU", "").lower() in {"1", "true", "yes"}:
+    if os.getenv("SPACES_ZERO_GPU", "").lower() in {"1", "t", "true"}:
         return RuntimeInfo(runtime="zerogpu", device="cuda", dtype="auto")
 
     try:

src/lost_found_desk/runtime/gpu.py

@@ -1,24 +1,26 @@
 from __future__ import annotations
 
-import os
 from typing import Callable, TypeVar
 
 F = TypeVar("F", bound=Callable)
 
 
 def maybe_zerogpu(duration: int = 60, size: str | None = None):
-    """Return @spaces.GPU on ZeroGPU; otherwise return an identity decorator.
+    """Return the official @spaces.GPU decorator when the package is present.
 
-    Keeping this import isolated makes local CUDA/MPS development independent
-    from the Hugging Face Spaces-only `spaces` module.
+    The `spaces` package decides whether the current runtime is ZeroGPU from
+    Hugging Face's `SPACES_ZERO_GPU` environment variable. Outside ZeroGPU it
+    returns the original function unchanged, so local CUDA/MPS/CPU runs keep the
+    same call path while HF startup can still discover GPU-decorated functions.
     """
-    is_zero = os.getenv("ZEROGPU", "").lower() in {"1", "true", "yes"}
-    if not is_zero:
+
+    try:
+        import spaces  # type: ignore
+    except ImportError:
         def identity(fn: F) -> F:
             return fn
-        return identity
 
-    import spaces  # type: ignore
+        return identity
 
     kwargs = {"duration": duration}
     if size: