Add @spaces.GPU decorator for ZeroGPU Space compatibility

ZeroGPU Spaces do not attach a GPU to the main process; they require at least one @spaces.GPU-decorated function so the scheduler knows what to attach a GPU to for the duration of the call, or startup fails with "No @spaces.GPU function detected during startup".

Decorate the button event handler (run_alignment), and import spaces before torch per ZeroGPU's required import order. Model loading already happened lazily on first align() call, so it now runs inside the GPU-attached window instead of at startup.

app.py

@@ -9,6 +9,11 @@ Serves both:
 
 from __future__ import annotations
 
+# `spaces` must be imported before `torch` (imported transitively via
+# `aligner`) so it can patch CUDA initialization for ZeroGPU Spaces. On
+# non-ZeroGPU hardware (or local dev), `spaces.GPU` is a harmless no-op.
+import spaces
+
 import gradio as gr
 
 from aligner import MAX_AUDIO_SECONDS, SUPPORTED_LANGUAGES, align
@@ -21,11 +26,19 @@ EXAMPLE_AUDIO_URL = (
 EXAMPLE_TEXT = "甚至出现交易几乎停滞的情况。"
 
 
+@spaces.GPU(duration=120)
 def run_alignment(audio, text: str, language: str):
     """Event handler for the UI button, and the function exposed as the API.
 
     Returns (dataframe_rows, raw_json) so the UI shows both a readable
     table and the full structured payload; API callers get the same tuple.
+
+    `@spaces.GPU` is required on ZeroGPU Spaces: it's how the ZeroGPU
+    scheduler detects which functions need a GPU attached and allocates one
+    for the duration of the call. Model loading happens lazily inside
+    `aligner.get_model()`, on first call to `align()` below, so it also
+    runs inside this GPU-attached window (required — CUDA calls made
+    outside a `@spaces.GPU` call fail on ZeroGPU hardware).
     """
     try:
         spans = align(audio=audio, text=text, language=language)

pyproject.toml

@@ -12,6 +12,9 @@ dependencies = [
     # via qwen-asr) — see requirements.txt for details. Matches README.md's
     # sdk_version.
     "gradio==6.17.3",
+    # ZeroGPU SDK: provides the `@spaces.GPU` decorator app.py uses. A
+    # no-op on non-ZeroGPU hardware / local dev.
+    "spaces==0.50.4",
     "qwen-asr==0.0.6",
     "torch==2.11.0",
     "numpy==2.4.6",

requirements.txt

@@ -1,10 +1,11 @@
 # This is what the Hugging Face Space build actually installs (Spaces read
 # requirements.txt, not pyproject.toml). Keep in sync with pyproject.toml.
 #
-# gradio itself is intentionally NOT pinned here: the Space builder already
-# installs the exact `sdk_version` from README.md's frontmatter, and adding
-# a second, looser requirement for it here only grows the resolver's search
-# space.
+# gradio and spaces are intentionally NOT pinned here: on a ZeroGPU Space,
+# the builder already force-installs an exact `gradio[oauth,mcp]==...`
+# (matching README.md's `sdk_version`) and `spaces==...` (the ZeroGPU SDK
+# app.py imports for the `@spaces.GPU` decorator). Adding a second, looser
+# requirement for either here only grows the resolver's search space.
 #
 # qwen-asr's own metadata leaves several transitive deps unpinned
 # (librosa, qwen-omni-utils, flask, sox, pytz), which combined with a bare