fix build: torch 2.8.0, python 3.12.12, module-level load

README.md

@@ -5,6 +5,7 @@ colorFrom: green
 colorTo: indigo
 sdk: gradio
 app_file: app.py
+python_version: "3.12.12"
 pinned: false
 short_description: "Taiwan-Mandarin ASR: Traditional + zh-en code-switch"
 models:

app.py

@@ -1,18 +1,18 @@
 """TEA-ASR demo (Hugging Face Spaces, ZeroGPU). Taiwan-Mandarin ASR: Traditional script + Taiwanese lexicon +
 Mandarin-English code-switch, adapted from Qwen3-ASR. No runtime OpenCC (the Traditional decode is baked into the
-model's own tokenizer). `import spaces` MUST come before torch for ZeroGPU."""
+model's own tokenizer).
+
+ZeroGPU rules followed: `import spaces` before torch; models placed on cuda at MODULE level (CUDA-emulated at
+startup); GPU-dependent fn decorated with @spaces.GPU. torch pinned to a ZeroGPU-supported version (see
+requirements.txt) and Python pinned to 3.12 (see README)."""
 import os
 import spaces
+import torch
 import gradio as gr
+import numpy as np
+import librosa
 
-MODELS = {
-    "TEA-ASR-1.7B (best)": "JacobLinCool/TEA-ASR-1.7B",
-    "TEA-ASR-0.6B (fast)": "JacobLinCool/TEA-ASR-0.6B",
-}
-LANGS = ["auto", "Chinese", "English"]
-_cache = {}
-
-# private model repos -> authenticate from the HF_TOKEN Space secret (also lets transformers find it)
+# private model repos -> need an HF_TOKEN secret (read) in the Space settings
 if os.environ.get("HF_TOKEN"):
     try:
         from huggingface_hub import login
@@ -20,26 +20,34 @@ if os.environ.get("HF_TOKEN"):
     except Exception as e:
         print("HF login failed:", e)
 
+REPOS = {
+    "TEA-ASR-1.7B (best)": "JacobLinCool/TEA-ASR-1.7B",
+    "TEA-ASR-0.6B (fast)": "JacobLinCool/TEA-ASR-0.6B",
+}
+LANGS = ["auto", "Chinese", "English"]
 
-def _load(repo):
-    import torch
+# load both models on cuda at module level (recommended ZeroGPU pattern)
+MODELS, LOAD_ERR = {}, None
+try:
     from qwen_asr import Qwen3ASRModel
-    return Qwen3ASRModel.from_pretrained(repo, dtype=torch.bfloat16, device_map="cuda:0")
+    for _name, _repo in REPOS.items():
+        MODELS[_name] = Qwen3ASRModel.from_pretrained(_repo, dtype=torch.bfloat16, device_map="cuda:0")
+except Exception as e:
+    LOAD_ERR = repr(e)
+    print("model load failed:", LOAD_ERR)
 
 
 @spaces.GPU(duration=120)
 def transcribe(audio_path, model_choice, language, context):
+    if LOAD_ERR:
+        return ("Model load failed. If the models are private, add an HF_TOKEN secret (read) in this Space's "
+                f"Settings.\n\n{LOAD_ERR}")
     if not audio_path:
         return "請提供音訊 / Please provide audio."
-    import librosa, numpy as np
-    repo = MODELS[model_choice]
-    if repo not in _cache:
-        _cache[repo] = _load(repo)
-    model = _cache[repo]
     wav, _ = librosa.load(audio_path, sr=16000, mono=True)
     lang = None if language == "auto" else language
-    out = model.transcribe(audio=[(np.asarray(wav, dtype="float32"), 16000)],
-                           context=(context or ""), language=lang)[0]
+    out = MODELS[model_choice].transcribe(
+        audio=[(np.asarray(wav, dtype="float32"), 16000)], context=(context or ""), language=lang)[0]
     return out.text
 
 
@@ -53,14 +61,14 @@ DESC = (
     "**TEA-ASR (Taiwan Everyday Audio)** — Traditional-script + Taiwanese-lexicon ASR with robust "
     "Mandarin–English code-switch, adapted from Qwen3-ASR with a tokenizer-first procedure. "
     "Output is Traditional Chinese (軟體/網路/影片…) with **no runtime OpenCC**. "
-    "Pick a model, optionally set the language hint to *Chinese* (best for Taiwan speech), and transcribe."
+    "Set the language hint to *Chinese* for Taiwan speech (best results)."
 )
 
 demo = gr.Interface(
     fn=transcribe,
     inputs=[
         gr.Audio(type="filepath", label="Audio (upload or record)"),
-        gr.Dropdown(list(MODELS), value="TEA-ASR-1.7B (best)", label="Model"),
+        gr.Dropdown(list(REPOS), value="TEA-ASR-1.7B (best)", label="Model"),
         gr.Dropdown(LANGS, value="Chinese", label="Language hint"),
         gr.Textbox(label="Context / hotwords (optional)", placeholder="例如：台積電 員工 名稱…"),
     ],

requirements.txt

@@ -1,10 +1,2 @@
-spaces
-torch==2.4.0
-transformers==4.57.6
-tokenizers==0.22.2
+torch==2.8.0
 qwen-asr==0.0.6
-qwen-omni-utils==0.0.9
-accelerate
-soundfile
-librosa
-numpy