Convert to CPU-compatible Space: remove ZeroGPU dependency

- Remove spaces import and @spaces.GPU decorator
- Auto-detect device (CUDA > MPS > CPU) instead of hardcoded "cuda"
- Use float32 dtype on CPU, bfloat16 on CUDA/MPS
- Pass DEVICE variable to tts_model.inference() instead of inline check
- Remove spaces package from requirements.txt
- Update README title/description for CPU variant

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

README.md

@@ -1,5 +1,5 @@
 ---
-title: MOSS-TTS-Nano
+title: MOSS-TTS-Nano CPU
 emoji: 🎙️
 colorFrom: blue
 colorTo: purple
@@ -7,7 +7,7 @@ sdk: gradio
 sdk_version: 6.5.1
 app_file: app.py
 pinned: false
-short_description: 0.1B multilingual TTS with voice cloning
+short_description: 0.1B multilingual TTS with voice cloning (CPU)
 ---
 
 # MOSS-TTS-Nano

app.py

@@ -8,20 +8,22 @@ from pathlib import Path
 import gradio as gr
 import torch
 
-try:
-    import spaces
-except ImportError:
-    class _SpacesFallback:
-        @staticmethod
-        def GPU(*args, **kwargs):
-            def decorator(func):
-                return func
-            return decorator
-    spaces = _SpacesFallback()
-
 logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s: %(message)s")
 logger = logging.getLogger(__name__)
 
+# ─── Device detection ─────────────────────────────────────────────────────────
+def _get_device() -> str:
+    if torch.cuda.is_available():
+        return "cuda"
+    if torch.backends.mps.is_available():
+        return "mps"
+    return "cpu"
+
+DEVICE = _get_device()
+# bfloat16 works on CUDA/MPS; use float32 on CPU for stability
+DTYPE = torch.bfloat16 if DEVICE in ("cuda", "mps") else torch.float32
+logger.info("Running on device: %s, dtype: %s", DEVICE, DTYPE)
+
 MODEL_ID = "OpenMOSS-Team/MOSS-TTS-Nano-100M"
 AUDIO_TOKENIZER_ID = "OpenMOSS-Team/MOSS-Audio-Tokenizer-Nano"
 OUTPUT_DIR = Path("/tmp/moss-tts-output")
@@ -29,18 +31,18 @@ SAMPLE_DIR = Path("/tmp/moss-tts-samples")
 OUTPUT_DIR.mkdir(parents=True, exist_ok=True)
 SAMPLE_DIR.mkdir(parents=True, exist_ok=True)
 
-# ─── Load models at module level (ZeroGPU packs them at startup) ──────────────
+# ─── Load models ──────────────────────────────────────────────────────────────
 from transformers import AutoModelForCausalLM, AutoModel, AutoTokenizer
 
 logger.info("Loading TTS model: %s", MODEL_ID)
 tts_model = AutoModelForCausalLM.from_pretrained(
     MODEL_ID,
     trust_remote_code=True,
-    dtype=torch.bfloat16,
+    torch_dtype=DTYPE,
     low_cpu_mem_usage=True,
 )
 tts_model.eval()
-tts_model._set_attention_implementation("sdpa")  # flash_attn not installed; sdpa is always available
+tts_model._set_attention_implementation("sdpa")
 
 logger.info("Loading audio tokenizer: %s", AUDIO_TOKENIZER_ID)
 audio_tokenizer = AutoModel.from_pretrained(
@@ -52,9 +54,9 @@ audio_tokenizer.eval()
 logger.info("Loading text tokenizer")
 text_tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
 
-logger.info("Moving models to CUDA (ZeroGPU packing) …")
-tts_model.to("cuda")
-audio_tokenizer.to("cuda")
+logger.info("Moving models to %s …", DEVICE)
+tts_model.to(DEVICE)
+audio_tokenizer.to(DEVICE)
 logger.info("All models ready.")
 
 # ─── Sample audio files ───────────────────────────────────────────────────────
@@ -108,7 +110,6 @@ LANG_TO_SAMPLE_AUDIO: dict[str, str | None] = {
 
 
 # ─── Inference ────────────────────────────────────────────────────────────────
-@spaces.GPU(duration=120)
 def generate_speech(
     text: str,
     reference_audio: str | None,
@@ -124,10 +125,9 @@ def generate_speech(
     seed_int = int(seed) if seed else 0
     if seed_int != 0:
         torch.manual_seed(seed_int)
-        if torch.cuda.is_available():
+        if DEVICE == "cuda":
             torch.cuda.manual_seed(seed_int)
 
-    device = "cuda" if torch.cuda.is_available() else "cpu"
     output_path = str(OUTPUT_DIR / f"output_{os.getpid()}_{abs(hash(text)) % 1_000_000}.wav")
 
     try:
@@ -139,7 +139,7 @@ def generate_speech(
             text_tokenizer=text_tokenizer,
             audio_tokenizer=audio_tokenizer,
             audio_tokenizer_type="moss-audio-tokenizer-nano",
-            device=device,
+            device=DEVICE,
             max_new_frames=int(max_new_frames),
             do_sample=bool(do_sample),
             use_kv_cache=True,

requirements.txt

@@ -5,5 +5,4 @@ safetensors>=0.4.3
 soundfile>=0.13.1
 torchaudio==2.7.1
 gradio==6.5.1
-spaces
 accelerate>=0.30.0