Update app.py

app.py

@@ -69,22 +69,39 @@ def _device() -> str:
 
 print("Loading model (this can take a minute on first launch)...")
 DEVICE = _device()
-DTYPE = torch.float16 if DEVICE in ("cuda", "mps") else torch.float32
+DTYPE = torch.float32  # ZeroCPU 只有 CPU，强制 fp32
+# DTYPE = torch.float16 if DEVICE in ("cuda", "mps") else torch.float32
 
 tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
 
 tokenizer.pad_token = tokenizer.eos_token
 
 _base = AutoModelForCausalLM.from_pretrained(
-    MODEL_ID, torch_dtype=DTYPE, trust_remote_code=True
+    MODEL_ID,
+    torch_dtype=torch.float32,
+    device_map="cpu",
+    trust_remote_code=True,
 )
-# Keep the adapter separate (not merged) so we can toggle it on/off.
+
+# adapter 加载前临时 patch 掉 spaces 的 torch 拦截
+import safetensors.torch as _st
+_orig = _st.load_file
+_st.load_file = lambda f, **kw: _orig(f, device="cpu")
 model = PeftModel.from_pretrained(_base, ADAPTER_DIR)
-model = model.to(DEVICE)
+_st.load_file = _orig  # 还原
+
 model.eval()
+
+# _base = AutoModelForCausalLM.from_pretrained(
+#     MODEL_ID, torch_dtype=DTYPE, trust_remote_code=True
+# )
+# # Keep the adapter separate (not merged) so we can toggle it on/off.
+# model = PeftModel.from_pretrained(_base, ADAPTER_DIR)
+# model = model.to(DEVICE)
+# model.eval()
 print(f"Model loaded on {DEVICE}.")
 
-@spaces.GPU(duration=120)
+# @spaces.GPU(duration=120)
 def _generate(text: str, max_new_tokens: int = 400) -> str:
     """Run one generation with the current adapter state."""
     messages = [