Update app.py

app.py

@@ -17,13 +17,18 @@ print("Loading model...")
 DTYPE = torch.bfloat16 if torch.cuda.is_available() else torch.float32
 base = AutoModelForCausalLM.from_pretrained(
     BASE_MODEL,
-    dtype=DTYPE,
-    device_map="auto",
+    torch_dtype=DTYPE,
+    device_map="cpu",          # load to CPU first, avoids meta tensor issue
     trust_remote_code=True,
-    low_cpu_mem_usage=True,
+    low_cpu_mem_usage=False,   # disable — this is what triggers meta tensors
+)
+model = PeftModel.from_pretrained(
+    base,
+    ADAPTER_PATH,
+    is_trainable=False,
 )
-model = PeftModel.from_pretrained(base, ADAPTER_PATH)
 model = model.merge_and_unload()
+model = model.to(DTYPE)
 model.eval()
 print("Ready!")