Update app.py

app.py

@@ -14,9 +14,32 @@ BASE_MODEL = "stanford-crfm/BioMedLM"  # or the path you used originally
 ADAPTER_PATH = "Tufan1/BioMedLM-Cardio-Fold10-CPU"
 tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)
 
+# Load base model with safe settings
+base_model = AutoModelForCausalLM.from_pretrained(
+    BASE_MODEL,
+    torch_dtype=torch.float32,
+    low_cpu_mem_usage=True,  # Reduces memory spikes
+    device_map="cpu"  # Force CPU loading
+)
+
+# Load PEFT adapter with proper weight mapping
+model = PeftModel.from_pretrained(
+    base_model,
+    ADAPTER_PATH,
+    device_map="cpu",
+    is_trainable=False
+)
+
+# Explicitly load adapter weights to CPU
+adapter_weights = torch.load(
+    f"{ADAPTER_PATH}/adapter_model.bin", 
+    map_location=torch.device('cpu')
+)
+model.load_state_dict(adapter_weights, strict=False)
+
 # Force CPU-safe model loading
-base_model = AutoModelForCausalLM.from_pretrained(BASE_MODEL, torch_dtype=torch.float32)
-model = PeftModel.from_pretrained(base_model, ADAPTER_PATH, device_map=None).to("cpu")
+#base_model = AutoModelForCausalLM.from_pretrained(BASE_MODEL, torch_dtype=torch.float32)
+#model = PeftModel.from_pretrained(base_model, ADAPTER_PATH, device_map=None).to("cpu")
 
 
 # Dictionaries to decode user inputs