Fix LoRA model loading

- Load base Mistral-7B model first
- Apply LoRA adapter with PeftModel
- Add peft to requirements
- Fix torch_dtype deprecation warning

app.py

@@ -1,21 +1,27 @@
 import gradio as gr
 from transformers import AutoModelForCausalLM, AutoTokenizer
+from peft import PeftModel
 import torch
 
 print("🔄 Loading CCISD TEKS Educational Assistant...")
 print("   This may take 1-2 minutes on first launch...")
 
-# Load your trained model
-MODEL_NAME = "robworks-software/ccisd-teks-educator-mistral7b"
+# Load base model + LoRA adapter
+BASE_MODEL = "mistralai/Mistral-7B-Instruct-v0.2"
+ADAPTER_MODEL = "robworks-software/ccisd-teks-educator-mistral7b"
 
-tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
+print(f"📥 Loading base model: {BASE_MODEL}...")
+tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)
 model = AutoModelForCausalLM.from_pretrained(
-    MODEL_NAME,
-    torch_dtype=torch.float16,
+    BASE_MODEL,
+    dtype=torch.float16,
     device_map="auto",
     low_cpu_mem_usage=True
 )
 
+print(f"🔧 Loading LoRA adapter: {ADAPTER_MODEL}...")
+model = PeftModel.from_pretrained(model, ADAPTER_MODEL)
+
 print("✅ Model loaded successfully!")
 
 def chat(message, history):

requirements.txt

@@ -2,5 +2,6 @@ gradio>=4.0.0
 transformers>=4.35.0
 torch>=2.0.0
 accelerate>=0.24.0
+peft>=0.7.0
 sentencepiece>=0.1.99
 protobuf>=3.20.0