fix

app.py

@@ -15,9 +15,16 @@ from transformers import (
 import torch.nn.functional as F
 from collections import defaultdict
 HF_TOKEN = os.getenv("HF_token")
-CHAT_MODEL_ID = "NousResearch/Meta-Llama-3-8B-Instruct"
+CHAT_MODEL_ID = "meta-llama/Llama-3.1-8B-Instruct"
 EMB_MODEL_ID  = "mixedbread-ai/mxbai-embed-large-v1"
 MAX_PROMPT_TOKENS = 8192
+import transformers
+pipeline = transformers.pipeline(
+    "text-generation",
+    model=CHAT_MODEL_ID,
+    model_kwargs={"torch_dtype": torch.bfloat16},
+    device_map="auto",
+)
 
 # ---------- new defaults & helper ------------------
 DEFAULT_TEMP        = 0.7