Added support for large models

main_backend_lighteval.py

@@ -63,8 +63,16 @@ def run_auto_eval():
     # This needs to be done
     #instance_size, instance_type = get_instance_for_model(eval_request)
     # For GPU
-    # instance_size, instance_type = "small", "g4dn.xlarge" 
-    instance_size, instance_type = "medium", "g5.2xlarge" 
+    if not eval_request or eval_request.params < 0:
+        raise ValueError("Couldn't detect number of params, please make sure the metadata is available")
+    elif eval_request.params < 4:
+        instance_size, instance_type = "small", "g4dn.xlarge" 
+    elif eval_request.params < 9:
+        instance_size, instance_type = "medium", "g5.2xlarge" 
+    elif eval_request.parmas < 24:
+        instance_size, instance_type = "large", "g4dn.12xlarge"
+    else:
+        raise ValueError("Number of params too big, can't run this model")
     # For CPU
     # instance_size, instance_type = "medium", "c6i"