Fix high VRAM usage caused by workaround for scalar type error

aiserver.py

@@ -3172,7 +3172,7 @@ def load_model(use_gpu=True, gpu_layers=None, disk_layers=None, initial_load=Fal
                             else:
                                 raise RuntimeError(f"4-bit load failed. Model type {koboldai_vars.model_type} not supported in 4-bit")
 
-                            model = model.float()
+                            model = model.half()
                         else:
                             try:
                                 tokenizer = AutoTokenizer.from_pretrained(koboldai_vars.custmodpth, revision=koboldai_vars.revision, cache_dir="cache", use_fast=False)