ExLLaMA fixes

aiserver.py

@@ -1737,10 +1737,6 @@ def load_model(model_backend, initial_load=False):
         #Have to add a sleep so the server will send the emit for some reason
         time.sleep(0.1)    
     
-    if 'model' in globals():
-        model.unload()
-    
-    
     # If transformers model was selected & GPU available, ask to use CPU or GPU
     if(not koboldai_vars.use_colab_tpu and koboldai_vars.model not in ["InferKit", "Colab", "API", "CLUSTER", "OAI", "GooseAI" , "ReadOnly", "TPUMeshTransformerGPTJ", "TPUMeshTransformerGPTNeoX"]):
         # loadmodelsettings()
@@ -6337,6 +6333,9 @@ def UI_2_resubmit_model_info(data):
 @socketio.on('load_model')
 @logger.catch
 def UI_2_load_model(data):
+    logger.debug("Unloading previous model")
+    if 'model' in globals():
+        model.unload()
     logger.debug("Loading model with user input of: {}".format(data))
     model_backends[data['plugin']].set_input_parameters(data)
     load_model(data['plugin'])
@@ -8568,6 +8567,8 @@ def put_model(body: ModelSelectionSchema):
             backend = "Huggingface"
 
     try:
+        if 'model' in globals():
+            model.unload()
         load_model(backend)
     except Exception as e:
         koboldai_vars.model = old_model

modeling/inference_models/exllama/class.py

@@ -221,7 +221,7 @@ class model_backend(InferenceModel):
         self.tokenizer._koboldai_header = self.tokenizer.encode("")
 
     def unload(self):
-        self.model_config = None
+        #self.model_config = None # This breaks more than it fixes - Henk
 
         self.model = None
         self.tokenizer = None
@@ -289,7 +289,7 @@ class model_backend(InferenceModel):
             torch.manual_seed(seed)
 
         bad_words_ids = [self.tokenizer.bos_token_id]
-        if utils.koboldai_vars.use_default_badwordids:
+        if utils.koboldai_vars.use_default_badwordsids:
             bad_words_ids.append(self.tokenizer.eos_token_id)
             bad_words_ids.extend(self.bracket_tokens)
         if single_line: