Update run.py

run.py

@@ -2,7 +2,7 @@
 # Title:  German AI-Interface with advanced RAG
 # Author: Andreas Fischer
 # Date:   January 31st, 2023
-# Last update: February 21st, 2024
+# Last update: February 25st, 2024
 ##########################################################################################
 
 #https://github.com/abetlen/llama-cpp-python/issues/306
@@ -187,18 +187,25 @@ else:
 #------------
 
 def extend_prompt(message="", history=None, system=None, RAGAddon=None, system2=None, zeichenlimit=None,historylimit=4): #float("Inf")
+  startOfString=""
   if zeichenlimit is None: zeichenlimit=1000000000 # :-)
-  template0="[INST] {system} [/INST]</s>" #<s>
-  template1="[INST] {message} [/INST] "
-  template2="{response}</s>"
+  template0=" [INST]{system}\n  [/INST] </s>" if onPrem else "<s> [INST] {system} [/INST] </s>" 
+  template1=" [INST] {message} [/INST]"
+  template2=" {response}</s>"
+  if("gemma-" in modelPath): # https://huggingface.co/mistralai/Mixtral-8x7B-Instruct-v0.1
+    template0="<start_of_turn>user{system}</end_of_turn>" 
+    template1="<start_of_turn>user{message}</end_of_turn><start_of_turn>model"
+    template2="{response}</end_of_turn>"
   if("mixtral-8x7b-instruct" in modelPath): # https://huggingface.co/mistralai/Mixtral-8x7B-Instruct-v0.1
-    template0="[INST] {system} [/INST]</s>" #<s>
-    template1="[INST] {message} [/INST] "
-    template2="{response}</s>"
+    startOfString="<s>"
+    template0=" [INST]{system}\n  [/INST] </s>" if onPrem else "<s> [INST]{system}\n  [/INST] </s>" 
+    template1=" [INST] {message} [/INST]"
+    template2=" {response}</s>"
   if("Mistral-7B-Instruct" in modelPath): #https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2
-    template0="[INST] {system} [/INST]</s>" #<s>
-    template1="[INST] {message} [/INST] "
-    template2="{response}</s>"
+    startOfString="<s>"
+    template0="[INST]{system}\n [/INST]</s>" if onPrem else "<s>[INST]{system}\n [/INST]</s>" 
+    template1="[INST] {message} [/INST]"
+    template2=" {response}</s>"
   if("openchat-3.5" in modelPath): #https://huggingface.co/TheBloke/openchat-3.5-0106-GGUF
     template0="GPT4 Correct User: {system}<|end_of_turn|>GPT4 Correct Assistant: Okay.<|end_of_turn|>"
     template1="GPT4 Correct User: {message}<|end_of_turn|>GPT4 Correct Assistant: "
@@ -231,7 +238,7 @@ def extend_prompt(message="", history=None, system=None, RAGAddon=None, system2=
   if message is not None: prompt += template1.format(message=message[:zeichenlimit])                #"[INST] {message} [/INST]"
   if system2 is not None:
     prompt += system2
-  return prompt
+  return startOfString+prompt
 
 import gradio as gr
 import requests