wizardcoder-ggml

Paused

matthoffner commited on Jun 21, 2023

Commit

ebbbded

1 Parent(s): 4efb547

Update main.py

Files changed (1) hide show

main.py CHANGED Viewed

@@ -66,7 +66,7 @@ async def chat(request: ChatCompletionRequest):
     tokens = llm.tokenize(combined_messages)
     try:
-        chat_chunks = llm.generate(tokens, max_new_tokens=request.max_tokens)
     except Exception as e:
         raise HTTPException(status_code=500, detail=str(e))

     tokens = llm.tokenize(combined_messages)
     try:
+        chat_chunks = llm.generate(tokens)
     except Exception as e:
         raise HTTPException(status_code=500, detail=str(e))