test

app.py

@@ -4,10 +4,9 @@ from langchain_community.llms import CTransformers
 from transformers import AutoModel
 
 # load the model
-
-TinyLlamaModel = AutoModel.from_pretrained("NousResearch/Llama-2-7b-hf")
+PretrainedModel = AutoModel.from_pretrained("TinyLlama/TinyLlama-1.1B-Chat-v1.0")
 def getLlamaResponse(input_text, no_words, blog_style):
-    llm=CTransformers(model=TinyLlamaModel,
+    llm=CTransformers(model=PretrainedModel,
         model_type='llama',
         config={'max_new_tokens':256, 'temperature':0.2})
 

requirements.txt

@@ -1,5 +1,6 @@
 sentence-transformers
 uvicorn
+transformers
 ctransformers
 langchain
 langchain-community