Create app.py

app.py

@@ -0,0 +1,71 @@
+import gradio as gr
+
+import torch
+from transformers import AutoTokenizer, AutoModelForCausalLM, TextIteratorStreamer, pipeline
+from threading import Thread
+
+# Function that accepts a prompt and generates text using the phi2 pipeline
+def generate(message, chat_history, max_new_tokens=256):
+
+  history = []
+
+  for sent, received in chat_history:
+    history.append({"role": "user", "content": sent})
+    history.append({"role": "assistant", "content": received})
+
+  history.append({"role": "user", "content": message})
+  #print(history)
+
+  if len(tokenizer.apply_chat_template(history)) > 512:
+    yield "chat history is too long"
+  else:
+    # Streamer
+    streamer = TextIteratorStreamer(tokenizer=tokenizer, skip_prompt=True, skip_special_tokens=True, timeout=300.0)
+    thread = Thread(target=llama_am,
+                    kwargs={
+                        "text_inputs":history,
+                        "max_new_tokens":max_new_tokens,
+                        "temperature":0.0,
+                        "repetition_penalty":1.25,
+                        "streamer":streamer
+                        }
+                    )
+    thread.start()
+
+    generated_text = ""
+    for word in streamer:
+      generated_text += word
+      response = generated_text.strip()
+
+      yield response
+
+# Chat interface with gradio
+with gr.Blocks() as demo:
+  gr.Markdown("""
+  # Llama 3.2 Amharic Chatbot Demo
+
+  This chatbot was created using a finetuned version of my 180 million parameter Llama 3.2 Amharic transformer model, [Llama-3.2-Amharic-Instruct-Beta](https://huggingface.co/rasyosef/Phi-1_5-Instruct-v0.1).
+  """)
+
+  tokens_slider = gr.Slider(8, 256, value=64, label="Maximum new tokens", info="A larger `max_new_tokens` parameter value gives you longer text responses but at the cost of a slower response time.")
+
+  chatbot = gr.ChatInterface(
+    chatbot=gr.Chatbot(height=400),
+    fn=generate,
+    additional_inputs=[tokens_slider],
+    stop_btn=None,
+    examples=[
+        ["የኢትዮጵያ ዋና ከተማ ስም ምንድን ነው?"],
+        ["የፈረንሳይ ዋና ከተማ ስም ምንድን ነው?"],
+        ["የኢትዮጵያ የመጨረሻው ንጉስ ማን ነበሩ?"],
+        ["የ አሜሪካ ፕሬዝዳንት ማን ነው?"],
+        ["የ እስራኤል ጠቅላይ ሚንስትር ማን ነው?"],
+        ["ሶስት የአፍሪካ ሃገራት ጥቀስልኝ"],
+        ["፫ የአውሮፓ ሃገራት ጥቀስልኝ"],
+        ["፫ የእስያ ሃገራት ጥቀስልኝ"],
+        ["ሶስት የአሜሪካ መሪዎችን ስም ጥቀስ"],
+        ["አምስት የአሜሪካ ከተማዎችን ስም ጥቀስ"]
+      ]
+  )
+
+demo.queue().launch(debug=True)