Change how we ask the model questions

Brings it closer to the moondream2 example space. I really wish I could test locally.

app.py

@@ -9,6 +9,12 @@ moondream_model = AutoModelForCausalLM.from_pretrained(
     moondream_id, trust_remote_code=True, code_revision=moondream_revision
 )
 
+
+def answer_question(_img, _prompt):
+    image_embeds = moondream_model.encode_image(_img)
+    return moondream_model.answer_question(image_embeds, _prompt, moondream_tokenizer)
+
+
 with gr.Blocks() as app:
     gr.Markdown(
         """
@@ -24,8 +30,8 @@ with gr.Blocks() as app:
         img = gr.Image(label="Image")
         output = gr.TextArea(label="Output")
 
-    submit.click(moondream_model.answer_question, [img, prompt, moondream_tokenizer], output)
-    prompt.submit(moondream_model.answer_question, [img, prompt, moondream_tokenizer], output)
+    submit.click(answer_question, [img, prompt], output)
+    prompt.submit(answer_question, [img, prompt], output)
 
 if __name__ == "__main__":
     app.launch()