Try loading moondream2 using parts of their example

app.py

@@ -1,18 +1,13 @@
 import gradio as gr
-from transformers import pipeline
-
-# from transformers import AutoModelForCausalLM, AutoTokenizer
-
-# from PIL import Image
-
+from transformers import AutoModelForCausalLM, AutoTokenizer
 
+# Moondream does not support the HuggingFace pipeline system, so we have to do it manually
 moondream_id = "vikhyatk/moondream2"
 moondream_revision = "2024-04-02"
-# moondream_model = AutoModelForCausalLM.from_pretrained(
-#     moondream_id, trust_remote_code=True, code_revision=moondream_revision
-# )
-# tokenizer = AutoTokenizer.from_pretrained(moondream_id, code_revision=moondream_revision)
-pipe = pipeline(model=moondream_id, revision=moondream_revision, trust_remote_code=True)
+moondream_tokenizer = AutoTokenizer.from_pretrained(moondream_id, code_revision=moondream_revision)
+moondream_model = AutoModelForCausalLM.from_pretrained(
+    moondream_id, trust_remote_code=True, code_revision=moondream_revision
+)
 
 with gr.Blocks() as app:
     gr.Markdown(
@@ -26,7 +21,11 @@ with gr.Blocks() as app:
         prompt = gr.Textbox(label="Input", value="Describe this image.")
         submit = gr.Button(label="Submit")
     with gr.Row():
-        img = gr.Image(type="pil")
+        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)
 
 if __name__ == "__main__":
     app.launch()