almost

app.py

@@ -15,10 +15,12 @@ model.eval()
 processor = ChameleonProcessor.from_pretrained(model_path)
 tokenizer = processor.tokenizer
 
+# file_name, alt
 multimodal_file = tuple[str, str]
+# {'text': 'message here', 'files': []}
 multimodal_message = list[str | multimodal_file] | multimodal_file
 # todo: verify this type with gr.ChatInterface
-message_t = str | multimodal_message
+message_t = dict[str, str | list[multimodal_file]]
 history_t = list[tuple[str, str] | list[tuple[multimodal_message, multimodal_message]]]
 
 def history_to_prompt(
@@ -28,9 +30,10 @@ def history_to_prompt(
         image_placeholder = "<image>"
     ):
 
-    prompt = ""
-    images = []
-    for turn in history + (message, None):
+    prompt = message["text"]
+    images = [Image.open(f) for f in message["files"]]
+    
+    for turn in history:
         print("turn:", turn)
         # turn should be a tuple of user message and assistant message
         for message in turn:
@@ -51,7 +54,7 @@ def history_to_prompt(
             prompt += eot_id
     return prompt, images
 
-@spaces.GPU(duration=90)
+@spaces.GPU(duration=30)
 def respond(
     message,
     history: history_t,
@@ -72,7 +75,7 @@ def respond(
 
     inputs = processor(prompt, images=images, return_tensors="pt").to(model.device, dtype=torch.bfloat16)
 
-    streamer = TextIteratorStreamer(tokenizer)
+    streamer = TextIteratorStreamer(tokenizer, skip_prompt=True)
     generation_kwargs = dict(inputs, streamer=streamer, max_new_tokens=20)
 
     try: