Upload moondream.py

moondream.py

@@ -124,8 +124,11 @@ class Moondream(PreTrainedModel):
     ):
         prompt = f"<image>\n\n{chat_history}Question: {question}\n\nAnswer:"
         streamer = TextStreamer(tokenizer)
+
+
         output_ids = self.text_model.generate(
             inputs_embeds=self.input_embeds(prompt, image_embeds, tokenizer),
+            attention_mask=torch.ones(self.inputs_embeds.shape, dtype=torch.long, device=self.device), #maybe remove
             streamer=streamer,
             **kwargs,
         )