Llama 70b returning incomplete responses

Inference Endpoints on the Hub
1

Hello! All of a sudden, the 70b chat llama model started to return incomplete text results. Even using the same exact input, the result returned is an incomplete version of the correct one.

Example:

curl --request POST \
  --url /static-proxy?url=https%3A%2F%2Fapi-inference.huggingface.co%2Fmodels%2Fmeta-llama%2FLlama-2-70b-chat-hf \
  --header 'Authorization: Bearer API_KEY' \
  --header 'Content-Type: application/json' \
  --data '{"inputs": "Please write a long, very long poem about mars and cats"}'

Result:

[
	{
		"generated_text": "Please write a long, very long poem about mars and cats.\n\nMars, the red planet, a world so vast,\nA place where no"
	}
]

Has something changed in the last few days?