The model began to add </s > to each main and server response

[slavonnet]

git revision from master 3fec68b

usage cmd lines

./server -ngl 16000 --numa numactl --model models/MOE/mixtral-8x7b-instruct-v0.1.Q4_K_M.gguf --slots-endpoint-disable -c 4096

./main --color --threads 8 \
       --file $1 \
       -m models/7B/saiga_mistral-q8_0.gguf \
       --reverse-prompt "### Instruction:\n" \
       --in-prefix "###RESPONCE:" \
       --batch_size 512 --ctx_size 16384 -n -1 \
       --keep -1 \
       --numa numactl \
       --repeat_last_n 64 --repeat_penalty 1.3 \
       -ngl 1048576 \
       --n_predict -2 --temp 0.1 --top_k 260 --top_p 0.98 -s 5

</s> appeared in all responses.

[ggerganov]

Fixed in #6860

[MB7979]

I'm still getting stop tokens printed at the end of outputs using main.

[ggerganov]

These are currently disabled only in server. main still renders them