Generate crashed by repeatedly generating <think>
#35
by qpz - opened
During the model generation process (usually near the end), it suddenly crashes and repeatedly generates tokens. This occurs when generating on the AIME test set responses, and it is triggered in about 20% of cases.
qpz changed discussion title from Generate crashed by repeatedly generate <think> to Generate crashed by repeatedly generating <think>
do you use vvlm?
try add the args --enforce-eager