Speculative Decoding for gpt-oss models

[edyapd]

I'm trying to use speculative decoding with the gpt-oss-120b model, using gpt-oss-20b as the draft model.
However, I haven't been successful, as not a single token proposed by the draft model is being accepted by the target model.
I'm wondering if I'm doing something wrong in my setup, or if this combination is not supposed to work. Any insights would be appreciated.