Closed hamishivi closed 4 months ago
Could you provide some context on why manual tokenizer loading is necessary for vllm? The changes look good to me though.
Ah, it strictly isn't, but I feel it's preferable to expose it rather than rely on vllm internals? Happy to revert it back. There are some other small fixes in this PR too (bug in toxigen eval).
Fix tokenizer issue with vllm, have to load manually. Also add budget to the eval job script, and send to multiple clusters.