allenai / mmda

multimodal document analysis
Apache License 2.0
158 stars 18 forks source link

Caps max allowed version of transformers for vila. #279

Closed cmwilhelm closed 8 months ago

cmwilhelm commented 8 months ago

v4.34.0 release did a complete refactor of the tokenizer module, see:

https://github.com/huggingface/transformers/pull/23909

Something about the difference is causing vila to produce literally billions of lines of log warning messages to Datadog in prod. I don't know if these warnings are meaningful, but they are expensive.

Example logs: https://app.datadoghq.com/logs?query=service%3Avila-v0%20&cols=host%2Cservice&index=%2A&messageDisplay=inline&refresh_mode=paused&stream_sort=desc&viz=stream&from_ts=1697556761689&to_ts=1697557153857&live=false