Open Models Almanac
Methodssoftware implementationinfrastructure service

single mention · software implementation

Write-ahead log (WAL)

A write-ahead log here is a token-granular log kept per generation request in a rollout service.1 Each newly generated token is appended immediately, and on preemption or hardware failure the persisted WAL tokens let decoding resume or the KV cache be reconstructed.2

Trade-off Avoids the length bias and re-decoding cost of regenerating unfinished requests from scratch.3

How this definition was written

Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.

  1. we implement a token-granular Write-Ahead Log (WAL) for each generation request — DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence
  2. Whenever a new token is generated for a request, we immediately append it to that request’s WAL — DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence
  3. it is mathematically incorrect to regenerate unfinished requests from scratch, as this introduces length bias — DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence

Also called Write-Ahead Log.

Filed under software implementation › infrastructure service