Skip to content

server: reprocess instead of aborting after failing to restore unified KV cache checkpoint - #23723

Closed
NickM-27 wants to merge 2 commits into
ggml-org:masterfrom
NickM-27:fix-kv-error
Closed

server: reprocess instead of aborting after failing to restore unified KV cache checkpoint#23723
NickM-27 wants to merge 2 commits into
ggml-org:masterfrom
NickM-27:fix-kv-error

Conversation

@NickM-27

@NickM-27 NickM-27 commented May 26, 2026

Copy link
Copy Markdown
Contributor

Overview

This PR updates the context checkpoint functions to return a boolean if the checkpoint was able to be restored, allowing falling back to full reprocessing instead of aborting.

Additional information

This fixes #23720 which I provided more information and reproduction steps for.

I understand there may be other reasons for the abort, so perhaps this should be fixed somewhere else to avoid getting to that point. It does seem preferable though to fully reprocess vs abort.

Requirements

  • I have read and agree with the contributing guidelines
  • AI usage disclosure: yes, used AI to troubleshoot the crash and gain an initial understanding of this area of the code

@NickM-27
NickM-27 requested review from a team as code owners May 26, 2026 14:10
@NickM-27 NickM-27 changed the title Fix backend crash after failing to restore unified KV cache checkpoint server: fix backend crash after failing to restore unified KV cache checkpoint May 26, 2026
@ggerganov ggerganov self-assigned this May 26, 2026
@NickM-27 NickM-27 changed the title server: fix backend crash after failing to restore unified KV cache checkpoint server: reprocess instead of aborting after failing to restore unified KV cache checkpoint May 28, 2026
@ggerganov

Copy link
Copy Markdown
Member

Superseded by #23981

@ggerganov ggerganov closed this Jun 2, 2026
@NickM-27
NickM-27 deleted the fix-kv-error branch July 27, 2026 14:19
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Eval bug: Backend crash due to fragmented unified KV cache

2 participants