server: fix SWA prompt reuse boundary condition

[1oridevs]

Description

9.4.26 1oridevs

While testing long prompts with SWA-enabled models, I noticed that the server sometimes falls back to full prompt reprocessing even when it looks like the cached state should still be reusable.

After digging into it, it seems to come down to how the boundary is checked:

• The current condition treats pos_min == pos_next - n_swa as invalid
• Checkpoints at that same boundary can also get skipped

So even when the cache is technically still valid, the server ends up not using it.

What this change does

• Changes the reuse condition from:

  pos_min >= pos_min_thold

  to:

  pos_min > pos_min_thold

• Updates the checkpoint lookup to allow:

  cur.pos_min <= pos_min_thold

This basically makes the boundary inclusive instead of exclusive, so we don’t throw away valid reuse cases right at the edge.

Also removed a leftover debug comment while I was there.

Why it matters

Without this, some requests end up doing a full prompt reprocess even though they don’t need to, which can hurt performance — especially with longer contexts.

With this change, reuse behaves more consistently at the sliding window boundary.

---

Notes

Tested locally with the server build — behavior looks correct and avoids unnecessary resets in cases where reuse should be possible.

---

Checklist

• Built and tested locally
• Change is minimal and scoped
• No behavior change outside SWA boundary handling

---

AI usage

I used AI tools to help reason about the issue and review the change, but the debugging, testing, and final code changes were done manually.

[1oridevs]

Ive used cursor AI to help me write this pl description

[ggml-gh-bot]

Hi @1oridevs, thanks for your contribution!

Per our contribution guidelines, the automated PR checker found the following issue(s) that need your attention:

• Multiple open PRs from a new contributor: We limit new contributors (those without a previously merged PR) to 1 open PR at a time. You currently have 2 open PRs.

• AI-generated content: This project does not accept PRs, descriptions or commit messages that are fully or predominantly AI-generated. If you have used AI to assist you in writing code, please make sure to disclose that explicitly.

---

Please note that maintainers reserve the right to make final decisions on PRs. If you believe there is a mistake, please comment below.