Repository navigation
server: save the draft state in the slot file - #30574
Merged
Merged
Conversation
The slot file held the target context and the context checkpoints, but never the draft context, so a restored slot drafted with an empty or stale draft KV until the cache refilled. The file now ends with the live draft state, the full draft context and the speculative state, in the same appendix style as the checkpoints. A draft state that does not fit the draft context is dropped and the draft starts empty, while older files and older servers keep working.
Contributor
Author
|
I think the file plumbing (open for append, flush and error check on save, open and seek to the offset on load) could be shared between the checkpoint appendix and the draft block with two small helpers, but I left the checkpoint code from #26004 untouched to keep this change focused, happy to factor it if you prefer. |
Member
Yes, we can improve plumbing in general. Something we should refactor in follow-ups - for now it's nice to patch the gaps. |
ggerganov
approved these changes
Oct 15, 2026
ggerganov
added
the
merge ready
axh-xecoy
added a commit
to axh-xecoy/llama.cpp
that referenced
this pull request
Oct 21, 2026
… start With --slot-save-path set, the live slot is written to auto-slot0.bin when the server stops and read back after the model loads, so a long session resumes with a prefix hit instead of a full re-prefill. Measured on a 5900-token prompt: prompt_n 5900 -> 516 after a restart. The draft memory goes to an auto-slot0.drft sidecar. Without it a restored slot keeps an empty draft and accepts far fewer speculative tokens (llama.cpp ggml-org#30574, merged upstream one day after this tree). The target file keeps the /slots save format byte for byte, so a snapshot taken here can be restored through the REST API and vice versa. Anything that does not fit the running context is a warning plus a cold start.
Garretsw
pushed a commit
to Garretsw/llama.cpp
that referenced
this pull request
Oct 26, 2026
Upstream master now persists context checkpoints (and the draft state) in slot save files itself (ggml-org#26004, ggml-org#30574), and load_tgt/load_dft return false instead of aborting. Resolve the conflicts in favour of master's implementation: - common: take master's bool load_tgt/load_dft and drop the now redundant try_load_tgt/try_load_dft - server: drop the branch's file-scope "CKPC" appendix writer/reader in favour of master's save/load_slot_checkpoints + save/load_slot_draft, and master's checkpoint restore fallback (failed slot-file checkpoints force a full re-process and are invalidated) - keep reporting n_ckpt_saved / n_ckpt_restored in the slot save/restore responses and the slot_checkpoints /props flag Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0186dV229bL7yh4joD2nXXAF
auRose94
pushed a commit
to auRose94/llama.cpp
that referenced
this pull request
Oct 28, 2026
The slot file held the target context and the context checkpoints, but never the draft context, so a restored slot drafted with an empty or stale draft KV until the cache refilled. The file now ends with the live draft state, the full draft context and the speculative state, in the same appendix style as the checkpoints. A draft state that does not fit the draft context is dropped and the draft starts empty, while older files and older servers keep working. [no release]
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Overview
Follow-up to #30514, fixes #28619 (thanks @dgabehar for the report).
A slot restored with --slot-save-path gets back the main model memory but not the draft model memory, so speculative decoding restarts with an empty draft and accepts far fewer tokens.
The slot file now also stores the draft memory, the same way the RAM prompt cache already does, and restores it. A draft that does not match is skipped and starts empty, and old slot files and old servers keep working.
Additional information
Draft acceptance rate on the turn after save, erase and restore (the output is identical in all cases):
test_slot_restore_draft_state fails without this PR.
Requirements