llama.cpp context checkpoint 64-token minimum tuning

Facts

• A prompt shorter than 64 tokens gets no checkpoint, so a follow-up that extends it has nothing to restore on a hybrid model. [source]
• Real agent prompts run to thousands of tokens, so the 64 floor is not what breaks them; spacing and boundary placement are. [source]
• Since the PR 22929 / 24176 / 25472 line of work, the tunable that controls how many checkpoints survive is `-cms/--checkpoint-min-step`, where 0 means no minimum. [source]
• 2026-04-26: issue 22384 proposed lowering the minimum to 4 tokens for recurrent and hybrid models; the issue was closed as completed the same day while the maintainer said no server change was needed (use `preserve_thinking`). [source]
• 2026-05: PR 22929 moved creation to the position before the latest user message and avoided periodic mid-prompt checkpoints when that position is known. [source]
• Build b9309 still logged cadence-style creation ("512 tokens since last checkpoint"); b9354 (first bad commit e98cb51) created one checkpoint at the end of the prompt. [source]
• On a 2,800-token prompt, b9309 made checkpoints at 2047, 2299 and 2811 tokens; b9354 made one at 2800, which became invalid when the next prompt was shorter than that position. [source]
• A user tuning `--checkpoint-min-step 512` found it did nothing to prevent full re-processing in b9354; removing the flag did not help either. [source]
• Maintainer reading: no checkpoint change needed, enable `preserve_thinking` (22384). Reporter reading: a short first turn gets no checkpoint even with `preserve_thinking=true`, so a separate bug exists (22384, Tongas test). The test used tiny prompts, so it does not show the floor matters at agent scale. [source]
• Whether the literal `>= 64` test still exists in current master `server-context.cpp`; the 2026-06-25 diff context in issue 24055 shows the min-spacing check but not the 64 test. [source]
• No source gives a measured optimum for `-cms` on a Mac. [source]
• In the 22384 test with `preserve_thinking=true` and no patch, turn 1 created 0 checkpoints because the 64-token threshold was not met, and turn 2 logged `forcing full prompt re-processing` with 75 tokens processed. [source]
• With the patch, turn 1 created a checkpoint and turn 2 logged `restored context checkpoint` and processed 51 tokens. [source]
• The 22384 maintainer reply was "The fix is incorrect. There is nothing to fix in llama-server" and pointed to `preserve_thinking`; the issue state is closed as completed (2026-04-26). [source]
• The README documents `-cms, --checkpoint-min-step N` as minimum spacing between context checkpoints in tokens, default 8192, 0 meaning no minimum, with env `LLAMA_ARG_CHECKPOINT_MIN_SPACING_NT`. [source]
• A b9354 startup line reads `context checkpoints enabled, max = 64, min spacing = 512` for `--ctx-checkpoints 64 --checkpoint-min-step 512`. [source]
• A b9309 log line reads `512 tokens since last checkpoint at 0, creating new checkpoint during processing at position 2300`, with checkpoints created at n_tokens 2048, 2300 and 2812. [source]
• In b9354 the same prompt produced one checkpoint at n_tokens 2800 (`pos_min = 2799, pos_max = 2799, size = 149.626 MiB`), which was erased on the next request. [source]
• Issue 24055 names commit e98cb51 as the first bad commit for this change of behavior. [source]
• A commenter on 24055 attributes the change to PR 22929: checkpoints are created only at the last message, so any edit before it forces a full reprocess. [source]
• PR 22929 lists its design as: extract `message_spans` from chat templates, find the prompt token position before the latest user message, split prompt batching there, create a checkpoint before the latest user input, and avoid periodic mid-prompt checkpoints when that position is known. [source]
• The author of PR 22929 tested with `--ctx-checkpoints 24 --cache-ram 65536 --keep 4096 -b 8192 --parallel 1 --checkpoint-min-step 256 --chat-template-kwargs '{"preserve_thinking":true}'`, a min-step far below the 8192 default. [source]
• The diff posted on 24055 on 2026-06-25 shows the creation guard `slot.prompt.checkpoints.empty() || is_last_user_message || n_tokens_start > slot.prompt.checkpoints.back().n_tokens + params_base.checkpoint_min_step`. [source]
• The ik_llama.cpp fork uses `--ctx-checkpoints-interval 1024` and shares the same two bugs (hybrid restore test and the 64-token creation floor) in issue 1762. [source]
• ik_llama.cpp issue 1762 (Qwen3.6-27B Q8_0, 8 RTX 2080 Ti, 66,293-token turns) measured 195 s of prompt eval per turn at 339.70 tokens per second. [source]
• Inferred: for agent prompts above a few hundred tokens the 64-token floor is irrelevant; tune `-cms` toward the median turn length and keep `-ctxcp` large enough that checkpoints are not all evicted. [source]