Hybrid GatedDeltaNet state rollback and checkpoint prefix caching

Facts

• llama.cpp's hybrid memory is two sub-caches: an attention KV cache and a recurrent cache with one cell per sequence. [source]
• A checkpoint stores only the part that cannot be rebuilt by truncation (`LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY`); the attention KV prefix stays in the slot. [source]
• The recurrent cache also keeps a short ring of per-token snapshots (`n_rs_seq`), used when speculative decoding rejects part of a draft. [source]
• Restore picks the latest checkpoint at or before the shared prefix and reprocesses from there. [source]
• 2026-04-26: issue 22384 shows the restore test never succeeds for recurrent state. [source]
• 2026-06: PRs 24176, 24411 and 25472 repair placement, erasure and eviction; PR 24797 (draft) targets the `seq_pos_min()` root cause. [source]
• ik_llama.cpp keeps the same logic and the same failure (issue 1762, May 2026). [source]
• Speculative decoding on hybrid models uses checkpoints: the server logs `speculative decoding will use checkpoints` at load. Partial acceptance beyond `n_rs_seq` used to fail `seq_rm` and leave the attention cache with stale entries. [source]
• The recurrent snapshot is not precise after a fall-through rollback: the model reprocesses from the last valid checkpoint. [source]
• The hybrid `seq_pos_min()` was written as `max()` of both sub-caches; the PR 24797 author argues it should return only the attention cache's minimum, as non-hybrid memory does. Whether other hybrid types (Mamba-based) should follow is untested. [source]
• Merge state and effect of PR 24797 on the residual misses reported after 24411 and 24176. [source]
• Whether Metal's behavior differs; every cited log is CUDA or ROCm. [source]
• PR 24797 states hybrid memory consists of two independent cache subsystems with different positional semantics: for the attention cache `pos_min` is the earliest cached KV position, and for the recurrent cache `pos_min == pos_max ==` the current position. [source]
• PR 24797 says the old `seq_pos_min()` returned the `std::max()` of both caches' minimums, which made a cached prefix look absent and forced full reprocessing on any prompt variation (stripped thinking, injected timestamp). [source]
• PR 24797 keeps `seq_pos_max` with intersection semantics because it bounds valid attention entries by where the recurrent state has advanced. [source]
• PR 24797 notes other hybrid types with recurrent memory would probably benefit from the same `seq_pos_min()` change but were not tested. [source]
• On hybrid models, when speculative decoding partially accepts draft tokens and the rollback distance exceeds `n_rs_seq`, the pre-24797 recurrent cache returned false from `seq_rm`, leaving the attention cache with stale entries. [source]
• Particula states the bounded recurrent rewind succeeds only when `cell.pos - (p0 - 1)` is between 1 and `n_rs_seq` (context type `SEQ_RM_TYPE_RS`). [source]
• Hybrid checkpoints in 24055 logs satisfy `pos_min == pos_max == n_tokens - 1` (for example 2799, 2799, 2800 tokens), so each stores one recurrent position. [source]
• A 27B hybrid checkpoint measured 149.626 MiB at 2,800 tokens and 151.055 MiB at 364 tokens, so size is nearly independent of prompt length. [source]
• The load log of a Qwen3.6-27B MTP model reads `speculative decoding will use checkpoints` followed by `common_speculative_init: no implementations specified for speculative decoding` when no draft type is given. [source]
• ik_llama.cpp issue 1762 (qwen3next, Gated DeltaNet) reports the same restore-test failure and 64-token creation floor as llama.cpp 22384 and says the same fix is needed in the fork's server checkpoint code. [source]
• Reporters in 24055 use `--spec-type draft-mtp` with `--spec-draft-n-max 3` (and 6) on hybrid Qwen3.6 models while relying on checkpoints, with 85% draft acceptance reported on one 27B run. [source]
• Inferred: restore cost on a hybrid model is the token gap between the resume point and the nearest earlier checkpoint, so placing a checkpoint at each user-message boundary (PR 24176) bounds typical re-prefill to one turn's new text plus the last assistant turn. [source]