Record: llama.cpp version review — 2026-08-26

The owner rebuilt llama.cpp from the latest GitHub source and restarted both routers. Before: beast 0.2.0-dev (10628), athena 10129. After: beast 0.3.0-dev (10657, a58d57cea), athena 0.3.0-dev (10643, d6cf9b091) — the athena version gap is closed.

Verified unchanged

• The tuned presets work as-is on the new builds: /v1/models still lists 13 (beast) / 15 (athena) models with the aliases and --parallel 4 (checked on beast).
• --spec-type option list unchanged (draft-simple/eagle3/mtp/dflash/dspark/ngram-*); gpt-oss native MTP still auto-enables. --draft-max was renamed --spec-draft-n-max (no config impact).
• KV cache quantization (-ctk/-ctv) and --fit unchanged.

New in this build — changed recommendations

1. --cache-reuse N is now available (KV shifting reuse across requests; “requires prompt caching”, default 0). This is the headline change: it lets repeated request prefixes (DSH’s system prompt, aider’s injected rules) reuse cached KV instead of reprocessing. The 120B’s prompt processing measured 2.4 t/s; cache-reuse directly attacks that. Recommendation: add cache-reuse = 256 under [*] on both hosts (not yet applied — pending owner approval).
2. --parallel default is now auto (-1) instead of 1. The explicit parallel = 4 in the tuned configs still overrides it; no change needed.
3. --n-cpu-moe N (keep first N MoE layers’ weights in CPU) is new — relevant only to the MoE model gemma-4-26B-A4B on the 16 GB MI25, as an alternative to naive layer spill. Optional experiment; not recommended as a default.

Standing recommendations (unchanged)

• KV cache quant (cache-type-k/v = q8_0) — the memory lever for gemma-26b at 81920 ctx; speed/quality tradeoff for the owner.
• EAGLE-3 drafter trial for gemma-26b; ngram for the CPU 70Bs (both optional).
• Avoid IQ quants on the MI25 (Vulkan int dot: 0).

Verification

• Alias routing verified on the new build: coder32 chat completion answered correctly (2026-08-26).
• Captures refreshed; make test passes.