Your local LLM forgets its cache after every restart. I measured how much time that wastes.
vimalnakrani
• build_tag string | cache_type_k string | cache_type_v string | created_at timestamp[s] | flash_attn string | llamacpp_commit string | model_sha256 string | n_saved_tokens int64 | probe_hash string | probe_n_predict int64 | probe_sampler_json string | prompt_prefix_sha256 string | save_file string | save_point string | sidecar_version int64 | size_bytes int64 | swa_full bool |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
b9871 | f16 | f16 | 2026-07-06T03:03:16 | on | 65b8fcd92af6b4fefa935c625d1ac27ea29dcb6ee14589c55a8f115ceaaa1423 | 8,219 | b0a33786750c33d04eedb9e8e4b8b8af55aa588651e479035912baee0830dcef | 64 | {"ignore_eos": true, "seed": 42, "temperature": 0, "top_k": 1} | e1405ae5ae1b20ed2b75e3386bed818d054dc74ced0ba80ca4c05d5e617d3df2 | frankenstein_8k_qwen25_7b_f16.bin | doc_prefill | 1 | 471,442,540 | false |
A downloadable KV-cache save file — with the honest math
One llama-server slot save: the first 8,192 Llama-tokens of Frankenstein (public domain), prefilled by Qwen2.5-7B-Instruct Q4_K_M (Apache-2.0 model — chosen over Llama specifically for artifact licensing) and saved with a stillwarm sidecar.
This file is USELESS unless your setup matches the sidecar exactly:
| field | value |
|---|---|
| llama.cpp build | b9871 (ef2d770117db45b05aa7ecd1b0acca36370c5470) — advisory: ±5 weeks measured compatible |
| model | Qwen2.5-7B-Instruct-Q4_K_M.gguf, sha256 65b8fcd92af6b4fefa935c625d1ac27ea29dcb6ee14589c55a8f115ceaaa1423 (bartowski/Qwen2.5-7B-Instruct-GGUF) |
| server flags | -fa on -ctk f16 -ctv f16 (fa mismatch = clean refusal), ctx ≥ 8219 |
| state | 8,219 tokens (doc_prefill — no sampled tokens), 471,442,540 bytes |
| verify probe | 64-token pinned-greedy hash in the sidecar (stillwarm inspect shows it) |
The honest math (measured numbers, Apple M3 Max)
Recomputing this state from the raw text takes ~17.6 s (measured cold-TTFT
median for this model/rung, blockE_low_kv_head_8k_f16.csv in the results dataset). Restoring it from local disk takes ~0.2–0.3 s. But if you
have to download it first (471.4 MB):
| your bandwidth | download time | download + restore | vs recompute (~17.6 s) |
|---|---|---|---|
| 50 Mbps | ~75 s | ~76 s | recompute wins ~4.3× |
| 100 Mbps | ~38 s | ~38 s | recompute wins ~2.2× |
| 215 Mbps | ~17.5 s | ~18 s | ≈ break-even |
| 1 Gbps | ~3.8 s | ~4 s | download wins ~4× |
The symmetric statement: below ≈ 215 Mbps, recomputing wins; above it, downloading wins. That threshold is just the two measured inputs divided: 471,442,540 bytes (this file) ÷ 17.6 s (Qwen2.5-7B's measured 8K cold median on an M3 Max) ≈ 26.8 MB/s ≈ 215 Mbps. Faster prefill hardware pushes the threshold up; slower hardware pulls it down. Downloading caches pays on fast links, LAN transfer, or prefill-slow hardware — otherwise local persistence (save once, restore forever) is where the 71–201× wins live.
Author: Vimal Nakrani. Data license CC-BY-4.0. Source text public domain. No model weights included.
- Downloads last month
- 19