DeepSeek Harness - Save KV-Cache
Save and load KV cache slots on a llama.cpp server from within DeepSeek Harness (DSH).
0.2.0 adds a slot capability probe (auto-save goes dormant on servers without
--slot-save-pathinstead of posting blind 404s, with honest command errors) and bundle metadata for the DSH Plugins panel ("Save KV-Cache" under Personal, with the standard toggle).
Features
/save-kv-cache— Save the current KV cache slot on the active llama.cpp server assession-<id>.bin. Download icon./load-kv-cache— Restore a saved KV cache file on the active llama.cpp server. Data icon.- Auto-save — Fires on every
turn/endevent and saves slot0assession-<id>.bin. - Capability probe — Before saving, the plugin probes
POST /slots/0?action=restore(10-minute per-server cache). On servers without--slot-save-path(501not_supported_erroror a router-style 404) auto-save stays dormant — no blind POST traffic at every turn — and the manual commands answer with the fix ("start llama-server with--slot-save-path") instead of a raw404 The model 'None' does not exist. - Plugin panel — Ships
dsh.bundle.patch+locale/en|zh.json+icon.svg, so Plugins lists it under Personal as "Save KV-Cache" (保存 KV 缓存) with the same enable/disable toggle as the built-in plugins.
How It Works
Save
POST /slots/{slot_id}?action=save
Body: { "filename": "session-<id>.bin" }
The llama.cpp server writes the KV cache from the specified slot into --slot-save-path.
Load
POST /slots/{slot_id}?action=restore
Body: { "filename": "session-<id>.bin" }
The server restores the saved sequence state into the slot. The command reports
n_restored, n_read, and restore_ms from the server response.
Provider selection
Commands and auto-save resolve the active provider from the session's latest
request/header (or model/selection event), then read that provider's
baseURL from llm-pi-ai.providers. This is important when more than one
llama.cpp server is configured: saving against the wrong server writes an
unrelated slot cache. The first configured provider is used only when no
selection exists yet.
Commands
| Command | Description |
|---|---|
/save-kv-cache |
Save KV cache as session-<id>.bin |
/load-kv-cache |
Restore KV cache from session-<id>.bin |
| Auto-save | Saves slot 0 on every turn/end using the active provider |
Filenames
The code tries session.title, then falls back to String(session.id). In DSH
0.2.0-rc.2 the title is a session event (session/title), not a property on the
session object, so the effective filename is the session ID (e.g.
session-<session-uuid>.bin).
Requirements
- DeepSeek Harness (DSH)
0.2.0-rc.2+ - llama.cpp server with
--slot-save-pathconfigured - Plugin installed in the profile's
node_modules(see INSTALL.md)
Installation
From npm:
dsh plugin --profile <profile> add @residenthiago/dsh-save-kv-cache
Device-specific deployment (manual copy/symlink, profile bundle wiring): see INSTALL.md.
Repository layout
The npm package lives in plugin/ (package.json, lib/,
locale/, bundle patch, package README/LICENSE, test/); this root holds the
full documentation — this README and INSTALL.md.
No comments yet. Be the first to write one.