Move llama.cpp chat and local embeddings onto a verified externally managed llama-server runtime. Remove the in-process native runtime, forked embedding workers, and node-llama-cpp dependency while preserving guided setup, local GGUF models, tool-capable agent runs, diagnostics, and operator docs.
6.0 KiB
summary, read_when, title, sidebarTitle
| summary | read_when | title | sidebarTitle | ||||
|---|---|---|---|---|---|---|---|
| Run local GGUF chat and memory embeddings with an OpenClaw-managed llama.cpp server |
|
llama.cpp Provider | llama.cpp Provider |
The llama-cpp plugin manages a loopback-only llama-server for local GGUF
chat and embeddings. OpenClaw installs a pinned, integrity-verified llama.cpp
release, starts it only when a request needs it, reuses it across concurrent
chat and embedding requests, and stops it after an idle period.
Install the official plugin before using either local inference or local memory embeddings:
openclaw plugins install @openclaw/llama-cpp-provider
Guided setup
Choose llama.cpp once during interactive onboarding or configuration. OpenClaw then:
- Selects the verified llama-server build for the Gateway platform.
- Verifies the archive SHA-256 and the extracted server version.
- Downloads and verifies the default chat and embedding models after consent.
- Writes a durable OpenAI-compatible provider with a loopback
baseUrlandlocalServiceprocess definition. - Live-tests the candidate before saving it.
The default chat model remains:
hf:unsloth/gemma-4-E4B-it-GGUF/gemma-4-E4B-it-Q4_K_M.gguf
Gemma 4 E4B IT Q4_K_M is about 5.0 GB. OpenClaw offers that download only on machines with at least 16 GiB of RAM. The default context cap remains 8,192 tokens. The bundled EmbeddingGemma model is about 0.3 GB.
Discovery is read-only. It reports a prepared choice only when the managed binary, server preset, and selected model already exist; it never installs or downloads during discovery.
How requests run
The provider uses OpenClaw's normal OpenAI-compatible chat, image, streaming,
and tool transport. llama-server applies the GGUF chat template; OpenClaw
executes tool calls and returns their results to the model. The existing
llamacpp-gbnf tool-schema compatibility profile remains enabled.
One managed router owns separate presets for chat and embeddings. This lets
memory.search.provider: "local" use a dedicated embedding GGUF through
/v1/embeddings without creating a second OpenClaw process supervisor.
Use another GGUF model
Add the model to models.providers.llama-cpp.models, set params.modelPath,
make it the selected llama-cpp/<model-id>, then run interactive llama.cpp
setup again. modelPath accepts:
- an absolute or
~/local GGUF path; - a cache-relative GGUF filename;
- a full
hf:file URI, including#branchwhen needed; - an HTTPS GGUF URL that publishes a SHA-256 response digest.
Example model entry:
{
id: "my-local-model",
name: "My local GGUF",
reasoning: false,
input: ["text"],
cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
contextWindow: 8192,
maxTokens: 2048,
params: {
modelPath: "~/Models/my-model.Q4_K_M.gguf",
contextSize: 8192,
},
compat: { supportsTools: true },
}
The default managed cache is ~/.openclaw/models/llama.cpp. Existing
modelCacheDir settings still win, and setup recognizes the former
~/.node-llama-cpp/models default cache so upgrades do not redownload a model
that is already present.
Local memory embeddings
Set the memory provider to local:
{
memory: {
search: {
provider: "local",
local: {
modelPath: "hf:ggml-org/embeddinggemma-300m-qat-q8_0-GGUF/embeddinggemma-300m-qat-Q8_0.gguf",
},
},
},
}
The plugin preserves the historical local provider/model cache identity, so
the transport migration does not require a SQLite schema change or automatic
memory reindex. A custom embedding modelPath remains its literal index
identity. Run openclaw memory status --index if you intentionally change it.
Diagnostics
Run:
openclaw memory status --deep
openclaw doctor
After the managed embedding server has handled a request, deep status reports
facts observed from /health, /models, /props, and /metrics: server build,
model id and path, endpoint state, and configured capabilities. Vision is
reported only when /props confirms it. Draft and multimodal projector support
are not currently configured by this plugin and are never inferred from a model
name.
Local-service startup and exit logs include bounded, redacted stderr tails. See Logging and Local model services.
Platform requirements
- macOS arm64 uses the official Metal build. macOS x64 uses the CPU build.
- Linux x64 needs glibc 2.34 or newer; Linux arm64 needs glibc 2.38 or newer.
Install the OpenMP runtime (
libgomp1on Debian/Ubuntu orlibgompon Fedora) if the version probe reportslibgomp.so.1missing. - Windows x64 and arm64 use the CPU build and require the Microsoft Visual C++ 2015-2022 Redistributable.
- Alpine/musl and platforms without a pinned official build fail with an actionable manual-server path rather than silently skipping setup.
OpenClaw intentionally does not auto-select CUDA, ROCm, SYCL, OpenVINO, or Vulkan archives. Those builds add driver and companion-runtime contracts that cannot be verified safely from onboarding alone.
Troubleshooting
Binary missing or wrong version: run interactive llama.cpp setup again. It
reinstalls the pinned build and rewrites the absolute localService.command.
Model missing: configure a local GGUF path or rerun setup and approve the verified default download.
Server starts but the model fails to load: inspect openclaw logs --follow
and openclaw memory status --deep. The managed service error includes the
bounded server stderr tail.
Only keyword memory matches: run openclaw memory status --deep, repair the
reported endpoint/model issue, then run openclaw memory index --force only if
status reports an index identity mismatch.