llama-cpp
Runs LLM inference on CPU, Apple Silicon, and consumer GPUs without NVIDIA hardware. Use for edge deployment, M1/M2/M3 Macs, AMD/Intel GPUs, or when CUDA is unavailable. Supports GGUF quantization (1.5-8 bit) for reduced memory and 4-10× speedup vs PyTorch on CPU.
Install this skill
Works with Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, OpenCode, and 70+ other harnesses.
npx skills add KalarisLabs/research-agent-skills --skill llama-cppuvx research-agent-skills install llama-cppDocumentation & Instructions
llama.cpp
Pure C/C++ LLM inference with minimal dependencies, optimized for CPUs and non-NVIDIA hardware.
When to use llama.cpp
Use llama.cpp when:
- Running on CPU-only machines
- Deploying on Apple Silicon (M1/M2/M3/M4)
- Using AMD or Intel GPUs (no CUDA)
- Edge deployment (Raspberry Pi, embedded systems)
- Need simple deployment without Docker/Python
Use TensorRT-LLM instead when:
- Have NVIDIA GPUs (A100/H100)
- Need maximum throughput (100K+ tok/s)
- Running in datacenter with CUDA
Use vLLM instead when:
- Have NVIDIA GPUs
- Need Python-first API
- Want PagedAttention
Quick start
Installation
# macOS/Linux
brew install llama.cpp
# Or build from source
git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp
make
# With Metal (Apple Silicon)
make LLAMA_METAL=1
# With CUDA (NVIDIA)
make LLAMA_CUDA=1
# With ROCm (AMD)
make LLAMA_HIP=1
Download model
# Download from HuggingFace (GGUF format)
huggingface-cli download \
TheBloke/Llama-2-7B-Chat-GGUF \
llama-2-7b-chat.Q4_K_M.gguf \
--local-dir models/
# Or convert from HuggingFace
python convert_hf_to_gguf.py models/llama-2-7b-chat/
Run inference
# Simple chat
./llama-cli \
-m models/llama-2-7b-chat.Q4_K_M.gguf \
-p "Explain quantum computing" \
-n 256 # Max tokens
# Interactive chat
./llama-cli \
-m models/llama-2-7b-chat.Q4_K_M.gguf \
--interactive
Server mode
# Start OpenAI-compatible server
./llama-server \
-m models/llama-2-7b-chat.Q4_K_M.gguf \
--host 0.0.0.0 \
--port 8080 \
-ngl 32 # Offload 32 layers to GPU
# Client request
curl http://localhost:8080/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "llama-2-7b-chat",
"messages": [{"role": "user", "content": "Hello!"}],
"temperature": 0.7,
"max_tokens": 100
}'
Quantization formats
GGUF format overview
| Format | Bits | Size (7B) | Speed | Quality | Use Case |
|---|---|---|---|---|---|
| Q4_K_M | 4.5 | 4.1 GB | Fast | Good | Recommended default |
| Q4_K_S | 4.3 | 3.9 GB | Faster | Lower | Speed critical |
| Q5_K_M | 5.5 | 4.8 GB | Medium | Better | Quality critical |
| Q6_K | 6.5 | 5.5 GB | Slower | Best | Maximum quality |
| Q8_0 | 8.0 | 7.0 GB | Slow | Excellent | Minimal degradation |
| Q2_K | 2.5 | 2.7 GB | Fastest | Poor | Testing only |
Choosing quantization
# General use (balanced)
Q4_K_M # 4-bit, medium quality
# Maximum speed (more degradation)
Q2_K or Q3_K_M
# Maximum quality (slower)
Q6_K or Q8_0
# Very large models (70B, 405B)
Q3_K_M or Q4_K_S # Lower bits to fit in memory
Hardware acceleration
Apple Silicon (Metal)
# Build with Metal
make LLAMA_METAL=1
# Run with GPU acceleration (automatic)
./llama-cli -m model.gguf -ngl 999 # Offload all layers
# Performance: M3 Max 40-60 tokens/sec (Llama 2-7B Q4_K_M)
NVIDIA GPUs (CUDA)
# Build with CUDA
make LLAMA_CUDA=1
# Offload layers to GPU
./llama-cli -m model.gguf -ngl 35 # Offload 35/40 layers
# Hybrid CPU+GPU for large models
./llama-cli -m llama-70b.Q4_K_M.gguf -ngl 20 # GPU: 20 layers, CPU: rest
AMD GPUs (ROCm)
# Build with ROCm
make LLAMA_HIP=1
# Run with AMD GPU
./llama-cli -m model.gguf -ngl 999
Common patterns
Batch processing
# Process multiple prompts from file
cat prompts.txt | ./llama-cli \
-m model.gguf \
--batch-size 512 \
-n 100
Constrained generation
# JSON output with grammar
./llama-cli \
-m model.gguf \
-p "Generate a person: " \
--grammar-file grammars/json.gbnf
# Outputs valid JSON only
Context size
# Increase context (default 512)
./llama-cli \
-m model.gguf \
-c 4096 # 4K context window
# Very long context (if model supports)
./llama-cli -m model.gguf -c 32768 # 32K context
Performance benchmarks
CPU performance (Llama 2-7B Q4_K_M)
| CPU | Threads | Speed | Cost |
|---|---|---|---|
| Apple M3 Max | 16 | 50 tok/s | $0 (local) |
| AMD Ryzen 9 7950X | 32 | 35 tok/s | $0.50/hour |
| Intel i9-13900K | 32 | 30 tok/s | $0.40/hour |
| AWS c7i.16xlarge | 64 | 40 tok/s | $2.88/hour |
GPU acceleration (Llama 2-7B Q4_K_M)
| GPU | Speed | vs CPU | Cost |
|---|---|---|---|
| NVIDIA RTX 4090 | 120 tok/s | 3-4× | $0 (local) |
| NVIDIA A10 | 80 tok/s | 2-3× | $1.00/hour |
| AMD MI250 | 70 tok/s | 2× | $2.00/hour |
| Apple M3 Max (Metal) | 50 tok/s | ~Same | $0 (local) |
Supported models
LLaMA family:
- Llama 2 (7B, 13B, 70B)
- Llama 3 (8B, 70B, 405B)
- Code Llama
Mistral family:
- Mistral 7B
- Mixtral 8x7B, 8x22B
Other:
- Falcon, BLOOM, GPT-J
- Phi-3, Gemma, Qwen
- LLaVA (vision), Whisper (audio)
Find models: https://huggingface.co/models?library=gguf
References
- Quantization Guide - GGUF formats, conversion, quality comparison
- Server Deployment - API endpoints, Docker, monitoring
- Optimization - Performance tuning, hybrid CPU+GPU
Resources
- GitHub: https://github.com/ggerganov/llama.cpp
- Models: https://huggingface.co/models?library=gguf
- Discord: https://discord.gg/llama-cpp
Agent operating procedure
- Check the environment. Confirm hardware, framework and server versions, model format, and expected load.
- Pin down the inputs. Confirm formats, identifiers and parameters from the data or the user. Ask rather than guess any value that changes the result.
- Run a small version first. Serve or log a single request or run end to end before scaling.
- Execute the full task using the instructions and references above.
- Validate the result. Measure latency, throughput and output correctness against a reference; check resource usage and costs.
- Report. State what was run (versions, commands, parameters), what was checked, and what is still uncertain.
| If this happens | Do this |
|---|---|
| The server fails to start or OOMs | Check model size versus memory, quantization and parallelism settings. |
| A function, flag or endpoint in these instructions is missing in the installed version | Check the installed version's own documentation (help(), --help, official docs), adapt, and tell the user. Never invent an API. |
| A required input, identifier or parameter is ambiguous | Ask the user, or state the assumption explicitly before running. |
Integrity rules
- Never fabricate results, parameters, identifiers, citations or statistics. If something cannot be run or verified, say so plainly.
- Do not expose services or credentials publicly; confirm cloud costs before provisioning.
- Treat version-specific details here as possibly outdated: confirm them against the official documentation for the installed version.
- Ask before actions that cost money, consume shared GPUs or cloud quota, touch personal or patient data, or cannot be undone.
Related skills
gguf-quantization: GGUF format and llama.cpp quantization for efficient CPU/GPU inference.tensorrt-llm: Optimizes LLM inference with NVIDIA TensorRT for maximum throughput and lowest latency.gptq: Quantizes LLMs to 4-bit (also 3-bit) with GPTQ using group-wise quantization (group size 128 by default), via AutoGPTQ and transformers.
Trigger Prompts & Evaluation Corpus
These evaluation prompts test that the agent correctly selects llama-cpp for relevant tasks and leaves it out for near misses.
Should trigger (3)
- should trigger
“How do I run a 7B chat model on my M2 MacBook Air without a GPU server?”
expects Recommends llama.cpp with the Metal backend, a GGUF model at about Q4_K_M, and gives install, download and run commands. Notes the 8-16 GB RAM limit when choosing the model and quantization, and says to check the installed version's --help for the flags.
- should trigger
“I'm on a Ryzen 9 7950X workstation with 64 GB RAM and no GPU. I downloaded Meta-Llama-3-8B-Instruct.Q5_K_M.gguf from Hugging Face and want to serve it locally for a team script. When I start the server with a 32k context it gets very slow and then the process is killed. What should I check and how do I fix it?”
expects Diagnoses memory use from the KV cache at 32k context plus model size, and suggests lowering the context size, a smaller quantization, thread count tuning, or parallel slot settings. Gives llama-server flags, says to start with a small test request first, and confirms the flags against the installed version's --help instead of inventing them.
- should trigger
“Our lab has a few old Raspberry Pi 5 boards and an Intel Arc laptop but no NVIDIA cards. We want a small language model to summarize field notes offline. What's a realistic way to set that up, and what speed should we expect?”
expects Suggests a CPU-efficient C/C++ runtime (llama.cpp) with a small, heavily quantized GGUF model, and Vulkan or SYCL for the Intel GPU. Gives modest speed expectations, says to measure them on the actual hardware instead of quoting the table figures, and states assumptions about model size and RAM.
Near misses (2)
- near miss
“I have an 8xH100 node and need to serve Llama 3 70B at the highest possible throughput with FP8 and in-flight batching for a production API. How should I set it up?”
expects Recognizes this as NVIDIA datacenter serving and points to TensorRT-LLM (or vLLM) rather than llama.cpp, with FP8, in-flight batching and tensor parallelism.
- near miss
“I have a 4-bit GPTQ checkpoint of Mistral 7B and want to load it with AutoGPTQ in transformers and pick between the Marlin and ExLlamaV2 kernels on my RTX 4090. Which should I use?”
expects Identifies this as GPTQ loading and kernel selection in Python with transformers on an NVIDIA GPU, which is the gptq skill, not llama.cpp GGUF inference. Compares the Marlin and ExLlamaV2 kernels.