🏎️ Dyno — llama.cpp Auto-Tuner & Benchmark

Dyno is an open-source CLI that auto-tunes and benchmarks llama.cpp / ik_llama.cpp inference on NVIDIA, AMD (ROCm), and Apple Silicon GPUs, producing reproducible, shareable results.

pipx install llama-dyno

dyno detect
dyno tune ~/models/mistral-7b.Q4_K_M.gguf --quick
dyno bench ~/models/mistral-7b.Q4_K_M.gguf --ngl 99 --fa
dyno bench --lmstudio llama-3.2-3b  # benchmark an LM Studio model
dyno report ~/models/mistral-7b.Q4_K_M.gguf
dyno submit ~/models/mistral-7b.Q4_K_M.gguf

30-Second Quickstart

# 1. Install
pipx install llama-dyno

# 2. Check your hardware
dyno detect

# 3. Auto-tune a model
dyno tune ~/Downloads/my-model.q4_k_m.gguf

# 4. Run the final benchmark
dyno bench ~/Downloads/my-model.q4_k_m.gguf --ngl 99 --fa

# 5. Generate a report
dyno report ~/Downloads/my-model.q4_k_m.gguf

Prerequisites

• Python 3.11+
• A GPU: NVIDIA (drivers + CUDA), AMD (ROCm, via rocm-smi), or Apple Silicon (M-series, uses Metal + unified memory)
• llama-bench binary from llama.cpp or ik_llama.cpp — built with the matching backend (CUDA or Metal)

Install llama.cpp:

brew install llama.cpp                        # macOS / Linux
# or build from source:
git clone https://github.com/ggml-org/llama.cpp
cd llama.cpp && cmake -B build && cmake --build build --config Release

ik_llama.cpp (optional, for MoE models):

git clone https://github.com/ikawrakow/ik_llama.cpp
cd ik_llama.cpp && cmake -B build && cmake --build build --config Release
# Symlink the ik binaries
ln -sf $(pwd)/build/bin/ik_llama-bench ~/.local/bin/

Commands

dyno detect

Fingerprint hardware: GPU model, VRAM, driver/CUDA version, CPU, RAM, and detect which llama.cpp backend is installed (with commit hash).

dyno tune <model.gguf>

Find the fastest config for your GPU + model.

Flag	Default	Description
--quick	(default)	~10 trials, fast iteration
--thorough		~25 trials, best accuracy

Search strategy:

1. Validate model loads correctly
2. Coarse sweep — ngl (layer offload), flash attention on/off, KV cache quant (f16 / q8_0 / q4_0)
3. Hill-climb — batch size (128–4096), threads (auto / half / full cores)
4. ik_llama.cpp extras — -fmoe, -rtr, -amb toggles (thorough only)

OOM configs are discarded gracefully with ngl backoff. Live progress table shows every trial.

ik_llama.cpp Support

Dyno automatically detects ik_llama.cpp when ik_llama-bench is on your PATH. Extra flags tuned:

Flag	Description	Best for
-fmoe	Fast MoE computation	MoE models (Mixtral, DeepSeek)
-rtr	Runtime tensor reorder	Memory-bandwidth-bound scenarios
-amb	Attention memory bound	Large context / many KV heads

Dyno detects which flags your build supports and only searches those. MoE models get fmoe enabled by default; Dense models skip MoE flags entirely.

LM Studio Support

Dyno can benchmark models served by LM Studio through its OpenAI-compatible API (http://localhost:1234/v1).

# Show LM Studio status in hardware detection
dyno detect

# Benchmark a loaded model
dyno bench --lmstudio llama-3.2-3b

# Tune and create a report
dyno tune --lmstudio llama-3.2-3b --json-out report.json

LM Studio does not expose engine-reported tok/s — throughput is measured client-side from the streamed SSE response (wall-clock). Results are approximate and may vary with system load.

Flag	Default	Description
--lmstudio		Benchmark an LM Studio model
--host	http://localhost:1234/v1	LM Studio API base URL

dyno bench <model.gguf>

Run a specific config 3× and report median tokens/sec with variance.

Flag	Default	Description
--ngl	99	GPU layers to offload
--fa/--no-fa	true	Flash attention
--ctk	f16	K cache quant (f16, q8_0, q4_0)
--ctv	f16	V cache quant (f16, q8_0, q4_0)
--batch	512	Batch size
--ubatch	512	Micro batch size
--threads	0 (auto)	Thread count
--runs	3	Number of benchmark runs
--fmoe	false	Fast MoE (ik_llama.cpp)
--rtr	false	Runtime reorder (ik_llama.cpp)
--amb	false	Attention memory bound (ik_llama.cpp)

dyno report <model.gguf>

Generate a shareable report including:

• Full hardware fingerprint
• Model details (name, quant, SHA-256)
• Winning config
• Median scores with variance
• Reproducible llama-server command
• Shareable markdown snippet
• JSON output

dyno submit <model.gguf>

Submit your results to the community results repo (llama-dyno-results):

1. Opens a PR via gh CLI
2. Falls back to a GitHub Gist
3. Saves locally if neither works

Quality & Reproducibility

• Fixed bench params: Every run uses pp=512 / tg=128 (fixed prompt/gen tokens) so results are comparable
• 3-run median + variance reported
• No fabricated numbers — real hardware detection, real subprocess results
• OOM handling — graceful backoff of ngl
• Clear install hints — suggests brew or git clone when binaries missing

Results Table (Example)

GPU	Model	Quant	Backend	TG tok/s
RTX 4090	Mixtral-8x7B	Q4_K_M	ik_llama.cpp	58.7
RTX 4090	Llama-3-70B	Q4_K_M	ik_llama.cpp	42.3
RTX 4090	Llama-3-70B	Q4_K_M	llama.cpp	35.1
RTX 3090	Mistral-7B	Q4_K_M	llama.cpp	112.8
RTX 4060	Phi-3-mini	Q4_K_M	llama.cpp	68.5

Shell Completions

Dyno supports shell completions via Typer/Click:

# Install completions for your shell
dyno --install-completion

# Show completion script (to manually install)
dyno --show-completion

Supported shells: bash, zsh, fish, powershell.

Architecture

src/llama_dyno/
├── __init__.py    # Package metadata
├── cli.py         # Typer CLI (detect, tune, bench, report, submit)
├── types.py       # Data types (BenchParams, TrialResult, etc.)
├── detect.py      # Hardware fingerprinting (pynvml, nvidia-smi, /proc)
├── bench.py       # llama-bench subprocess driver + parser
├── tune.py        # Search strategy (coarse sweep + hill climb)
├── ollama.py      # Ollama REST API runner
├── lmstudio.py    # LM Studio OpenAI-compatible API runner
├── report.py      # Report generation (JSON + markdown)
└── submit.py      # GitHub PR / Gist submission

Development

git clone https://github.com/lachy/llama-dyno
cd llama-dyno
pip install -e ".[dev]"
pytest

Out of Scope (v1)

• GUI
• Multi-GPU
• Intel backends (planned)
• Server hosting for results

License

Apache 2.0