llama.cpp-style CLI example

Path: example/llamadart_cli · Platforms: Dart CLI on macOS, Linux and Windows

A terminal chat CLI that accepts llama.cpp-style arguments, with harnesses that compare its output against upstream llama.cpp.

Run#

cd example/llamadart_cli
dart pub get
dart run bin/llamadart_cli.dart --model /path/to/model.gguf

Variants:

# Download from Hugging Face into ./models, with Unsloth's GLM settings
dart run bin/llamadart_cli.dart \
  -hf unsloth/GLM-4.7-Flash-GGUF:UD-Q4_K_XL \
  --jinja --ctx-size 16384 \
  --temp 1.0 --top-p 0.95 --min-p 0.01 --fit on

# One prompt from a file, in llama.cpp simple-io style
dart run bin/llamadart_cli.dart --model /path/to/model.gguf \
  --file prompt.txt --simple-io

# List every flag
dart run bin/llamadart_cli.dart --help

What it demonstrates#

• llama.cpp flag names and aliases (-c, -ngl, -n, --top_p) mapped onto ModelParams and GenerationParams (Runtime parameters).
• -hf repo[:file-hint] resolution and download into a local model folder (Downloads and cache).
• Interactive streaming chat with a thinking stream and slash commands (Generation and streaming).
• --fit on, which trims old turns and caps the output to the remaining context.
• Transcript parity against llama-cli and tool-call parity against llama-server, run through example/llamadart_server (Tool calling).

Test#

cd example/llamadart_cli
dart test

The real-model parity gates are tagged local-only and need local llama.cpp builds and model files.

Full options: parity harness setup, environment overrides and CLI notes are in the example README.