Bittune
Local agent that deploys and benchmarks vLLM/SGLang inference on your GPUs with auditable evidence.
Tested on Linux
Use it
Linux x86_64
You need first
- Runtimes:
Python >= 3.10,Node.js >= 22.19.0 - System packages:
OpenSSL CLI - A GPU is optional
- Keys or accounts: BITTUNE_AGENT_LLM_API_KEY
step 4 still names /home/sbx inside single quotes, where $HOME would not work: put your own home folder there before running it.
These steps ran from a folder named work in the home folder, which already existed. Make it and run them from there:
mkdir -p "$HOME/work" && cd "$HOME/work"clone Bittune source
git clone --depth 1 https://github.com/bitcloud-ai/Bittune bittune-srcnpm ci plus venv Python Core install
cd bittune-src && npm ci && python3 -m venv .venv && .venv/bin/python -m pip install -e ./packages/bittune-core && .venv/bin/python -c "import bittune_core; print('core ok')"build portable dist bundle
cd bittune-src && npm run build:shellinstall bittune shim on PATH
mkdir -p ~/.local/bin && printf '%s\n' '#!/bin/sh' 'export BITTUNE_CORE_PYTHON="/home/sbx/work/bittune-src/.venv/bin/python"' 'export BITTUNE_TEST_PYTHON="/home/sbx/work/bittune-src/.venv/bin/python"' 'exec node "/home/sbx/work/bittune-src/dist/bittune.js" "$@"' > ~/.local/bin/bittune && chmod +x ~/.local/bin/bittune
Where the README and reality differ
- README: npm ci; python3 -m venv .venv; pip install -e ./packages/bittune-core; npm run bittune
What worked: After npm ci plus venv pip install, also ran npm run build:shell and installed a ~/.local/bin/bittune shim that sets BITTUNE_CORE_PYTHON/BITTUNE_TEST_PYTHON to the venv python and execs node dist/bittune.js; direct tsx entry fails outside the repo with Cannot find package 'tsx'
Verified in pickfit: the commands exactly as its sandbox ran them
For the record only: they name the sandbox's home, /home/sbx.
git clone --depth 1 https://github.com/bitcloud-ai/Bittune bittune-srccd bittune-src && npm ci && python3 -m venv .venv && .venv/bin/python -m pip install -e ./packages/bittune-core && .venv/bin/python -c "import bittune_core; print('core ok')"cd bittune-src && npm run build:shellmkdir -p ~/.local/bin && printf '%s\n' '#!/bin/sh' 'export BITTUNE_CORE_PYTHON="/home/sbx/work/bittune-src/.venv/bin/python"' 'export BITTUNE_TEST_PYTHON="/home/sbx/work/bittune-src/.venv/bin/python"' 'exec node "/home/sbx/work/bittune-src/dist/bittune.js" "$@"' > ~/.local/bin/bittune && chmod +x ~/.local/bin/bittunebittune --helpbittune versionCheck that it works
bittune --helpThe output should contain “Usage:”.
First thing to try
bittune versionRun bittune version in a terminal to confirm the tool starts, then launch bittune to open the interface.
What it does
These come from the README; pickfit didn’t check them one by one.
- Deploy and manage vLLM or SGLang servicesNot tested
- Check your GPU setup and find modelsNot tested
- Benchmark services and save the resultsNot tested
- Record tuning experiments and compare candidatesNot tested
How far pickfit got
Linux x86_64
- Installs
- Starts
- Main function works
- Works offline
Good for
- Deploy Qwen model to vLLM and benchmark throughput
- Inspect GPU machine and generate hardware snapshot
- Compare benchmark results with deterministic deltas
How pickfit tested it
Each run starts from a clean machine and follows the steps above. How pickfit verifies
Linux x86_64
Its main function works
- Replayed from scratch
- Works offline
- Checked on 30 Oct 2026
- Version 0.1.0
Differences the test machine hit
- README: tar -xzf bittune-<version>-linux-x86_64.tar.gz; cd bundle; sudo ./install.sh
In the test: Skipped ranked sudo ./install.sh and installed from source because the box has no root/sudo and install.sh installs to /usr/local/bin and /opt/bittune - README: export BITTUNE_AGENT_LLM_API_KEY='your-api-key'; bittune configure --base-url https://endpoint.example.com/v1 --model-id your-tool-capable-model; bittune doctor
In the test: bittune doctor exits 0 with a dummy OpenAI-compatible base-url and BITTUNE_AGENT_LLM_API_KEY=dummy, reporting Python Core and Runtime assembled without any real endpoint, Docker, GPU, or EvalScope