docs/bitnet.md

Kinetic + BitNet (bitnet.cpp)

Use Microsoft bitnet.cpp as Kinetic’s Language Reasoner for 1.58-bit LLM inference on CPU (GPU kernels available upstream).

Why this fits “light + high performance”

Property	How Kinetic uses it
1.58-bit weights	Ternary BitNet; official lossless kernels
CPU speedups	Upstream reports ~1.4–6×; large energy savings
Event-only	BitNet runs on new instruction / replan, never in the control loop
Compact context	Prompt = system rules + WorldState.summary_for_llm()
Fallback	Rule reasoner if BitNet is missing

Control / policy / vision stay separate — BitNet only emits intent JSON.

---

Two official scripts (do not confuse -p)

Script	Purpose	-p means	-n means
run_inference.py	Real text generation / chat	Text prompt (string)	Tokens to generate
utils/e2e_benchmark.py	Throughput / latency bench	Prompt token count (int, default 512)	Generated tokens (default 128)

Kinetic reasoning uses run_inference.py.
Kinetic bitnet-bench uses e2e_benchmark.py.

---

Local repo (this project)

Kinetic expects BitNet at:

F:\Kinetic\BitNet

Lightweight defaults: short JSON prompts, n_predict=64, ctx_size=1024, prefer Falcon-E-1B (then 0.7B large, then 2B).

One-shot setup (recommended)

# Prefer VS2022 Developer PowerShell (clang + MSVC toolset)
cd F:\Kinetic
powershell -ExecutionPolicy Bypass -File scripts\setup_bitnet_light.ps1
# Optional: -Model BitNetLarge | BitNet2B

This runs submodule init, setup_env.py for a small HF model, cmake build, and points configs/bitnet.yaml at the GGUF.

Official install (bitnet.cpp) — manual

Requirements: Python ≥ 3.9, CMake ≥ 3.22, Clang ≥ 18.
Windows: always use a VS 2022 Developer Command Prompt / PowerShell.

cd F:\Kinetic\BitNet
git submodule update --init --recursive
pip install huggingface_hub
# Lightest instruct option
python setup_env.py -hr tiiuae/Falcon-E-1B-Instruct -md models -q i2_s -p

Inference smoke test (run_inference.py)

python run_inference.py -m models/BitNet-b1.58-2B-4T/ggml-model-i2_s.gguf -p "You are a helpful assistant" -cnv

Convert from .safetensors (bf16 checkpoint)

huggingface-cli download microsoft/bitnet-b1.58-2B-4T-bf16 --local-dir ./models/bitnet-b1.58-2B-4T-bf16
python ./utils/convert-helper-bitnet.py ./models/bitnet-b1.58-2B-4T-bf16

---

Official e2e benchmark (utils/e2e_benchmark.py)

Setup the environment for running inference benchmarks.

Arguments

Flag	Meaning	Default
-m, --model	Required. Path to model GGUF	—
-n, --n-token	Number of generated tokens	128
-p, --n-prompt	Number of prompt tokens (not text!)	512
-t, --threads	Thread count	2
-h, --help	Help	—

Examples

# Real model
python utils/e2e_benchmark.py -m /path/to/model.gguf -n 200 -p 256 -t 4

# Dummy model (kernel layout not in a public HF model)
python utils/generate-dummy-bitnet-model.py models/bitnet_b1_58-large `
  --outfile models/dummy-bitnet-125m.tl1.gguf --outtype tl1 --model-size 125M

python utils/e2e_benchmark.py -m models/dummy-bitnet-125m.tl1.gguf -p 512 -n 128

Via Kinetic CLI (wrapper)

# After setting KINETIC_BITNET_REPO / model_path
python -m kinetic.cli bitnet-bench -m F:\BitNet\models\BitNet-b1.58-2B-4T\ggml-model-i2_s.gguf -n 128 -p 512 -t 4

# Or from config
python -m kinetic.cli bitnet-bench --config configs/bitnet.yaml -n 200 -p 256 -t 8

Use benchmark numbers to pick threads / model size for replan latency, not control-loop Hz.

---

Wire into Kinetic

Config (configs/bitnet.yaml)

language:
  backend: bitnet
  bitnet:
    repo_dir: F:/BitNet
    model_path: F:/BitNet/models/BitNet-b1.58-2B-4T/ggml-model-i2_s.gguf
    threads: 8
    n_predict: 128
    temperature: 0.1
    strict: false

Environment

$env:KINETIC_BITNET_REPO = "F:\BitNet"
$env:KINETIC_BITNET_MODEL = "F:\BitNet\models\BitNet-b1.58-2B-4T\ggml-model-i2_s.gguf"

Run

cd F:\Kinetic
python -m pip install -e ".[dev]"

python -m kinetic.cli bitnet-info --config configs/bitnet.yaml
python -m kinetic.cli reason -i "Pick up the blue bottle" --config configs/bitnet.yaml --json
python -m kinetic.cli bitnet-bench --config configs/bitnet.yaml -n 128 -p 256 -t 4
python -m kinetic.demos.pick_bottle --config configs/bitnet.yaml

---

Architecture placement

User instruction
      │
      ▼  (event only)
 BitNet b1.58  (run_inference.py)  →  GoalSpec JSON
      │
      ▼
 Task Planner → World Model → Policy → Controller → Robot

Never call BitNet every control tick.

---

FAQ (from upstream BitNet + Kinetic notes)

Q1: Build dies in llama.cpp log.cpp / std::chrono?

Known llama.cpp integration issue. Check BitNet GitHub discussions for the referenced commit fix (upstream BitNet FAQ Q1).

Q2: Clang / conda on Windows — 'clang' is not recognized?

Your shell is not initialized for VS tools. Verify:

clang -v

Command Prompt:

"C:\Program Files\Microsoft Visual Studio\2022\Professional\Common7\Tools\VsDevCmd.bat" -startdir=none -arch=x64 -host_arch=x64

(Use Community or Enterprise instead of Professional if that is your edition.)

PowerShell:

Import-Module "C:\Program Files\Microsoft Visual Studio\2022\Professional\Common7\Tools\Microsoft.VisualStudio.DevShell.dll"
Enter-VsDevShell -SkipAutomaticLocation -DevCmdArguments "-arch=x64 -host_arch=x64"

Then rebuild BitNet from that shell.

Q3: Kinetic falls back to rules without error?

Set language.bitnet.strict: true, or run kinetic reason ... and inspect bitnet_error / backend.

Q4: Benchmark works but reason returns garbage?

• Bench only measures kernels (-p is token count).
• Reasoning needs a real instruct GGUF + run_inference.py with a text -p.
• Keep temperature low (0.1) and ask for JSON only (Kinetic’s prompt already does).

---

Performance tuning for Kinetic

Knob	Suggested for replan	Notes
Model	2B (BitNet-b1.58-2B-4T)	0.7B if RAM-limited
threads	physical cores	Measure with bitnet-bench -t …
n_predict / -n	64–128	Short JSON only
Replan rate	< 1–2 Hz events	Not 20 Hz control
Bench -p	256–512	Approximates world-summary size

---

References

• microsoft/BitNet — bitnet.cpp
• utils/e2e_benchmark.py, utils/generate-dummy-bitnet-model.py, utils/convert-helper-bitnet.py
• Technical reports in the BitNet README (1-bit AI Infra, BitNet b1.58, GPU kernels)