jwatch 0.1.0

Jetson Workload Analysis & Thermal Hardware Correlation - CLI-based LLM benchmarking and hardware profiler for NVIDIA Jetson

jWATCH

jWATCH is a command-line benchmarking tool for NVIDIA Jetson.

It runs LLM inference through Ollama and samples hardware metrics in parallel, so you can inspect throughput/latency together with thermal and power behavior.

What it does

• Benchmarks single models (bench)
• Compares multiple models (compare)
• Streams live hardware metrics (monitor)
• Saves benchmark history to SQLite (history)
• Compares two historical runs (diff)
• Exports runs to JSON (export)
• Supports --mock mode for development on non-Jetson machines

Why this exists

On edge devices, raw tokens/sec can be misleading. A model can look slower simply because the board is hotter, power-limited, or memory-constrained. jWATCH captures inference and hardware state together so comparisons are easier to trust.

Installation

python3 -m pip install -U pip
python3 -m pip install .

For development:

python3 -m pip install pytest pytest-asyncio pytest-cov black mypy

Quick start

1. Validate environment

python3 -m edgewatch check

2. Run a quick mock benchmark (works on Mac/Linux)

python3 -m edgewatch bench \
  --model qwen3.5:0.8b \
  --prompt "Hello" \
  --runs 2 \
  --mock

3. Compare two models

python3 -m edgewatch compare \
  --model qwen3.5:0.8b \
  --model qwen:4b \
  --prompt "Write a short summary" \
  --runs 3 \
  --mock

4. Query stored runs

python3 -m edgewatch history --last 10
python3 -m edgewatch diff <run_id_a> <run_id_b>
python3 -m edgewatch export --output report.json --last 10

Running on Jetson vs non-Jetson

• Jetson: run without --mock to use real tegrastats sampling
• Mac/Linux (non-Jetson): use --mock for telemetry
• Ollama: optional in mock mode, required for real inference runs

CLI commands

bench    Benchmark a single model
compare  Compare multiple models
monitor  Monitor hardware metrics
check    Check Ollama + tegrastats availability
history  Show saved benchmark runs
diff     Compare two run IDs
export   Export runs to JSON

Project layout

edgewatch/
  analysis/
  correlation/
  ollama/
  report/
  storage/
  tegrastats/
  utils/
  edgewatch/cli.py
tests/
assets/

Roadmap

• Better throttle event classification and reporting
• Stronger statistical summaries for long benchmark sessions
• Packaging cleanup for straightforward pip install -e .

Contributing

Issues and PRs are welcome. If you plan to add a feature, open an issue first so we can align on scope.

License

MIT — see LICENSE.