dgxarley 0.0.49

Integration testing, streaming utilities, and repetition detection for distributed LLM inference on DGX Spark clusters

dgxarley

Tooling for the DGX Arley K3s inference cluster — integration tests, streaming utilities, and CLI entry points for SGLang, Ollama, and OpenWebUI services.

Heureka! — Qwen3-235B-A22B MoE (AWQ 4-bit) running distributed inference across both DGX Sparks:

sglang-raw — Dual-panel SSE stream viewer

Dual-panel Rich TUI for inspecting SGLang's OpenAI-compatible streaming API in real time. The top half renders the AI response as it arrives, while the bottom half displays the raw JSON SSE stream chunks — showing fields like chat_completion_chunk, choices, delta, finish_reason, and model. Useful for debugging streaming behaviour, verifying token delivery, and understanding the wire format of the API.

sglang-raw — Think/text token classification

Token-level stream inspection with per-chunk breakdown in a structured table. Columns show the token type (think vs text), content, finish reason, and cumulative token count — visualizing how reasoning tokens (from <think>...</think> blocks) are separated from the actual output tokens. This view helps when tuning thinking budgets, verifying reasoning_parser behaviour, or diagnosing unexpected token classification.

What's included

CLI tools

Command	Description
sglang-raw	Interactive SSE stream viewer with dual-panel Rich display (interpreted output + raw JSON chunks)
sglang-test	Direct SGLang client with sequential and parallel load testing (live Rich TUI)
sglang-bench	Benchmark launcher for SGLang with persistent dataset caching and concurrency-sweep mode
openwebui-test	OpenWebUI / LLM client with preset management and streaming
ollama-test	Ollama API health, model, embedding, and chat completions tests
comfyui-test	ComfyUI integration tests for the local image-generation playground
kceve-kvm	RS232 serial control for KCEVE KVM1001A KVM switches — 4-port and 10-port variants, --ports/KCEVE_KVM_PORTS (switch ports, query state, sniff)
kceve-kvm-web	Web UI for KCEVE KVM control (FastAPI, requires dgxarley[web])
kceve-kvm-web-plain	Lightweight web UI for KCEVE KVM control (stdlib http.server, no extra dependencies)

kceve-kvm-web — KCEVE KVM1001A Web UI

Browser-based control panel for the KCEVE KVM1001A KVM switch via RS232 serial. Shows the currently active input port on a virtual 7-segment display and allows switching between inputs with a single click. The port count defaults to 10 and is selectable for the 4-port variant via -n/--ports (or the KCEVE_KVM_PORTS env var). Commands are sent over a USB-to-RS232 adapter at 115200 baud using the X<channel>,1$ ASCII protocol.

Demo video

Remote test setup (serial over SSH tunnel)

If the KVM is connected to a remote host, you can tunnel the serial port via TCP:

# 1. On k3smaster (where the USB-RS232 adapter is connected): expose serial as TCP server
root@k3smaster ~ # socat tcp-listen:7000,reuseaddr,fork /dev/ttyACM0,b115200,raw,echo=0

# 2. On workstation: SSH tunnel to remote TCP port
user@workstation ~ $ ssh -N -L 7000:127.0.0.1:7000 root@k3smaster &

# 3. On workstation: create local PTY from TCP tunnel
user@workstation ~ $ socat pty,link=/tmp/kvm-serial,raw,echo=0 tcp:127.0.0.1:7000 &

# 4. On workstation: start web UI on the local PTY
user@workstation ~ $ kceve-kvm-web -d /tmp/kvm-serial -p 8080

Then open http://localhost:8080 in a browser.

Libraries

Module	Description
dgxarley.integration.repetition_detector	Offline n-gram, sentence, and loop repetition analysis for completed LLM outputs
dgxarley.integration.streaming_repetition_guard	Real-time repetition detection for token streams with configurable thresholds

Installation

pip install dgxarley

Quick start

from dgxarley.integration.repetition_detector import detect_repetition

report = detect_repetition(llm_output)
print(report.summary())
# [LOW] score=0.12 — N-Gram 'this is a test' x2

from dgxarley.integration.streaming_repetition_guard import RepetitionGuard

guard = RepetitionGuard()
for chunk in llm_stream:
    token = chunk.choices[0].delta.content or ""
    result = guard.feed(token)
    if result.should_stop:
        print(f"STOP: {result.reason}")
        break

Requirements

• Python >= 3.14

Source & documentation

Full documentation, network architecture, and Ansible playbooks: GitHub

License

This project is licensed under the LGPL where applicable/possible — see LICENSE.md. Some files/parts may use other licenses: MIT | GPL | LGPL. Always check per‑file headers/comments.

Authors

• Repo owner (primary author)
• Additional attributions are noted inline in code comments

Acknowledgments

• Inspirations and snippets are referenced in code comments where appropriate.

⚠️ Note

This is a development/experimental project. For production use, review security settings, customize configurations, and test thoroughly in your environment. Provided "as is" without warranty of any kind, express or implied, including but not limited to the warranties of merchantability, fitness for a particular purpose and noninfringement. In no event shall the authors or copyright holders be liable for any claim, damages or other liability, whether in an action of contract, tort or otherwise, arising from, out of or in connection with the software or the use or other dealings in the software. Use at your own risk.