Console

Streamlit dashboard for live gateway traffic, optional registry replica counts, and vLLM Slurm log telemetry. Launch with:

literegistry console \
  --logs /path/to/logs \
  --registry redis://login-node:6379 \
  --server-address 127.0.0.1 \
  --server-port 8765

By default the launcher also tries to start an ngrok tunnel to the Streamlit port (skipped cleanly if ngrok is not on PATH). Disable with --ngrok=False.

How it works

gateway *.log  →  listener / seed parse  →  metric queue  →  Streamlit charts
slurm *.log/.err →  vLLM parser (optional panel)
literegistry summary →  registry replica table (optional poll)

It looks for gateway summary lines such as:

Request counts (last 5.0s): Qwen/Qwen3-8B: 7, python: 8
Completion stats (last 5.0s): Qwen/Qwen3-8B: 68 reqs, avg: 14.402s, max: 60.860s

Initial files are tailed from the end after seeding only the latest ~50 lines per file, so you do not replay full historical logs. Newly discovered files are read from the start so startup lines are not missed.

Launcher arguments (literegistry console)

Streamlit bind (either naming style works):

Argument	Default	Meaning
server_address / address	127.0.0.1	Streamlit bind host
server_port / port	8765	Streamlit bind port
ngrok	True	Open public tunnel if ngrok is installed

App options (forwarded after -- to app.py):

Argument	Default	Meaning
logs	derived	One path, or Fire list of log roots to scan
logs_dir	package logs/gateway	Gateway log directory (legacy single-path)
registry	cluster Redis URL	For literegistry summary polling
slurm_logs_dir	package logs/slurmcompose	Slurm / compose log dir
vllm_logs_dir	None	Alias override for vLLM log dir
seed_recent	True	Seed charts from recent tails
poll_seconds	0.5	Log listener poll interval
window	"1 hour"	Gateway chart time window label
refresh	True	Auto-refresh UI
refresh_seconds	5	UI refresh period
poll_registry	True	Shell out to literegistry summary
registry_poll_seconds	5	Registry poll period
show_vllm	True	Show vLLM telemetry panel
vllm_newest_files	80	Max newest vLLM log files to consider
vllm_tail_lines	1000	Tail depth per vLLM file
vllm_window	"2 hours"	vLLM chart window label

Examples

# Minimal local UI
literegistry console --ngrok=False --server-port 8765

# Point at your job logs + registry
literegistry console \
  --logs /mmfs1/.../logs \
  --registry redis://klone-login03.hyak.local:6379 \
  --server-address 127.0.0.1 \
  --server-port 8765

# Multiple roots
literegistry console \
  --logs='["/path/gateway","/path/slurmcompose"]' \
  --registry redis://login-node:6379

From a checkout without the entrypoint:

streamlit run literegistry/console/app.py \
  --server.address 127.0.0.1 --server.port 8765 -- \
  --logs /path/to/logs --registry redis://host:6379

Needs Python 3.9+ with Streamlit and pandas (see literegistry/console/requirements.txt).

What you see

• Request / completion rates, backlog pressure, avg & max latency
• Per-model and tool (python, etc.) breakdowns
• Registry replica counts (from literegistry summary)
• vLLM throughput, KV-cache, and request pressure from Slurm logs
• Recent parsed raw events and listener/queue status

Window labels map to seconds ("5 min" → 300, "all parsed" → no cutoff).

Ops checklist

1. Gateway logging to a directory the console can read.
2. Same --registry the cluster uses, if you want replica counts.
3. For remote viewing on HPC: either SSH tunnel to server_port, or leave ngrok=True with a configured ngrok binary.

Next: Gateway · docs index