Command-line interface

The PLLM command-line interface inspects public metadata, runs bounded local benchmarks, and starts the gateway, inference, and preparation service roles. Service commands do not imply production orchestration or make incomplete compiled plans executable.

Install the command

uv tool install pllm.run
pllm --help

This installs the pllm.run distribution from PyPI in an isolated tool environment. Use uv tool upgrade pllm.run to update it. The installed command remains pllm.

Choose a task

• Run private inference through the trusted gateway, then connect an SDK or raw HTTP client.
• Operate provider roles as independently configured inference and preparation services.
• Benchmark and diagnose with the headless runner or loopback dashboard.
• Inspect configuration and components without starting a model.

Each guide links to exact parser-generated command help under Command reference.

Available commands

• pllm config validates and exports public configuration.
• pllm components lists built-in component descriptions.
• pllm benchmark run runs the real client, preparation, and inference roles on loopback and returns a text-free diagnostic record.
• pllm gateway starts the trusted loopback API gateway and can co-locate both provider roles for development.
• pllm serve starts one inference or preparation role from explicit deployment configuration.
• pllm dev contains development-only tools.

Run pllm COMMAND --help for the options accepted by a command. Use --format json when another program will consume the result.

Inspect configuration

Validate a public configuration or export its canonical form:

pllm config show examples/pllm.yaml
pllm config export examples/pllm.yaml --output experiment.json

See config show and config export.

Inspect components

List component descriptors or inspect one stable component identity:

pllm components list --format json
pllm components show pllm/cpu --format json

See components list and components show.

Run a benchmark

The default model is Qwen/Qwen2.5-0.5B-Instruct:

pllm benchmark run \
  --prompt-file prompt.txt \
  --max-output-tokens 24 \
  --repetitions 3 \
  --output benchmark.json

The command starts and stops all three roles. Use --tiny only for a random-weight transport smoke test. Reports exclude prompts, generated text, token IDs, activation payloads, masks, seeds, and credentials. They are single-host diagnostics, not canonical research EvidenceReport records. See benchmark run.

Open the development dashboard

The dashboard is development-only. --tiny uses random weights for a local transport smoke test:

pllm dev dashboard --tiny

See dev dashboard.

Continue

• CLI reference lists every current command and option.
• Local gateway gives the exact public serving commands and config boundaries.
• SDK configuration explains configuration files and content identity.
• Research papers and the research backlog document source status and unfinished work.