The milkwashy command

milkwashy uses Fire. Each command accepts -- --help, which shows all of its flags:

milkwashy run -- --help

Two flags work for all commands:

Flag	Default	Function
--project_dir	.	The directory that holds 01-indata/ and 02-testllm/.
--verbose	True	Shows the log messages. Give --verbose=False to hide them.

Configuration commands

milkwashy init_config          # write the default TOML configuration
milkwashy init_config --force   # replace an existing file
milkwashy config_path           # print the location of the file

Input commands

milkwashy gen_indata            # write the four example texts to 01-indata/

The texts are in English, Polish, German and Arabic. They come from the package, so the command needs no network. The file names hold the language code, for example en__en-rivers-and-machines.txt.

Model commands

milkwashy models                        # list the models of the default server
milkwashy models --provider cliproxy    # list the models of one other server
milkwashy lms_status                    # show the LM Studio state
milkwashy lms_unload_all                # unload all models, and confirm it

The list is in order of size, from the smallest model to the largest. --limit N in the run command then always selects the same models.

Plan command

milkwashy plan

This command shows the chunks and the prompt for each input text. It sends nothing to a model, so it is fast and it costs nothing. Use it to see the plan before a long sweep.

Sweep command

milkwashy run                        # all models of the default server
milkwashy run --limit 5              # the 5 smallest models
milkwashy run --max_size_mb 8000     # skip models above 8 GB
milkwashy run --provider chat2api    # one other server
milkwashy run --manage_lms=False     # do not load or unload models
milkwashy run --force                # write the output files again

Flag	Default	Function
--provider	the default in the configuration	Which server to use.
--limit	no limit	Use only the N smallest models.
--max_size_mb	no limit	Skip each model that is larger than this size.
--manage_lms	True	Let milkwashy load and unload LM Studio models.
--force	False	Make the requests again, and replace the output files.

The sweep writes to 02-testllm/<server>/<model>/<text>.txt, and it writes the plan next to each output as <text>.plan.json.

A sweep can continue. If an output file is present, milkwashy does not make the request again. If all of the output files of a model are present, milkwashy does not load that model. Give --force to ignore this rule.

All servers together

The sweep.sh script in the repository starts one process for each server. The processes run at the same time.

./sweep.sh                        # all servers
./sweep.sh --limit 5              # the 5 smallest models of each server
PROVIDERS="lmstudio" ./sweep.sh   # one server only
PARALLEL=0 ./sweep.sh             # one server after the other

Each process writes its own log file to 02-testllm/_logs/<server>.log. Use tail -f to watch a log.

Only one LM Studio sweep can run on a computer. The script uses a lock directory, because two sweeps would load and unload models at the same time. A second LM Studio sweep stops with a message. The other servers in that command continue.

From Python

from pathlib import Path

import milkwashy

text = "Rivers move. Machines stop. Nobody remembers the first winter."

seed = milkwashy.text_seed(text)
for plan in milkwashy.plan_chunks(text, "en"):
    print(plan.index, plan.prompt_index, plan.text)

config = milkwashy.load_config()
provider = config.providers[config.default_provider]
result = milkwashy.run_sweep(
    config,
    provider,
    project_dir=Path("."),
    limit=1,
)
print(result)

From Rust

use milkwashr::{assemble_chunks, text_seed};

let sentences: Vec<String> = vec!["Rivers move.".into(), "Machines stop.".into()];
let seed = text_seed(&sentences.join(" "));
let chunks = assemble_chunks(&sentences, seed, 2, 8, 6000);