lmprompt

#!/usr/bin/env -S java --enable-native-access=ALL-UNNAMED --source 25 -cp ./zbo/lightmetal.jar:./lightmetal.jar

import lm.catalog.boundary.ModelCatalog;
import lm.configuration.control.ZCfg;
import lm.configuration.entity.GenerationConfig;
import lm.generation.boundary.LightMetal;
import lm.generation.entity.Tps;

String name = MethodHandles.lookup().lookupClass().getName();
String version = "2026-06-10.1";
String DIM = "\033[2m";
String RESET = "\033[0m";

void main(String... args) throws Exception {
    IO.println(name + " " + version);
    ZCfg.load("lightmetal");
    var fileName = args.length > 0 ? matchModel(args[0]) : ZCfg.string("model");
    if (fileName == null) {
        System.err.println(name + ": no model configured (set model=<file> in ~/.lightmetal/app.properties)");
        System.exit(1);
    }
    var modelPath = ModelCatalog.resolve(fileName);
    if (!Files.exists(modelPath)) {
        System.err.println(name + ": model not found at " + modelPath);
        System.exit(1);
    }
    var label = fileName.replace(".gguf", "") + "> ";
    var prompt = System.console().readLine(label);
    if (prompt == null || prompt.isBlank()) {
        System.err.println(name + ": empty prompt");
        System.exit(1);
    }

    var count = new AtomicLong();
    var thoughtCount = new AtomicLong();
    var firstTokenNanos = new AtomicLong();

    try (var lm = LightMetal.load(modelPath)) {
        var modelName = lm.metadata().name().orElse(modelPath.getFileName().toString());
        try (var stream = lm.generate(prompt, GenerationConfig.fromProperties())) {
            stream.forEach(t -> {
                firstTokenNanos.compareAndSet(0L, System.nanoTime());
                count.incrementAndGet();
                if ("thought".equals(t.channel())) {
                    thoughtCount.incrementAndGet();
                    System.err.print(DIM + t.text() + RESET);
                    System.err.flush();
                    return;
                }
                IO.print(t.text());
                System.out.flush();
            });
        }
        IO.println("");

        if (count.get() < 2) {
            System.err.println(name + ": got " + count.get() + " tokens, need at least 2 for tps");
            System.exit(1);
        }

        var thought = thoughtCount.get() > 0 ? ", " + thoughtCount.get() + " thought" : "";
        IO.println("[" + modelName + ": " + Tps.measure(count.get(), firstTokenNanos.get()) + thought + "]");
    }
}

String matchModel(String fragment) {
    var matches = ModelCatalog.search(fragment);
    if (matches.size() == 1) {
        return matches.getFirst();
    }
    if (matches.isEmpty()) {
        System.err.println(name + ": no model matches \"" + fragment + "\"");
    } else {
        System.err.println(name + ": " + matches.size() + " models match \"" + fragment + "\":");
        matches.forEach(m -> System.err.println("  " + m));
    }
    System.exit(1);
    return null;
}