node-llama-cpp
Version:
Run AI models locally on your machine with node.js bindings for llama.cpp. Enforce a JSON schema on the model output on the generation level
534 lines • 25.7 kB
JavaScript
import * as readline from "readline";
import process from "process";
import path from "path";
import chalk from "chalk";
import fs from "fs-extra";
import prettyMilliseconds from "pretty-ms";
import { getLlama } from "../../bindings/getLlama.js";
import { LlamaLogLevel, LlamaLogLevelGreaterThan, nodeLlamaCppGpuOptions, parseNodeLlamaCppGpuOption } from "../../bindings/types.js";
import { LlamaCompletion } from "../../evaluator/LlamaCompletion.js";
import withOra from "../../utils/withOra.js";
import { TokenMeter } from "../../evaluator/TokenMeter.js";
import { printInfoLine } from "../utils/printInfoLine.js";
import { printCommonInfoLines } from "../utils/printCommonInfoLines.js";
import { resolveCommandGgufPath } from "../utils/resolveCommandGgufPath.js";
import { withProgressLog } from "../../utils/withProgressLog.js";
import { resolveHeaderFlag } from "../utils/resolveHeaderFlag.js";
import { withCliCommandDescriptionDocsUrl } from "../utils/withCliCommandDescriptionDocsUrl.js";
import { documentationPageUrls } from "../../config.js";
import { ConsoleInteraction, ConsoleInteractionKey } from "../utils/ConsoleInteraction.js";
import { DraftSequenceTokenPredictor } from "../../evaluator/LlamaContext/tokenPredictors/DraftSequenceTokenPredictor.js";
export const CompleteCommand = {
command: "complete [modelPath]",
describe: withCliCommandDescriptionDocsUrl("Generate a completion for a given text", documentationPageUrls.CLI.Complete),
builder(yargs) {
return yargs
.option("modelPath", {
alias: ["m", "model", "path", "url", "uri"],
type: "string",
description: "Model file to use for the completion. Can be a path to a local file or a URI of a model file to download. Leave empty to choose from a list of recommended models"
})
.option("header", {
alias: ["H"],
type: "string",
array: true,
description: "Headers to use when downloading a model from a URL, in the format `key: value`. You can pass this option multiple times to add multiple headers."
})
.option("gpu", {
type: "string",
// yargs types don't support passing `false` as a choice, although it is supported by yargs
choices: nodeLlamaCppGpuOptions,
coerce: (value) => {
if (value == null || value == "")
return undefined;
return parseNodeLlamaCppGpuOption(value);
},
defaultDescription: "Uses the latest local build, and fallbacks to \"auto\"",
description: "Compute layer implementation type to use for llama.cpp. If omitted, uses the latest local build, and fallbacks to \"auto\""
})
.option("systemInfo", {
alias: "i",
type: "boolean",
default: false,
description: "Print llama.cpp system info"
})
.option("text", {
type: "string",
description: "First text to automatically start generating completion for"
})
.option("textFile", {
type: "string",
description: "Path to a file to load text from and use as the first text to automatically start generating completion for"
})
.option("contextSize", {
alias: "c",
type: "number",
description: "Context size to use for the model context",
default: -1,
defaultDescription: "Automatically determined based on the available VRAM"
})
.option("batchSize", {
alias: "b",
type: "number",
description: "Batch size to use for the model context. The default value is the context size"
})
.option("flashAttention", {
alias: "fa",
type: "boolean",
default: false,
description: "Enable flash attention"
})
.option("threads", {
type: "number",
defaultDescription: "Number of cores that are useful for math on the current machine",
description: "Number of threads to use for the evaluation of tokens"
})
.option("temperature", {
alias: "t",
type: "number",
default: 0,
description: "Temperature is a hyperparameter that controls the randomness of the generated text. It affects the probability distribution of the model's output tokens. A higher temperature (e.g., 1.5) makes the output more random and creative, while a lower temperature (e.g., 0.5) makes the output more focused, deterministic, and conservative. The suggested temperature is 0.8, which provides a balance between randomness and determinism. At the extreme, a temperature of 0 will always pick the most likely next token, leading to identical outputs in each run. Set to `0` to disable."
})
.option("minP", {
alias: "mp",
type: "number",
default: 0,
description: "From the next token candidates, discard the percentage of tokens with the lowest probability. For example, if set to `0.05`, 5% of the lowest probability tokens will be discarded. This is useful for generating more high-quality results when using a high temperature. Set to a value between `0` and `1` to enable. Only relevant when `temperature` is set to a value greater than `0`."
})
.option("topK", {
alias: "k",
type: "number",
default: 40,
description: "Limits the model to consider only the K most likely next tokens for sampling at each step of sequence generation. An integer number between `1` and the size of the vocabulary. Set to `0` to disable (which uses the full vocabulary). Only relevant when `temperature` is set to a value greater than 0."
})
.option("topP", {
alias: "p",
type: "number",
default: 0.95,
description: "Dynamically selects the smallest set of tokens whose cumulative probability exceeds the threshold P, and samples the next token only from this set. A float number between `0` and `1`. Set to `1` to disable. Only relevant when `temperature` is set to a value greater than `0`."
})
.option("seed", {
type: "number",
description: "Used to control the randomness of the generated text. Only relevant when using `temperature`.",
defaultDescription: "The current epoch time"
})
.option("gpuLayers", {
alias: "gl",
type: "number",
description: "number of layers to store in VRAM",
default: -1,
defaultDescription: "Automatically determined based on the available VRAM"
})
.option("repeatPenalty", {
alias: "rp",
type: "number",
default: 1.1,
description: "Prevent the model from repeating the same token too much. Set to `1` to disable."
})
.option("lastTokensRepeatPenalty", {
alias: "rpn",
type: "number",
default: 64,
description: "Number of recent tokens generated by the model to apply penalties to repetition of"
})
.option("penalizeRepeatingNewLine", {
alias: "rpnl",
type: "boolean",
default: true,
description: "Penalize new line tokens. set `--no-penalizeRepeatingNewLine` or `--no-rpnl` to disable"
})
.option("repeatFrequencyPenalty", {
alias: "rfp",
type: "number",
description: "For n time a token is in the `punishTokens` array, lower its probability by `n * repeatFrequencyPenalty`. Set to a value between `0` and `1` to enable."
})
.option("repeatPresencePenalty", {
alias: "rpp",
type: "number",
description: "Lower the probability of all the tokens in the `punishTokens` array by `repeatPresencePenalty`. Set to a value between `0` and `1` to enable."
})
.option("maxTokens", {
alias: "mt",
type: "number",
default: 0,
description: "Maximum number of tokens to generate in responses. Set to `0` to disable. Set to `-1` to set to the context size"
})
.option("tokenPredictionDraftModel", {
alias: ["dm", "draftModel"],
type: "string",
description: "Model file to use for draft sequence token prediction (speculative decoding). Can be a path to a local file or a URI of a model file to download"
})
.option("tokenPredictionModelContextSize", {
alias: ["dc", "draftContextSize", "draftContext"],
type: "number",
description: "Max context size to use for the draft sequence token prediction model context",
default: 4096
})
.option("debug", {
alias: "d",
type: "boolean",
default: false,
description: "Print llama.cpp info and debug logs"
})
.option("meter", {
type: "boolean",
default: false,
description: "Log how many tokens were used as input and output for each response"
})
.option("timing", {
type: "boolean",
default: false,
description: "Print how how long it took to generate each response"
})
.option("noMmap", {
type: "boolean",
default: false,
description: "Disable mmap (memory-mapped file) usage"
})
.option("printTimings", {
alias: "pt",
type: "boolean",
default: false,
description: "Print llama.cpp's internal timings after each response"
});
},
async handler({ modelPath, header, gpu, systemInfo, text, textFile, contextSize, batchSize, flashAttention, threads, temperature, minP, topK, topP, seed, gpuLayers, repeatPenalty, lastTokensRepeatPenalty, penalizeRepeatingNewLine, repeatFrequencyPenalty, repeatPresencePenalty, maxTokens, tokenPredictionDraftModel, tokenPredictionModelContextSize, debug, meter, timing, noMmap, printTimings }) {
try {
await RunCompletion({
modelPath, header, gpu, systemInfo, text, textFile, contextSize, batchSize, flashAttention,
threads, temperature, minP, topK, topP, seed, gpuLayers, lastTokensRepeatPenalty,
repeatPenalty, penalizeRepeatingNewLine, repeatFrequencyPenalty, repeatPresencePenalty, maxTokens,
tokenPredictionDraftModel, tokenPredictionModelContextSize, debug, meter, timing, noMmap, printTimings
});
}
catch (err) {
await new Promise((accept) => setTimeout(accept, 0)); // wait for logs to finish printing
console.error(err);
process.exit(1);
}
}
};
async function RunCompletion({ modelPath: modelArg, header: headerArg, gpu, systemInfo, text, textFile, contextSize, batchSize, flashAttention, threads, temperature, minP, topK, topP, seed, gpuLayers, lastTokensRepeatPenalty, repeatPenalty, penalizeRepeatingNewLine, repeatFrequencyPenalty, repeatPresencePenalty, tokenPredictionDraftModel, tokenPredictionModelContextSize, maxTokens, debug, meter, timing, noMmap, printTimings }) {
if (contextSize === -1)
contextSize = undefined;
if (gpuLayers === -1)
gpuLayers = undefined;
const headers = resolveHeaderFlag(headerArg);
if (debug)
console.info(`${chalk.yellow("Log level:")} debug`);
const llamaLogLevel = debug
? LlamaLogLevel.debug
: LlamaLogLevel.warn;
const llama = gpu == null
? await getLlama("lastBuild", {
logLevel: llamaLogLevel
})
: await getLlama({
gpu,
logLevel: llamaLogLevel
});
const logBatchSize = batchSize != null;
const useMmap = !noMmap && llama.supportsMmap;
const resolvedModelPath = await resolveCommandGgufPath(modelArg, llama, headers, {
flashAttention,
useMmap
});
const resolvedDraftModelPath = (tokenPredictionDraftModel != null && tokenPredictionDraftModel !== "")
? await resolveCommandGgufPath(tokenPredictionDraftModel, llama, headers, {
flashAttention,
useMmap,
consoleTitle: "Draft model file"
})
: undefined;
if (systemInfo)
console.log(llama.systemInfo);
if (textFile != null && textFile !== "") {
if (text != null && text !== "")
console.warn(chalk.yellow("Both `text` and `textFile` were specified. `textFile` will be used."));
text = await fs.readFile(path.resolve(process.cwd(), textFile), "utf8");
}
if (batchSize != null && contextSize != null && batchSize > contextSize) {
console.warn(chalk.yellow("Batch size is greater than the context size. Batch size will be set to the context size."));
batchSize = contextSize;
}
let initialText = text ?? null;
const model = await withProgressLog({
loadingText: chalk.blue.bold("Loading model"),
successText: chalk.blue("Model loaded"),
failText: chalk.blue("Failed to load model"),
liveUpdates: !debug,
noProgress: debug,
liveCtrlCSendsAbortSignal: true
}, async (progressUpdater) => {
try {
return await llama.loadModel({
modelPath: resolvedModelPath,
gpuLayers: gpuLayers != null
? gpuLayers
: contextSize != null
? { fitContext: { contextSize } }
: undefined,
defaultContextFlashAttention: flashAttention,
useMmap,
ignoreMemorySafetyChecks: gpuLayers != null,
onLoadProgress(loadProgress) {
progressUpdater.setProgress(loadProgress);
},
loadSignal: progressUpdater.abortSignal
});
}
catch (err) {
if (err === progressUpdater.abortSignal?.reason)
process.exit(0);
throw err;
}
finally {
if (llama.logLevel === LlamaLogLevel.debug) {
await new Promise((accept) => setTimeout(accept, 0)); // wait for logs to finish printing
console.info();
}
}
});
const draftModel = resolvedDraftModelPath == null
? undefined
: await withProgressLog({
loadingText: chalk.blue.bold("Loading draft model"),
successText: chalk.blue("Draft model loaded"),
failText: chalk.blue("Failed to load draft model"),
liveUpdates: !debug,
noProgress: debug,
liveCtrlCSendsAbortSignal: true
}, async (progressUpdater) => {
try {
return await llama.loadModel({
modelPath: resolvedDraftModelPath,
defaultContextFlashAttention: flashAttention,
useMmap,
onLoadProgress(loadProgress) {
progressUpdater.setProgress(loadProgress);
},
loadSignal: progressUpdater.abortSignal
});
}
catch (err) {
if (err === progressUpdater.abortSignal?.reason)
process.exit(0);
throw err;
}
finally {
if (llama.logLevel === LlamaLogLevel.debug) {
await new Promise((accept) => setTimeout(accept, 0)); // wait for logs to finish printing
console.info();
}
}
});
const draftContext = draftModel == null
? undefined
: await withOra({
loading: chalk.blue("Creating draft context"),
success: chalk.blue("Draft context created"),
fail: chalk.blue("Failed to create draft context"),
useStatusLogs: debug
}, async () => {
try {
return await draftModel.createContext({
contextSize: { max: tokenPredictionModelContextSize }
});
}
finally {
if (llama.logLevel === LlamaLogLevel.debug) {
await new Promise((accept) => setTimeout(accept, 0)); // wait for logs to finish printing
console.info();
}
}
});
const context = await withOra({
loading: chalk.blue("Creating context"),
success: chalk.blue("Context created"),
fail: chalk.blue("Failed to create context"),
useStatusLogs: debug
}, async () => {
try {
return await model.createContext({
contextSize: contextSize != null ? contextSize : undefined,
batchSize: batchSize != null ? batchSize : undefined,
threads: threads === null ? undefined : threads,
ignoreMemorySafetyChecks: gpuLayers != null || contextSize != null,
performanceTracking: printTimings
});
}
finally {
if (llama.logLevel === LlamaLogLevel.debug) {
await new Promise((accept) => setTimeout(accept, 0)); // wait for logs to finish printing
console.info();
}
}
});
const draftContextSequence = draftContext?.getSequence();
const contextSequence = draftContextSequence != null
? context.getSequence({
tokenPredictor: new DraftSequenceTokenPredictor(draftContextSequence)
})
: context.getSequence();
const completion = new LlamaCompletion({
contextSequence
});
let lastDraftTokenMeterState = draftContextSequence?.tokenMeter.getState();
let lastTokenMeterState = contextSequence.tokenMeter.getState();
let lastTokenPredictionsStats = contextSequence.tokenPredictions;
await new Promise((accept) => setTimeout(accept, 0)); // wait for logs to finish printing
const padTitle = await printCommonInfoLines({
context,
draftContext,
useMmap,
minTitleLength: "Complete".length + 1,
logBatchSize,
tokenMeterEnabled: meter
});
printInfoLine({
title: "Complete",
padTitle: padTitle,
info: [{
title: "Repeat penalty",
value: `${repeatPenalty} (apply to last ${lastTokensRepeatPenalty} tokens)`
}, {
show: repeatFrequencyPenalty != null,
title: "Repeat frequency penalty",
value: String(repeatFrequencyPenalty)
}, {
show: repeatPresencePenalty != null,
title: "Repeat presence penalty",
value: String(repeatPresencePenalty)
}, {
show: !penalizeRepeatingNewLine,
title: "Penalize repeating new line",
value: "disabled"
}, {
show: timing,
title: "Response timing",
value: "enabled"
}]
});
// this is for ora to not interfere with readline
await new Promise((resolve) => setTimeout(resolve, 1));
const replHistory = [];
async function getPrompt() {
const rl = readline.createInterface({
input: process.stdin,
output: process.stdout,
history: replHistory.slice()
});
const res = await new Promise((accept) => rl.question(chalk.yellow("> "), accept));
rl.close();
return res;
}
while (true) {
const input = initialText != null
? initialText
: await getPrompt();
if (initialText != null) {
console.log(chalk.green("> ") + initialText);
initialText = null;
}
else
await replHistory.push(input);
if (input === ".exit")
break;
process.stdout.write(chalk.yellow("Completion: "));
const [startColor, endColor] = chalk.blue("MIDDLE").split("MIDDLE");
const abortController = new AbortController();
const consoleInteraction = new ConsoleInteraction();
consoleInteraction.onKey(ConsoleInteractionKey.ctrlC, async () => {
abortController.abort();
consoleInteraction.stop();
});
const timeBeforePrompt = Date.now();
try {
process.stdout.write(startColor);
consoleInteraction.start();
await completion.generateCompletion(input, {
temperature,
minP,
topK,
topP,
seed: seed ?? undefined,
signal: abortController.signal,
repeatPenalty: {
penalty: repeatPenalty,
frequencyPenalty: repeatFrequencyPenalty != null ? repeatFrequencyPenalty : undefined,
presencePenalty: repeatPresencePenalty != null ? repeatPresencePenalty : undefined,
penalizeNewLine: penalizeRepeatingNewLine,
lastTokens: lastTokensRepeatPenalty
},
maxTokens: maxTokens === -1
? context.contextSize
: maxTokens <= 0
? undefined
: maxTokens,
onTextChunk(chunk) {
process.stdout.write(chunk);
}
});
}
catch (err) {
if (!(abortController.signal.aborted && err === abortController.signal.reason))
throw err;
}
finally {
consoleInteraction.stop();
if (abortController.signal.aborted)
process.stdout.write(endColor + chalk.yellow("[generation aborted by user]"));
else
process.stdout.write(endColor);
console.log();
}
const timeAfterPrompt = Date.now();
if (printTimings) {
if (LlamaLogLevelGreaterThan(llama.logLevel, LlamaLogLevel.info))
llama.logLevel = LlamaLogLevel.info;
await context.printTimings();
await new Promise((accept) => setTimeout(accept, 0)); // wait for logs to finish printing
llama.logLevel = llamaLogLevel;
}
if (timing)
console.info(chalk.dim("Response duration: ") +
prettyMilliseconds(timeAfterPrompt - timeBeforePrompt, {
keepDecimalsOnWholeSeconds: true,
secondsDecimalDigits: 2,
separateMilliseconds: true,
compact: false
}));
if (meter) {
const newTokenMeterState = contextSequence.tokenMeter.getState();
const tokenMeterDiff = TokenMeter.diff(newTokenMeterState, lastTokenMeterState);
lastTokenMeterState = newTokenMeterState;
const showDraftTokenMeterDiff = lastDraftTokenMeterState != null && draftContextSequence != null;
const tokenPredictionsStats = contextSequence.tokenPredictions;
const validatedTokenPredictions = tokenPredictionsStats.validated - lastTokenPredictionsStats.validated;
const refutedTokenPredictions = tokenPredictionsStats.refuted - lastTokenPredictionsStats.refuted;
const usedTokenPredictions = tokenPredictionsStats.used - lastTokenPredictionsStats.used;
const unusedTokenPredictions = tokenPredictionsStats.unused - lastTokenPredictionsStats.unused;
lastTokenPredictionsStats = tokenPredictionsStats;
console.info([
showDraftTokenMeterDiff && (chalk.yellow("Main".padEnd("Drafter".length))),
chalk.dim("Input tokens:") + " " + String(tokenMeterDiff.usedInputTokens).padEnd(5, " "),
chalk.dim("Output tokens:") + " " + String(tokenMeterDiff.usedOutputTokens).padEnd(5, " "),
showDraftTokenMeterDiff && (chalk.dim("Validated predictions:") + " " + String(validatedTokenPredictions).padEnd(5, " ")),
showDraftTokenMeterDiff && (chalk.dim("Refuted predictions:") + " " + String(refutedTokenPredictions).padEnd(5, " ")),
showDraftTokenMeterDiff && (chalk.dim("Used predictions:") + " " + String(usedTokenPredictions).padEnd(5, " ")),
showDraftTokenMeterDiff && (chalk.dim("Unused predictions:") + " " + String(unusedTokenPredictions).padEnd(5, " "))
].filter(Boolean).join(" "));
if (lastDraftTokenMeterState != null && draftContextSequence != null) {
const newDraftTokenMeterState = draftContextSequence.tokenMeter.getState();
const draftTokenMeterDiff = TokenMeter.diff(newDraftTokenMeterState, lastDraftTokenMeterState);
lastDraftTokenMeterState = newDraftTokenMeterState;
console.info([
chalk.yellow("Drafter"),
chalk.dim("Input tokens:") + " " + String(draftTokenMeterDiff.usedInputTokens).padEnd(5, " "),
chalk.dim("Output tokens:") + " " + String(draftTokenMeterDiff.usedOutputTokens).padEnd(5, " ")
].join(" "));
}
}
}
}
//# sourceMappingURL=CompleteCommand.js.map