Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.
Python agentaichatgptdeepseekfine-tuninggemmaimage-generationllamallmllmsopenaipythonqwenreinforcement-learningself-hostedstable-diffusiontext-to-speechttsuiunsloth
// SPDX-License-Identifier: AGPL-3.0-only
// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
/** Mirrors llama_server_args.py override parsing; malformed input returns null instead of throwing. */
const SPLIT_MODE_FLAGS = new Set(["-sm", "--split-mode"]);
const GPU_LAYER_FLAGS = new Set(["-ngl", "--gpu-layers", "--n-gpu-layers"]);
/** `_LAYER_OFFLOAD_FLAGS | _MOE_OFFLOAD_FLAGS`, which manual mode owns and strips. */
const OFFLOAD_SHADOWING_FLAGS = new Set([
...GPU_LAYER_FLAGS,
"-fit",
"--fit",
"-ncmoe",
"--n-cpu-moe",
"-cmoe",
"--cpu-moe",
]);
/** Shorts start with a letter, so `-1` is a value. Underscores fold as in llama.cpp. */
function flagName(token: string): string | null {
const trimmed = token.trim();
if (!trimmed.startsWith("-") || trimmed === "-" || trimmed === "--") {
return null;
}
const second = trimmed[1];
if (second !== undefined && (/[0-9]/.test(second) || second === ".")) {
return null;
}
const name = trimmed.split("=", 1)[0];
return name.startsWith("--") ? name.replaceAll("_", "-") : name;
}
function lastFlagValue(
args: readonly string[] | null | undefined,
flags: ReadonlySet<string>,
): string | null {
let value: string | null = null;
for (let i = 0; i < (args?.length ?? 0); i += 1) {
const token = String(args?.[i]);
const flag = flagName(token);
if (flag === null || !flags.has(flag)) {
continue;
}
if (token.includes("=")) {
value = token.slice(token.indexOf("=") + 1);
continue;
}
const next = args?.[i + 1];
if (next === undefined || flagName(String(next)) !== null) {
return null;
}
value = String(next);
i += 1;
}
return value;
}
export function resolveTensorParallel(
extraArgs: readonly string[] | null | undefined,
tensorParallel: boolean,
): boolean {
const override = lastFlagValue(extraArgs, SPLIT_MODE_FLAGS);
return override === null
? tensorParallel
: override.trim().toLowerCase() === "tensor";
}
/** Malformed must stay distinct from no override, or the saved setting is lost silently. */
export type GpuLayersOverride =
| { kind: "absent" }
| { kind: "value"; layers: number }
| { kind: "invalid" };
export function parseGpuLayersOverride(
extraArgs: readonly string[] | null | undefined,
): GpuLayersOverride {
const raw = lastFlagValue(extraArgs, GPU_LAYER_FLAGS);
if (raw === null) {
return { kind: "absent" };
}
if (raw === "auto") {
return { kind: "value", layers: -1 };
}
if (raw === "all") {
return { kind: "value", layers: 2147483647 };
}
const value = Number.parseInt(raw, 10);
return Number.isInteger(value) && value >= -1 && String(value) === raw.trim()
? { kind: "value", layers: value }
: { kind: "invalid" };
}
/** Only under manual: in auto an inherited `-ngl` is respected. */
export function stripManagedOffloadFlags(
extraArgs: readonly string[] | null | undefined,
): string[] | null | undefined {
if (extraArgs == null) {
return extraArgs;
}
const kept: string[] = [];
for (let i = 0; i < extraArgs.length; i += 1) {
const token = String(extraArgs[i]);
const flag = flagName(token);
if (flag === null || !OFFLOAD_SHADOWING_FLAGS.has(flag)) {
kept.push(token);
continue;
}
if (!token.includes("=")) {
const next = extraArgs[i + 1];
if (next !== undefined && flagName(String(next)) === null) {
i += 1;
}
}
}
return kept;
}