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
import {
gpuMemoryTotalsGb,
resolveGpuVramUsedGb,
sharesHostMemoryDevice,
type VramReportingGpu,
} from "../hooks/gpu-vram.ts";
import type { GpuDevice } from "../hooks/use-system.ts";
export interface CardUsage {
key: string;
label: string;
name: string;
title: string;
usedGb: number | null;
totalGb: number;
}
const VENDOR_PREFIX = /^(?:AMD\s+Radeon(?:\(TM\))?|NVIDIA(?:\s+GeForce)?|Intel(?:\(R\))?\s+Arc(?:\(TM\))?)\s+/i;
export function shortCardName(name: string | undefined): string {
if (!name) return "GPU";
return name.replace(VENDOR_PREFIX, "") || name;
}
export interface VramCards {
cards: GpuDevice[];
totalGb: number;
/** null = unknown: one card without a reading must not read as 0. */
usedGb: number | null;
rows: CardUsage[];
showTotal: boolean;
}
/** Dedicated cards only: an iGPU borrows system RAM, and its missing reading would blank the total. */
export function vramCards(
devices: GpuDevice[],
usageGpu: VramReportingGpu | null | undefined,
): VramCards {
const dedicated = devices.filter((device) => !sharesHostMemoryDevice(device));
const cards = dedicated.length > 0 ? dedicated : devices;
const usedGb =
cards.length === devices.length
? resolveGpuVramUsedGb(usageGpu)
: resolveGpuVramUsedGb({ devices: cards });
const rows =
cards.length > 1
? cards.map((card, position) => ({
key: `${card.index ?? position}`,
// The ordinal the detail rows and Settings use, so a masked device keeps one name.
label: `GPU ${card.visible_ordinal ?? card.index ?? position}`,
name: shortCardName(card.name),
title: card.name ?? "GPU",
usedGb: Number.isFinite(card.vram_used_gb) ? (card.vram_used_gb as number) : null,
totalGb: gpuMemoryTotalsGb([card]).total,
}))
: [];
const showTotal =
rows.length === 0 || (usedGb !== null && rows.some((row) => row.usedGb === null));
return { cards, totalGb: gpuMemoryTotalsGb(cards).total, usedGb, rows, showTotal };
}
export function holdsModelsInVram(cards: readonly GpuDevice[]): boolean {
return cards.some((card) => card.shared_memory !== true && card.unified_memory !== true);
}