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
"""Persisted, feature-specific opt-ins for loading ML libraries during startup."""
from __future__ import annotations
import os
from collections.abc import Iterable
from typing import Literal
MlWarmupFeature = Literal["hardware", "text", "datasets", "media", "quantization"]
ML_WARMUP_FEATURES: tuple[MlWarmupFeature, ...] = (
"hardware",
"text",
"datasets",
"media",
"quantization",
)
ML_WARMUP_FEATURES_SETTING_KEY = "ml_library_warmup_features"
_HARDWARE_DEPENDENTS = frozenset({"text", "media", "quantization"})
def ml_warmup_disabled_by_env() -> bool:
"""Preserve the existing operator override for every optional warmup stage."""
return os.environ.get("UNSLOTH_STUDIO_DISABLE_TORCH_WARM") == "1"
def _validate_features(features: Iterable[str]) -> frozenset[str]:
if isinstance(features, (str, bytes, dict)):
raise ValueError("Startup warmup features must be a list of supported features.")
try:
selected = list(features)
except TypeError as exc:
raise ValueError("Startup warmup features must be a list of supported features.") from exc
if any(
not isinstance(feature, str) or feature not in ML_WARMUP_FEATURES for feature in selected
):
raise ValueError("Unknown startup warmup feature.")
return frozenset(selected)
def get_ml_warmup_features() -> frozenset[str]:
"""Missing, invalid or unreadable preferences keep optional imports cold."""
try:
from storage.studio_db import get_app_setting
stored = get_app_setting(ML_WARMUP_FEATURES_SETTING_KEY, None)
if not isinstance(stored, list):
return frozenset()
return _validate_features(stored)
except Exception:
return frozenset()
def set_ml_warmup_features(features: Iterable[str]) -> frozenset[str]:
"""Save explicit selections in canonical order; dependencies stay implicit."""
selected = _validate_features(features)
from storage.studio_db import upsert_app_settings
upsert_app_settings(
{
ML_WARMUP_FEATURES_SETTING_KEY: [
feature for feature in ML_WARMUP_FEATURES if feature in selected
],
}
)
return selected
def effective_ml_warmup_features(features: Iterable[str]) -> frozenset[str]:
"""Text, media and quantization share hardware setup; datasets do not need it."""
selected = _validate_features(features)
if selected & _HARDWARE_DEPENDENTS:
return selected | {"hardware"}
return selected
def get_startup_ml_warmup_features() -> frozenset[str]:
"""Resolve the launch's feature selection, respecting the broad disable flag."""
if ml_warmup_disabled_by_env():
return frozenset()
return effective_ml_warmup_features(get_ml_warmup_features())