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
"""family_inference_infos lists every auto-policy family; quantised estimates sit below bf16."""
from __future__ import annotations
from core.inference.diffusion_auto_policy import _FAMILY_BF16_GB, _QUANT_STEADY_FACTOR
from core.inference.diffusion_inference_info import family_inference_infos
def test_covers_every_auto_policy_family():
infos = family_inference_infos()
names = {info["family"] for info in infos}
assert names == set(_FAMILY_BF16_GB), "info must list exactly the auto-policy families"
assert [info["family"] for info in infos] == list(_FAMILY_BF16_GB)
def test_each_family_reports_all_schemes():
for info in family_inference_infos():
estimated = info["estimated_resident_gb"]
assert set(estimated) == {"bf16", *_QUANT_STEADY_FACTOR}
for value in estimated.values():
assert isinstance(value, float)
assert round(value, 1) == value
def test_component_sizes_match_the_table():
infos = {info["family"]: info for info in family_inference_infos()}
for name, (transformer, text_encoders, vae) in _FAMILY_BF16_GB.items():
info = infos[name]
assert info["transformer_bf16_gb"] == round(transformer, 1)
assert info["text_encoders_bf16_gb"] == round(text_encoders, 1)
assert info["vae_bf16_gb"] == round(vae, 1)
def test_quantised_estimate_is_below_bf16():
# Companions are shared and every steady factor is below 1, so quantised < bf16.
for info in family_inference_infos():
estimated = info["estimated_resident_gb"]
for scheme in _QUANT_STEADY_FACTOR:
assert estimated[scheme] < estimated["bf16"], f"{info['family']} {scheme}"
def test_nvfp4_is_below_int8():
for info in family_inference_infos():
estimated = info["estimated_resident_gb"]
assert estimated["nvfp4"] < estimated["int8"], info["family"]
def test_a_policy_family_reports_its_policy_nvfp4_footprint():
from core.inference.diffusion_auto_policy import _POLICY_STEADY_FACTOR, policy_steady_factor
infos = {info["family"]: info for info in family_inference_infos()}
for family, policy_id in (
("z-image", "zimg_rg76_v1"),
("flux.1", "flux_r420_v1"),
("qwen-image", "qwen_p02_v1"),
("qwen-image-2.1", "qwen21_r020_v1"),
):
transformer, text_encoders, vae = _FAMILY_BF16_GB[family]
expected = round(transformer * _POLICY_STEADY_FACTOR[policy_id] + text_encoders + vae, 1)
assert infos[family]["estimated_resident_gb"]["nvfp4"] == expected, family
whole = round(transformer * _QUANT_STEADY_FACTOR["nvfp4"] + text_encoders + vae, 1)
assert expected > whole, family
def test_a_family_without_a_policy_keeps_the_whole_model_factor():
from core.inference.diffusion_auto_policy import policy_steady_factor
infos = {info["family"]: info for info in family_inference_infos()}
for family in ("flux.1-kontext", "qwen-image-edit", "lumina-2"):
transformer, text_encoders, vae = _FAMILY_BF16_GB[family]
assert policy_steady_factor(family) is None, family
assert infos[family]["estimated_resident_gb"]["nvfp4"] == round(
transformer * _QUANT_STEADY_FACTOR["nvfp4"] + text_encoders + vae, 1
), family