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
from types import SimpleNamespace
import pytest
from models.training import TrainingStartRequest, TokenPreviewRequest
from utils.models.training_defaults import model_training_recommendations
def test_training_and_token_preview_default_to_assistant_targets_and_packing():
start = TrainingStartRequest(
model_name = "unlisted/model", training_type = "LoRA/QLoRA", format_type = "auto"
)
preview = TokenPreviewRequest(model_name = "unlisted/model")
assert start.packing is True
assert start.train_on_completions is True
assert preview.train_on_completions is True
@pytest.mark.parametrize("value", [True, False])
def test_worker_config_preserves_explicit_training_toggles(monkeypatch, value):
from core.training import training
monkeypatch.setattr(training, "get_device", lambda: SimpleNamespace(value = "cuda"))
request = TrainingStartRequest(
model_name = "unlisted/model",
training_type = "LoRA/QLoRA",
format_type = "auto",
packing = value,
train_on_completions = value,
)
config = training._build_training_worker_config(request.model_dump())
assert config["packing"] is value
assert config["train_on_completions"] is value
def test_internal_worker_config_defaults_match_api_defaults(monkeypatch):
from core.training import training
monkeypatch.setattr(training, "get_device", lambda: SimpleNamespace(value = "cuda"))
config = training._build_training_worker_config({"model_name": "unlisted/model"})
assert config["packing"] is True
assert config["train_on_completions"] is True
@pytest.mark.parametrize(
"flags, enabled",
[
({}, True),
({"is_vision": True}, True),
({"is_vision": True, "is_audio": True}, True),
({"is_audio": True}, False),
({"is_embedding": True}, False),
({"is_decision": True}, False),
],
)
def test_recommendations_ignore_legacy_recipe_opt_outs_but_respect_training_objective(
flags, enabled
):
config = {"training": {"packing": False, "train_on_completions": False, "batch_size": 2}}
defaults = {"is_vision": False, "is_audio": False, "is_embedding": False, "is_decision": False}
result = model_training_recommendations(config, **{**defaults, **flags})
assert result["training"]["packing"] is enabled
assert result["training"]["train_on_completions"] is enabled
assert result["training"]["batch_size"] == 2
# Imported YAML stays literal; only the model-recommendation response is normalized.
assert config["training"]["packing"] is False
assert config["training"]["train_on_completions"] is False
def test_vision_and_text_diffusion_packing_guards_remain_effective():
from core.training.dataset_bounds import effective_packing
from core.training.worker import _apply_diffusion_lm_overrides
assert effective_packing({"packing": True}, branch_never_packs = True) is False
config = {"packing": True}
_apply_diffusion_lm_overrides(SimpleNamespace(diffusion_lm_profile = "dream"), config)
assert config["packing"] is False
@pytest.mark.parametrize(
"overrides",
[
{"dataset_streaming": True, "max_steps": 2},
{"training_type": "Continued Pretraining"},
{"format_type": "raw"},
{"is_decision": True},
{"is_embedding": True},
],
)
def test_unsupported_objectives_do_not_enable_an_omitted_completion_option(overrides):
values = {
"model_name": "unlisted/model",
"training_type": "LoRA/QLoRA",
"format_type": "auto",
**overrides,
}
assert TrainingStartRequest(**values).train_on_completions is False
# Explicit values are preserved; the existing route/trainer checks remain authoritative.
assert TrainingStartRequest(**values, train_on_completions = True).train_on_completions is True