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.
"""FastSentenceTransformer.get_peft_model must hand every LoRA argument to LoraConfig on the fast encoder path."""
import math
import pytest
pytest.importorskip("sentence_transformers")
from unsloth import FastSentenceTransformer # noqa: E402
@pytest.fixture
def fast_encoder(tmp_path):
from sentence_transformers import SentenceTransformer
from sentence_transformers.models import Pooling, Transformer
from transformers import BertConfig, BertModel, BertTokenizerFast
vocab = tmp_path / "vocab.txt"
vocab.write_text(
"\n".join(["[PAD]", "[UNK]", "[CLS]", "[SEP]", "[MASK]", "a", "red", "square"]),
encoding = "utf-8",
)
config = BertConfig(
vocab_size = 8,
hidden_size = 16,
num_hidden_layers = 2,
num_attention_heads = 2,
intermediate_size = 32,
max_position_embeddings = 64,
)
model_dir = tmp_path / "model"
BertModel(config).save_pretrained(model_dir)
BertTokenizerFast(vocab_file = str(vocab)).save_pretrained(model_dir)
model = SentenceTransformer(
modules = [Transformer(str(model_dir)), Pooling(16)],
device = "cpu",
)
model._unsloth_fast_encoder = True
return model
def test_layers_rslora_and_modules_to_save_reach_the_adapter(fast_encoder):
r, alpha = 4, 8
model = FastSentenceTransformer.get_peft_model(
fast_encoder,
r = r,
lora_alpha = alpha,
target_modules = ["query", "key", "value"],
layers_to_transform = [1],
use_rslora = True,
modules_to_save = ["pooler"],
use_gradient_checkpointing = False,
)
peft_model = model[0].auto_model
config = peft_model.peft_config["default"]
assert config.use_rslora is True
assert config.layers_to_transform == [1]
assert set(config.modules_to_save) == {"pooler"}
names = [name for name, _ in peft_model.named_parameters() if ".lora_A." in name]
assert names and all(".layer.1." in name for name in names)
scalings = [
module.scaling["default"] for module in peft_model.modules() if hasattr(module, "lora_A")
]
assert scalings and all(s == pytest.approx(alpha / math.sqrt(r)) for s in scalings)
assert any(
"modules_to_save" in name and param.requires_grad
for name, param in peft_model.named_parameters()
)
def test_float16_modules_to_save_copy_trains_in_float32(fast_encoder):
import torch
fast_encoder.half()
model = FastSentenceTransformer.get_peft_model(
fast_encoder,
r = 4,
lora_alpha = 8,
target_modules = ["query", "key", "value"],
modules_to_save = ["pooler"],
use_gradient_checkpointing = False,
)
copies = [
param
for name, param in model[0].auto_model.named_parameters()
if ".modules_to_save." in name
]
assert copies and all(p.dtype == torch.float32 and p.requires_grad for p in copies)