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.
"""The fast-encoder merged export must merge a copy and leave the live LoRA model training (unsloth#14246)."""
import ast
import copy
from pathlib import Path
import pytest
import torch
peft = pytest.importorskip("peft")
transformers = pytest.importorskip("transformers")
ST_PY = Path(__file__).resolve().parents[1] / "unsloth" / "models" / "sentence_transformer.py"
def _helper():
src = ST_PY.read_text(encoding = "utf-8")
node = [
n
for n in ast.parse(src).body
if isinstance(n, ast.FunctionDef) and n.name == "_copy_sentence_encoder_for_merge"
]
assert node, "_copy_sentence_encoder_for_merge not found"
ns = {"copy": copy, "torch": torch}
exec(ast.get_source_segment(src, node[0]), ns)
return ns["_copy_sentence_encoder_for_merge"]
class _FakeQuantBert(transformers.BertModel):
"""Takes the bitsandbytes rebuild branch on CPU: flagged 4bit, dequantize is a no-op."""
def dequantize(self, dtype = None):
return self
def _bert(cls = transformers.BertModel):
torch.manual_seed(0)
cfg = transformers.BertConfig(
vocab_size = 64,
hidden_size = 32,
num_hidden_layers = 2,
num_attention_heads = 2,
intermediate_size = 64,
max_position_embeddings = 32,
)
return cls(cfg).eval()
def _lora(base, init = True):
cfg = peft.LoraConfig(
r = 4,
lora_alpha = 8,
target_modules = ["query", "value"],
init_lora_weights = init,
task_type = "FEATURE_EXTRACTION",
)
model = peft.get_peft_model(base, cfg)
with torch.no_grad():
for name, p in model.named_parameters():
if "lora_B" in name:
p.normal_(0, 0.05)
return model
def _out(model, ids):
with torch.no_grad():
return model(input_ids = ids).last_hidden_state
IDS = torch.tensor([[1, 5, 9, 13, 2, 7]])
def test_live_adapters_survive_and_merged_copy_matches():
live = _lora(_bert())
params = {n: p for n, p in live.named_parameters() if "lora_" in n}
expected = _out(live, IDS)
merged = _helper()(live, torch.float32)
assert isinstance(live, peft.PeftModel)
assert {n: p for n, p in live.named_parameters() if "lora_" in n} == params
assert not any("lora_" in n for n, _ in merged.named_parameters())
torch.testing.assert_close(_out(merged, IDS), expected)
torch.testing.assert_close(_out(live, IDS), expected)
@pytest.mark.parametrize("init", [True, "olora"])
def test_quantized_rebuild_merges_the_trained_weights(init):
base = _bert(_FakeQuantBert)
base.is_loaded_in_4bit = True
live = _lora(base, init)
expected = _out(live, IDS)
rng = torch.get_rng_state()
merged = _helper()(live, torch.float32)
assert torch.equal(torch.get_rng_state(), rng), "the export advanced the live run's RNG"
assert isinstance(live, peft.PeftModel)
assert not getattr(merged, "is_loaded_in_4bit", False)
torch.testing.assert_close(_out(merged, IDS), expected, atol = 1e-5, rtol = 1e-4)