File size: 3,837 Bytes
93adfea |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 |
# Copyright 2025 the LlamaFactory team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import pytest
import torch
from llamafactory.train.test_utils import (
check_lora_model,
compare_model,
load_infer_model,
load_reference_model,
load_train_model,
patch_valuehead_model,
)
TINY_LLAMA = os.getenv("TINY_LLAMA", "llamafactory/tiny-random-Llama-3")
TINY_LLAMA_ADAPTER = os.getenv("TINY_LLAMA_ADAPTER", "llamafactory/tiny-random-Llama-3-lora")
TINY_LLAMA_VALUEHEAD = os.getenv("TINY_LLAMA_VALUEHEAD", "llamafactory/tiny-random-Llama-3-valuehead")
TRAIN_ARGS = {
"model_name_or_path": TINY_LLAMA,
"stage": "sft",
"do_train": True,
"finetuning_type": "lora",
"dataset": "llamafactory/tiny-supervised-dataset",
"dataset_dir": "ONLINE",
"template": "llama3",
"cutoff_len": 1024,
"overwrite_cache": True,
"output_dir": "dummy_dir",
"overwrite_output_dir": True,
"fp16": True,
}
INFER_ARGS = {
"model_name_or_path": TINY_LLAMA,
"adapter_name_or_path": TINY_LLAMA_ADAPTER,
"finetuning_type": "lora",
"template": "llama3",
"infer_dtype": "float16",
}
@pytest.fixture
def fix_valuehead_cpu_loading():
patch_valuehead_model()
def test_lora_train_qv_modules():
model = load_train_model(lora_target="q_proj,v_proj", **TRAIN_ARGS)
linear_modules, _ = check_lora_model(model)
assert linear_modules == {"q_proj", "v_proj"}
def test_lora_train_all_modules():
model = load_train_model(lora_target="all", **TRAIN_ARGS)
linear_modules, _ = check_lora_model(model)
assert linear_modules == {"q_proj", "k_proj", "v_proj", "o_proj", "up_proj", "gate_proj", "down_proj"}
def test_lora_train_extra_modules():
model = load_train_model(additional_target="embed_tokens,lm_head", **TRAIN_ARGS)
_, extra_modules = check_lora_model(model)
assert extra_modules == {"embed_tokens", "lm_head"}
def test_lora_train_old_adapters():
model = load_train_model(adapter_name_or_path=TINY_LLAMA_ADAPTER, create_new_adapter=False, **TRAIN_ARGS)
ref_model = load_reference_model(TINY_LLAMA, TINY_LLAMA_ADAPTER, use_lora=True, is_trainable=True)
compare_model(model, ref_model)
def test_lora_train_new_adapters():
model = load_train_model(adapter_name_or_path=TINY_LLAMA_ADAPTER, create_new_adapter=True, **TRAIN_ARGS)
ref_model = load_reference_model(TINY_LLAMA, TINY_LLAMA_ADAPTER, use_lora=True, is_trainable=True)
compare_model(
model, ref_model, diff_keys=["q_proj", "k_proj", "v_proj", "o_proj", "up_proj", "gate_proj", "down_proj"]
)
@pytest.mark.usefixtures("fix_valuehead_cpu_loading")
def test_lora_train_valuehead():
model = load_train_model(add_valuehead=True, **TRAIN_ARGS)
ref_model = load_reference_model(TINY_LLAMA_VALUEHEAD, is_trainable=True, add_valuehead=True)
state_dict = model.state_dict()
ref_state_dict = ref_model.state_dict()
assert torch.allclose(state_dict["v_head.summary.weight"], ref_state_dict["v_head.summary.weight"])
assert torch.allclose(state_dict["v_head.summary.bias"], ref_state_dict["v_head.summary.bias"])
def test_lora_inference():
model = load_infer_model(**INFER_ARGS)
ref_model = load_reference_model(TINY_LLAMA, TINY_LLAMA_ADAPTER, use_lora=True).merge_and_unload()
compare_model(model, ref_model)
|