2024-06-15 17:54:33 +08:00
|
|
|
# Copyright 2024 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.
|
|
|
|
|
2024-06-08 06:46:09 +08:00
|
|
|
import os
|
|
|
|
|
2024-06-15 20:06:17 +08:00
|
|
|
import pytest
|
2024-06-08 06:46:09 +08:00
|
|
|
import torch
|
|
|
|
|
2024-07-19 01:06:27 +08:00
|
|
|
from llamafactory.train.test_utils import (
|
|
|
|
check_lora_model,
|
|
|
|
compare_model,
|
|
|
|
load_infer_model,
|
|
|
|
load_reference_model,
|
|
|
|
load_train_model,
|
|
|
|
patch_valuehead_model,
|
|
|
|
)
|
2024-06-08 06:46:09 +08:00
|
|
|
|
|
|
|
|
2024-06-10 21:24:15 +08:00
|
|
|
TINY_LLAMA = os.environ.get("TINY_LLAMA", "llamafactory/tiny-random-Llama-3")
|
2024-06-08 06:46:09 +08:00
|
|
|
|
2024-06-15 04:05:54 +08:00
|
|
|
TINY_LLAMA_ADAPTER = os.environ.get("TINY_LLAMA_ADAPTER", "llamafactory/tiny-random-Llama-3-lora")
|
|
|
|
|
2024-06-15 19:51:20 +08:00
|
|
|
TINY_LLAMA_VALUEHEAD = os.environ.get("TINY_LLAMA_VALUEHEAD", "llamafactory/tiny-random-Llama-3-valuehead")
|
|
|
|
|
2024-06-10 21:24:15 +08:00
|
|
|
TRAIN_ARGS = {
|
2024-06-08 06:46:09 +08:00
|
|
|
"model_name_or_path": TINY_LLAMA,
|
|
|
|
"stage": "sft",
|
|
|
|
"do_train": True,
|
|
|
|
"finetuning_type": "lora",
|
2024-06-10 21:24:15 +08:00
|
|
|
"dataset": "llamafactory/tiny-supervised-dataset",
|
2024-06-08 06:46:09 +08:00
|
|
|
"dataset_dir": "ONLINE",
|
|
|
|
"template": "llama3",
|
|
|
|
"cutoff_len": 1024,
|
|
|
|
"overwrite_cache": True,
|
|
|
|
"output_dir": "dummy_dir",
|
|
|
|
"overwrite_output_dir": True,
|
|
|
|
"fp16": True,
|
|
|
|
}
|
|
|
|
|
2024-06-15 04:05:54 +08:00
|
|
|
INFER_ARGS = {
|
|
|
|
"model_name_or_path": TINY_LLAMA,
|
|
|
|
"adapter_name_or_path": TINY_LLAMA_ADAPTER,
|
|
|
|
"finetuning_type": "lora",
|
|
|
|
"template": "llama3",
|
|
|
|
"infer_dtype": "float16",
|
|
|
|
}
|
2024-06-08 06:46:09 +08:00
|
|
|
|
2024-06-15 04:05:54 +08:00
|
|
|
|
2024-06-15 20:06:17 +08:00
|
|
|
@pytest.fixture
|
|
|
|
def fix_valuehead_cpu_loading():
|
2024-07-19 01:06:27 +08:00
|
|
|
patch_valuehead_model()
|
2024-06-15 20:06:17 +08:00
|
|
|
|
|
|
|
|
2024-06-15 19:51:20 +08:00
|
|
|
def test_lora_train_qv_modules():
|
2024-07-19 01:06:27 +08:00
|
|
|
model = load_train_model(lora_target="q_proj,v_proj", **TRAIN_ARGS)
|
|
|
|
linear_modules, _ = check_lora_model(model)
|
2024-06-15 19:51:20 +08:00
|
|
|
assert linear_modules == {"q_proj", "v_proj"}
|
|
|
|
|
|
|
|
|
2024-06-15 04:05:54 +08:00
|
|
|
def test_lora_train_all_modules():
|
2024-07-19 01:06:27 +08:00
|
|
|
model = load_train_model(lora_target="all", **TRAIN_ARGS)
|
|
|
|
linear_modules, _ = check_lora_model(model)
|
2024-06-08 06:46:09 +08:00
|
|
|
assert linear_modules == {"q_proj", "k_proj", "v_proj", "o_proj", "up_proj", "gate_proj", "down_proj"}
|
|
|
|
|
|
|
|
|
2024-06-15 04:05:54 +08:00
|
|
|
def test_lora_train_extra_modules():
|
2024-07-19 01:06:27 +08:00
|
|
|
model = load_train_model(additional_target="embed_tokens,lm_head", **TRAIN_ARGS)
|
|
|
|
_, extra_modules = check_lora_model(model)
|
2024-06-08 06:46:09 +08:00
|
|
|
assert extra_modules == {"embed_tokens", "lm_head"}
|
2024-06-15 04:05:54 +08:00
|
|
|
|
|
|
|
|
|
|
|
def test_lora_train_old_adapters():
|
2024-07-19 01:06:27 +08:00
|
|
|
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)
|
2024-06-15 04:05:54 +08:00
|
|
|
compare_model(model, ref_model)
|
|
|
|
|
|
|
|
|
|
|
|
def test_lora_train_new_adapters():
|
2024-07-19 01:06:27 +08:00
|
|
|
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)
|
2024-06-15 04:05:54 +08:00
|
|
|
compare_model(
|
|
|
|
model, ref_model, diff_keys=["q_proj", "k_proj", "v_proj", "o_proj", "up_proj", "gate_proj", "down_proj"]
|
|
|
|
)
|
|
|
|
|
|
|
|
|
2024-06-15 20:06:17 +08:00
|
|
|
@pytest.mark.usefixtures("fix_valuehead_cpu_loading")
|
2024-06-15 19:51:20 +08:00
|
|
|
def test_lora_train_valuehead():
|
2024-07-19 01:06:27 +08:00
|
|
|
model = load_train_model(add_valuehead=True, **TRAIN_ARGS)
|
|
|
|
ref_model = load_reference_model(TINY_LLAMA_VALUEHEAD, is_trainable=True, add_valuehead=True)
|
2024-06-15 19:51:20 +08:00
|
|
|
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"])
|
|
|
|
|
|
|
|
|
2024-06-15 04:05:54 +08:00
|
|
|
def test_lora_inference():
|
2024-07-19 01:06:27 +08:00
|
|
|
model = load_infer_model(**INFER_ARGS)
|
|
|
|
ref_model = load_reference_model(TINY_LLAMA, TINY_LLAMA_ADAPTER, use_lora=True).merge_and_unload()
|
2024-06-15 04:05:54 +08:00
|
|
|
compare_model(model, ref_model)
|