2024-06-15 11:51:20 +00:00
|
|
|
# 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.
|
|
|
|
|
|
|
|
import os
|
|
|
|
|
|
|
|
import torch
|
|
|
|
|
|
|
|
from llamafactory.extras.misc import get_current_device
|
2024-07-18 17:06:27 +00:00
|
|
|
from llamafactory.train.test_utils import load_train_model
|
2024-06-15 11:51:20 +00:00
|
|
|
|
|
|
|
|
|
|
|
TINY_LLAMA = os.environ.get("TINY_LLAMA", "llamafactory/tiny-random-Llama-3")
|
|
|
|
|
|
|
|
TRAIN_ARGS = {
|
|
|
|
"model_name_or_path": TINY_LLAMA,
|
|
|
|
"stage": "sft",
|
|
|
|
"do_train": True,
|
|
|
|
"finetuning_type": "lora",
|
|
|
|
"lora_target": "all",
|
|
|
|
"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,
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
def test_checkpointing_enable():
|
2024-07-18 17:06:27 +00:00
|
|
|
model = load_train_model(disable_gradient_checkpointing=False, **TRAIN_ARGS)
|
2024-06-15 11:51:20 +00:00
|
|
|
for module in filter(lambda m: hasattr(m, "gradient_checkpointing"), model.modules()):
|
|
|
|
assert getattr(module, "gradient_checkpointing") is True
|
|
|
|
|
|
|
|
|
|
|
|
def test_checkpointing_disable():
|
2024-07-18 17:06:27 +00:00
|
|
|
model = load_train_model(disable_gradient_checkpointing=True, **TRAIN_ARGS)
|
2024-06-15 11:51:20 +00:00
|
|
|
for module in filter(lambda m: hasattr(m, "gradient_checkpointing"), model.modules()):
|
|
|
|
assert getattr(module, "gradient_checkpointing") is False
|
|
|
|
|
|
|
|
|
|
|
|
def test_upcast_layernorm():
|
2024-07-18 17:06:27 +00:00
|
|
|
model = load_train_model(upcast_layernorm=True, **TRAIN_ARGS)
|
2024-06-15 11:51:20 +00:00
|
|
|
for name, param in model.named_parameters():
|
|
|
|
if param.ndim == 1 and "norm" in name:
|
|
|
|
assert param.dtype == torch.float32
|
|
|
|
|
|
|
|
|
|
|
|
def test_upcast_lmhead_output():
|
2024-07-18 17:06:27 +00:00
|
|
|
model = load_train_model(upcast_lmhead_output=True, **TRAIN_ARGS)
|
2024-06-15 11:51:20 +00:00
|
|
|
inputs = torch.randn((1, 16), dtype=torch.float16, device=get_current_device())
|
2024-06-19 13:27:00 +00:00
|
|
|
outputs: "torch.Tensor" = model.get_output_embeddings()(inputs)
|
2024-06-15 11:51:20 +00:00
|
|
|
assert outputs.dtype == torch.float32
|