3b71dd7300
* Update llama.py * offload * Update llama.py * Update llama.py * Update llama.py * Update llama.py * Update llama.py * Update llama.py * Update llama.py * continued pretraining trainer * Update trainer.py * Update trainer.py * Update trainer.py * Update trainer.py * is_bfloat16_supported * Update __init__.py * Update README.md * Update llama.py * is_bfloat16_supported * Update __init__.py
96 lines
3 KiB
Python
96 lines
3 KiB
Python
# Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
|
|
#
|
|
# 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.
|
|
|
|
from dataclasses import dataclass, field
|
|
from typing import Optional
|
|
from transformers import TrainingArguments
|
|
from trl import SFTTrainer
|
|
from . import is_bfloat16_supported
|
|
|
|
__all__ = [
|
|
"UnslothTrainingArguments",
|
|
"UnslothTrainer",
|
|
]
|
|
|
|
|
|
@dataclass
|
|
class UnslothTrainingArguments(TrainingArguments):
|
|
embedding_learning_rate : Optional[float] = field(
|
|
default = None,
|
|
metadata = {"help" : "Different learning rates for embeddings and lm_head."}
|
|
)
|
|
pass
|
|
|
|
|
|
def _create_unsloth_optimizer(
|
|
model,
|
|
optimizer_cls,
|
|
optimizer_kwargs,
|
|
embedding_lr = 5e-5,
|
|
):
|
|
lr = optimizer_kwargs["lr"]
|
|
weight_decay = optimizer_kwargs.get("weight_decay", 0.0)
|
|
|
|
param_groups = \
|
|
{
|
|
"non_embeddings" : {},
|
|
"embeddings" : {},
|
|
}
|
|
|
|
for name, param in model.named_parameters():
|
|
if not param.requires_grad: continue
|
|
if name.endswith("modules_to_save.default.weight"):
|
|
partial_name = name[:-len(".modules_to_save.default.weight")]
|
|
partial_name = partial_name[partial_name.rfind(".")+1:]
|
|
print(f"Unsloth: Setting lr = {embedding_lr:.2e} instead of {lr:.2e} for {partial_name}.")
|
|
param_groups["embeddings"] [name] = param
|
|
else:
|
|
param_groups["non_embeddings"][name] = param
|
|
pass
|
|
pass
|
|
|
|
optimizer_grouped_parameters = [
|
|
{
|
|
"params" : list(param_groups["non_embeddings"].values()),
|
|
"weight_decay" : weight_decay,
|
|
"lr" : lr,
|
|
},
|
|
{
|
|
"params" : list(param_groups["embeddings"].values()),
|
|
"weight_decay" : weight_decay,
|
|
"lr" : embedding_lr,
|
|
},
|
|
]
|
|
optimizer = optimizer_cls(optimizer_grouped_parameters, **optimizer_kwargs)
|
|
return optimizer
|
|
pass
|
|
|
|
|
|
class UnslothTrainer(SFTTrainer):
|
|
def create_optimizer(self):
|
|
embedding_learning_rate = getattr(self.args, "embedding_learning_rate", None)
|
|
if embedding_learning_rate is None: return super().create_optimizer()
|
|
|
|
if self.optimizer is None:
|
|
optimizer_cls, optimizer_kwargs = SFTTrainer.get_optimizer_cls_and_kwargs(self.args)
|
|
self.optimizer = _create_unsloth_optimizer(
|
|
self.model,
|
|
optimizer_cls,
|
|
optimizer_kwargs,
|
|
embedding_learning_rate,
|
|
)
|
|
pass
|
|
return self.optimizer
|
|
pass
|
|
pass
|