Fix: prevent loading best model when PEFT adapters are active (#3056)

[sahibpreetsingh12]

This pull request addresses issue #3056. When using load_best_model_at_end=True in combination with PEFT adapters, the trainer previously attempted to load a pytorch_model.bin file that doesn’t exist, resulting in a FileNotFoundError.

And the patch introduces a safeguard: it checks whether the model has any active adapters, and if so, skips reloading the best model. A clear user-facing warning is issued in these cases. This approach helps avoid crashes during training while retaining compatibility with adapter-based models.

The implementation closely follows the logic suggested in PR #3057.

[tomaarsen]

Hello!

My apologies, but about a month back I also looked into PEFT loading again in 3e7d2fa. In short, this _load_from_checkpoint works correctly with PEFT adapters, and I think it'd be wise to reuse it where possible.

With that approach, it's also possible to use PEFT with loading best model, so no need to return/give a warning.

I used this reproduction file

# -*- coding: utf-8 -*-
"""reproduce bug

Automatically generated by Colab.

Original file is located at
    https://colab.research.google.com/drive/1FfCUN1zwn0a7jG1WnDadXlmM4EFIL7wW

# Setup
"""

# !git clone https://github.com/GTimothee/transformers.git

# Commented out IPython magic to ensure Python compatibility.
# %cd transformers

# Commented out IPython magic to ensure Python compatibility.
# !git checkout sentencetransformers_load_peft
# !git pull
# %pip install .

# Commented out IPython magic to ensure Python compatibility.
# %pip install datasets sentence-transformers wandb

"""# Prepare data

code mostly copied from https://www.philschmid.de/fine-tune-embedding-model-for-rag
"""

from datasets import load_dataset

# Load dataset from the hub
dataset = load_dataset("philschmid/finanical-rag-embedding-dataset", split="train")

# rename columns
dataset = dataset.rename_column("question", "anchor")
dataset = dataset.rename_column("context", "positive")

# Add an id column to the dataset
dataset = dataset.add_column("id", range(len(dataset)))

# split dataset into a 10% test set
dataset = dataset.train_test_split(test_size=0.1)

# save datasets to disk
dataset["train"].to_json("train_dataset.json", orient="records")
dataset["test"].to_json("test_dataset.json", orient="records")

from datasets import load_dataset, concatenate_datasets
train_dataset = load_dataset("json", data_files="train_dataset.json")['train']#.select(range(50000))
val_dataset = load_dataset("json", data_files="test_dataset.json")['train']
corpus_dataset = concatenate_datasets([train_dataset, val_dataset])

corpus = dict(
    zip(corpus_dataset["id"], corpus_dataset["positive"])
)  # Our corpus (cid => document)
queries = dict(
    zip(val_dataset["id"], val_dataset["anchor"])
)  # Our queries (qid => question)
relevant_docs = {}  # Query ID to relevant documents (qid => set([relevant_cids])
for q_id in queries:
    relevant_docs[q_id] = [q_id]

from sentence_transformers.evaluation import (
    InformationRetrievalEvaluator
)
from sentence_transformers.util import cos_sim
ir_evaluator = InformationRetrievalEvaluator(
    queries=queries,
    corpus=corpus,
    relevant_docs=relevant_docs,
    score_functions={"cosine": cos_sim},
    batch_size=32,
    corpus_chunk_size=100,
    show_progress_bar=True
)

"""# Load model"""

from sentence_transformers import SentenceTransformer
from peft import LoraConfig, TaskType
import torch

model = SentenceTransformer("all-MiniLM-L6-v2", device="cuda" if torch.cuda.is_available() else "cpu")
peft_config = LoraConfig(
    task_type=TaskType.FEATURE_EXTRACTION,
    inference_mode=False,
    r=8,
    lora_alpha=32,
    lora_dropout=0.1,
)
model.add_adapter(peft_config)

embedding = model.encode("[bla] my name is [blub]")
print(embedding[:10])

from sentence_transformers import SentenceTransformerTrainingArguments
from sentence_transformers.training_args import BatchSamplers

args = SentenceTransformerTrainingArguments(
    output_dir="training_trial",
    num_train_epochs=0.1,
    per_device_train_batch_size=2,
    gradient_accumulation_steps=1,
    per_device_eval_batch_size=.5,
    warmup_ratio=0.1,

    learning_rate=1e-3,
    weight_decay=0.001,
    fp16=True,
    lr_scheduler_type='reduce_lr_on_plateau',
    report_to=None,

    batch_sampler=BatchSamplers.NO_DUPLICATES,
    eval_strategy="epoch",
    save_strategy="epoch",
    logging_steps=.2,
    save_total_limit=3,
    load_best_model_at_end=True,
    metric_for_best_model="eval_cosine_ndcg@10",
)

"""## it fails here:"""

from sentence_transformers import SentenceTransformerTrainer
from sentence_transformers.losses import MultipleNegativesRankingLoss

trainer = SentenceTransformerTrainer(
    model=model,
    args=args,
    train_dataset=train_dataset.select_columns(
        ["anchor", "positive"]
    ),
    loss=MultipleNegativesRankingLoss(model),
    evaluator=ir_evaluator
)

from transformers.utils.import_utils import is_peft_available
is_peft_available()

from packaging import version
import importlib
version.parse(importlib.metadata.version("peft"))

from transformers.integrations.peft import PeftAdapterMixin
print(isinstance(trainer.model, PeftAdapterMixin))
isinstance(trainer.model[0].auto_model, PeftAdapterMixin)

embedding = model.encode("[bla] my name is [blub]")
print("Before training:")
print(embedding[:10])

import os
os.environ["WANDB_MODE"] = "disabled"
trainer.train()

print(trainer.model[0].auto_model.active_adapters())

embedding = model.encode("[bla] my name is [blub]")
print("After training:")
print(embedding[:10])

# Save the adapter itself
model.save_pretrained("all-MiniLM-L6-v2-adapter")
# Load the adapter directly
loaded_model = SentenceTransformer("all-MiniLM-L6-v2-adapter")
embedding = loaded_model.encode("[bla] my name is [blub]")
print("After loading:")
print(embedding[:10])

"""
Before training:
[-0.05346012 -0.02178508 -0.00050329 -0.02896871 -0.09822005  0.00544653
  0.19883962  0.00645554 -0.04425742 -0.09591079]

After training:
[-0.05672589 -0.09013492 -0.0635585  -0.00599355 -0.12770522  0.0531613
  0.11739125 -0.03458377 -0.03921237 -0.06882482]

After loading:
[-0.05675213 -0.09006883 -0.0636415  -0.00605853 -0.12771909  0.05317839
  0.11733141 -0.03461097 -0.03919081 -0.06877207]
"""

I'll push the changes into this PR if you're okay with that.
also cc @GTimothee as this should resolve #3056 and #3057

• Tom Aarsen

[sahibpreetsingh12]

Perfect @tomaarsen since this was my first PR to Sentence Transformer
Thanks for sharing. I'll keep putting my best efforts