Parameter-efficient fine-tuning of Gemma with LoRA and QLoRA
Authors: Hongyu Chiu, Abheesht Sharma, Matthew Watson
Date created: 2024/08/06
Last modified: 2024/08/06
Description: Use KerasHub to fine-tune a Gemma LLM with LoRA and QLoRA.
Introduction
Large Language Models (LLMs) have been shown to be effective at a variety of NLP tasks. An LLM is first pre-trained on a large corpus of text in a self-supervised fashion. Pre-training helps LLMs learn general-purpose knowledge, such as statistical relationships between words. An LLM can then be fine-tuned on a downstream task of interest (such as sentiment analysis).
However, LLMs are extremely large in size, and we don't need to train all the parameters in the model while fine-tuning, especially because datasets on which the model is fine-tuned are relatively small. Another way of saying this is that LLMs are over-parametrized for fine-tuning. This is where Low-Rank Adaptation (LoRA) comes in; it significantly reduces the number of trainable parameters. This results in a decrease in training time and GPU memory usage, while maintaining the quality of the outputs.
Furthermore, Quantized Low-Rank Adaptation (QLoRA) extends LoRA to enhance efficiency through quantization techniques without performance degradation.
In this example, we will fine-tune KerasHub's Gemma model on the next token prediction task using LoRA and QLoRA.
Note that this example runs on all backends supported by Keras. TensorFlow is only used for data preprocessing.
Setup
Before we start implementing the pipeline, let's install and import all the libraries we need. We'll be using the KerasHub library.
Secondly, let's set the precision to bfloat16. This will help us reduce the memory usage and training time.
Also, ensure that KAGGLE_USERNAME and KAGGLE_KEY have been correctly
configured to access the Gemma model.
# We might need the latest code from Keras and KerasHub
!pip install -q git+https://github.com/keras-team/keras.git git+https://github.com/keras-team/keras-hub.git
import gc
import os
os.environ["KERAS_BACKEND"] = "jax"
os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3" # Suppress verbose logging from TF
# os.environ["KAGGLE_USERNAME"] = "..."
# os.environ["KAGGLE_KEY"] = "..."
import keras
import keras_hub
import tensorflow as tf
import tensorflow_datasets as tfds
keras.config.set_dtype_policy("bfloat16")
Dataset
We will use the MTNT (Machine Translation of Noisy Text) dataset, which is available from TensorFlow Datasets. In this example, we will use the French-to-English portion of the dataset.
train_ds = tfds.load("mtnt/fr-en", split="train")
We can print some samples. Each sample in the dataset contains two entries:
- src: the original French sentence.
- dst: the corresponding English translation.
examples = train_ds.take(3)
examples = examples.as_numpy_iterator()
for idx, example in enumerate(examples):
print(f"Example {idx}:")
for key, val in example.items():
print(f"{key}: {val}")
print()
Example 0:
dst: b'Yep, serious...'
src: b"Le journal l'est peut-\xc3\xaatre, mais m\xc3\xaame moi qui suit de droite je les trouve limite de temps en temps..."
Example 1:
dst: b'Finally, I explained to you in what context this copy-pasting is relevant: when we are told padamalgame etc.'
src: b"Enfin je t'ai expliqu\xc3\xa9 dans quel cadre ce copypasta est pertinent : quand on nous dit padamalgame etc."
Example 2:
dst: b'Gift of Ubiquity: Fran\xc3\xa7ois Baroin is now advisor to the Barclays Bank, mayor, president of the agglomeration, professor at HEC Paris, president of the Association of Mayors of France and Advocate Counselor, it must take him half a day each month.'
src: b"Don d'Ubiquit\xc3\xa9 : Fran\xc3\xa7ois Baroin est d\xc3\xa9sormais conseiller \xc3\xa0 la Banque Barclays, maire, pr\xc3\xa9sident d'agglom\xc3\xa9ration, professeur \xc3\xa0 HEC Paris, pr\xc3\xa9sident de l'association des maires de France et avocat Conseiller, \xc3\xa7a doit lui prendre une demi journ\xc3\xa9e par mois."
Since we will fine-tune our model to perform a French-to-English translation task, we should format the inputs for instruction tuning. For example, we could format the translation task in this example like:
<start_of_turn>user
Translate French into English:
{src}<end_of_turn>
<start_of_turn>model
{dst}<end_of_turn>
The special tokens such as <start_of_turn>user, <start_of_turn>model and
<end_of_turn> are used for Gemma models. You can learn more from
https://ai.google.dev/gemma/docs/formatting
train_ds = train_ds.map(
lambda x: tf.strings.join(
[
"<start_of_turn>user\n",
"Translate French into English:\n",
x["src"],
"<end_of_turn>\n",
"<start_of_turn>model\n",
"Translation:\n",
x["dst"],
"<end_of_turn>",
]
)
)
examples = train_ds.take(3)
examples = examples.as_numpy_iterator()
for idx, example in enumerate(examples):
print(f"Example {idx}:")
print(example)
print()
Example 0:
b"<start_of_turn>user\nTranslate French into English:\nLe journal l'est peut-\xc3\xaatre, mais m\xc3\xaame moi qui suit de droite je les trouve limite de temps en temps...<end_of_turn>\n<start_of_turn>model\nTranslation:\nYep, serious...<end_of_turn>"
Example 1:
b"<start_of_turn>user\nTranslate French into English:\nEnfin je t'ai expliqu\xc3\xa9 dans quel cadre ce copypasta est pertinent : quand on nous dit padamalgame etc.<end_of_turn>\n<start_of_turn>model\nTranslation:\nFinally, I explained to you in what context this copy-pasting is relevant: when we are told padamalgame etc.<end_of_turn>"
Example 2:
b"<start_of_turn>user\nTranslate French into English:\nDon d'Ubiquit\xc3\xa9 : Fran\xc3\xa7ois Baroin est d\xc3\xa9sormais conseiller \xc3\xa0 la Banque Barclays, maire, pr\xc3\xa9sident d'agglom\xc3\xa9ration, professeur \xc3\xa0 HEC Paris, pr\xc3\xa9sident de l'association des maires de France et avocat Conseiller, \xc3\xa7a doit lui prendre une demi journ\xc3\xa9e par mois.<end_of_turn>\n<start_of_turn>model\nTranslation:\nGift of Ubiquity: Fran\xc3\xa7ois Baroin is now advisor to the Barclays Bank, mayor, president of the agglomeration, professor at HEC Paris, president of the Association of Mayors of France and Advocate Counselor, it must take him half a day each month.<end_of_turn>"
We will take a subset of the dataset for the purpose of this example.
train_ds = train_ds.batch(1).take(100)
Model
KerasHub provides implementations of many popular model architectures.
In this example, we will use GemmaCausalLM, an end-to-end Gemma model for
causal language modeling. A causal language model predicts the next token based
on previous tokens.
Note that sequence_length is set to 256 to speed up the fitting.
preprocessor = keras_hub.models.GemmaCausalLMPreprocessor.from_preset(
"gemma_1.1_instruct_2b_en", sequence_length=256
)
gemma_lm = keras_hub.models.GemmaCausalLM.from_preset(
"gemma_1.1_instruct_2b_en", preprocessor=preprocessor
)
gemma_lm.summary()
Preprocessor: "gemma_causal_lm_preprocessor"
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓ ┃ Tokenizer (type) ┃ Vocab # ┃ ┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┩ │ gemma_tokenizer (GemmaTokenizer) │ 256,000 │ └────────────────────────────────────────────────────┴─────────────────────────────────────────────────────┘
Model: "gemma_causal_lm"
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓ ┃ Layer (type) ┃ Output Shape ┃ Param # ┃ Connected to ┃ ┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━┩ │ padding_mask (InputLayer) │ (None, None) │ 0 │ - │ ├───────────────────────────────┼───────────────────────────┼─────────────────┼────────────────────────────┤ │ token_ids (InputLayer) │ (None, None) │ 0 │ - │ ├───────────────────────────────┼───────────────────────────┼─────────────────┼────────────────────────────┤ │ gemma_backbone │ (None, None, 2048) │ 2,506,172,416 │ padding_mask[0][0], │ │ (GemmaBackbone) │ │ │ token_ids[0][0] │ ├───────────────────────────────┼───────────────────────────┼─────────────────┼────────────────────────────┤ │ token_embedding │ (None, None, 256000) │ 524,288,000 │ gemma_backbone[0][0] │ │ (ReversibleEmbedding) │ │ │ │ └───────────────────────────────┴───────────────────────────┴─────────────────┴────────────────────────────┘
Total params: 2,506,172,416 (4.67 GB)
Trainable params: 2,506,172,416 (4.67 GB)
Non-trainable params: 0 (0.00 B)
LoRA Fine-tuning
What exactly is LoRA?
Low-rank adaptation (LoRA) is a parameter-efficient fine-tuning technique for LLMs. It freezes the weights of the LLM, and injects trainable rank-decomposition matrices. Let's understand this more clearly.
Assume we have an n x n pre-trained dense layer (or weight matrix), W0. We
initialize two dense layers, A and B, of shapes n x rank, and rank x n,
respectively. rank is much smaller than n. In the paper, values between 1
and 4 are shown to work well.
LoRA equation
The original equation is output = W0x + b0, where x is the input, W0 and
b0 are the weight matrix and bias terms of the original dense layer (frozen).
The LoRA equation is: output = W0x + b0 + BAx, where A and B are the
rank-decomposition matrices.
LoRA is based on the idea that updates to the weights of the pre-trained
language model have a low "intrinsic rank" since pre-trained language models are
over-parametrized. Predictive performance of full fine-tuning can be replicated
even by constraining W0's updates to low-rank decomposition matrices.
Number of trainable parameters
Let's do some quick math. Suppose n is 768, and rank is 4. W0 has
768 x 768 = 589,824 parameters, whereas the LoRA layers, A and B together
have 768 x 4 + 4 x 768 = 6,144 parameters. So, for the dense layer, we go
from 589,824 trainable parameters to 6,144 trainable parameters!
Why does LoRA reduce memory footprint?
Even though the total number of parameters increase (since we are adding LoRA layers), the memory footprint reduces, because the number of trainable parameters reduces. Let's dive deeper into this.
The memory usage of a model can be split into four parts:
- Model memory: This is the memory required to store the model weights. This will be slightly higher for LoRA than the original model.
- Forward pass memory: This mostly depends on batch size, sequence length, etc. We keep this constant for both models for a fair comparison.
- Backward pass memory: This is the memory required to store the gradients. Note that the gradients are computed only for the trainable parameters.
- Optimizer memory: This is the memory required to store the optimizer state. For example, the Adam optimizer stores the "1st moment vectors" and "2nd moment vectors" for the trainable parameters.
Since, with LoRA, there is a huge reduction in the number of trainable parameters, the optimizer memory and the memory required to store the gradients for LoRA is much less than the original model. This is where most of the memory savings happen.
Why is LoRA so popular?
- Reduces GPU memory usage;
- Faster training; and
- No additional inference latency.
When using KerasHub, we can enable LoRA with an one-line API:
enable_lora(rank=4)
From gemma_lm.summary(), we can see enabling LoRA reduces the number of
trainable parameters significantly (from 2.5 billion to 1.3 million).
gemma_lm.backbone.enable_lora(rank=4)
gemma_lm.summary()
Preprocessor: "gemma_causal_lm_preprocessor"
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓ ┃ Tokenizer (type) ┃ Vocab # ┃ ┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┩ │ gemma_tokenizer (GemmaTokenizer) │ 256,000 │ └────────────────────────────────────────────────────┴─────────────────────────────────────────────────────┘
Model: "gemma_causal_lm"
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓ ┃ Layer (type) ┃ Output Shape ┃ Param # ┃ Connected to ┃ ┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━┩ │ padding_mask (InputLayer) │ (None, None) │ 0 │ - │ ├───────────────────────────────┼───────────────────────────┼─────────────────┼────────────────────────────┤ │ token_ids (InputLayer) │ (None, None) │ 0 │ - │ ├───────────────────────────────┼───────────────────────────┼─────────────────┼────────────────────────────┤ │ gemma_backbone │ (None, None, 2048) │ 2,507,536,384 │ padding_mask[0][0], │ │ (GemmaBackbone) │ │ │ token_ids[0][0] │ ├───────────────────────────────┼───────────────────────────┼─────────────────┼────────────────────────────┤ │ token_embedding │ (None, None, 256000) │ 524,288,000 │ gemma_backbone[0][0] │ │ (ReversibleEmbedding) │ │ │ │ └───────────────────────────────┴───────────────────────────┴─────────────────┴────────────────────────────┘
Total params: 2,507,536,384 (4.67 GB)
Trainable params: 1,363,968 (2.60 MB)
Non-trainable params: 2,506,172,416 (4.67 GB)
Let's fine-tune the LoRA model.
# To save memory, use the SGD optimizer instead of the usual AdamW optimizer.
# For this specific example, SGD is more than enough.
optimizer = keras.optimizers.SGD(learning_rate=1e-4)
gemma_lm.compile(
loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
optimizer=optimizer,
weighted_metrics=[keras.metrics.SparseCategoricalAccuracy()],
)
gemma_lm.fit(train_ds, epochs=1)
After fine-tuning, responses will follow the instructions provided in the prompt.
template = (
"<start_of_turn>user\n"
"Translate French into English:\n"
"{inputs}"
"<end_of_turn>\n"
"<start_of_turn>model\n"
"Translation:\n"
)
prompt = template.format(inputs="Bonjour, je m'appelle Morgane.")
outputs = gemma_lm.generate(prompt, max_length=256)
print("Translation:\n", outputs.replace(prompt, ""))
Translation:
Hello, my name is Morgane.
Release memory.
del preprocessor
del gemma_lm
del optimizer
gc.collect()
QLoRA Fine-tuning
Quantized Low-Rank Adaptation (QLoRA) extends LoRA to enhance efficiency by quantizing the model weights from high precision data types, such as float32, to lower precision data types like int8. This leads to reduced memory usage and faster computation. The saved model weights are also much smaller.
Note that the QLoRA implementation here is a simplified version compared to the original. The differences are:
- The 4-bit NormalFloat format is not used because no backend supports it.
- No double quantization.
- No Paged optimizer.
To enable QLoRA in KerasHub, follow these steps:
- Instantiate the model.
- Quantize the weights using dynamic int8 quantization.
- Enable LoRA.
Steps 2 and 3 are achieved with one-line APIs:
quantize("int8")enable_lora(...)
preprocessor = keras_hub.models.GemmaCausalLMPreprocessor.from_preset(
"gemma_1.1_instruct_2b_en", sequence_length=256
)
gemma_lm = keras_hub.models.GemmaCausalLM.from_preset(
"gemma_1.1_instruct_2b_en", preprocessor=preprocessor
)
gemma_lm.quantize("int8")
gemma_lm.backbone.enable_lora(rank=4)
gemma_lm.summary()
Preprocessor: "gemma_causal_lm_preprocessor_1"
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓ ┃ Tokenizer (type) ┃ Vocab # ┃ ┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┩ │ gemma_tokenizer (GemmaTokenizer) │ 256,000 │ └────────────────────────────────────────────────────┴─────────────────────────────────────────────────────┘
Model: "gemma_causal_lm_1"
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓ ┃ Layer (type) ┃ Output Shape ┃ Param # ┃ Connected to ┃ ┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━┩ │ padding_mask (InputLayer) │ (None, None) │ 0 │ - │ ├───────────────────────────────┼───────────────────────────┼─────────────────┼────────────────────────────┤ │ token_ids (InputLayer) │ (None, None) │ 0 │ - │ ├───────────────────────────────┼───────────────────────────┼─────────────────┼────────────────────────────┤ │ gemma_backbone │ (None, None, 2048) │ 2,508,502,016 │ padding_mask[0][0], │ │ (GemmaBackbone) │ │ │ token_ids[0][0] │ ├───────────────────────────────┼───────────────────────────┼─────────────────┼────────────────────────────┤ │ token_embedding │ (None, None, 256000) │ 524,544,000 │ gemma_backbone[0][0] │ │ (ReversibleEmbedding) │ │ │ │ └───────────────────────────────┴───────────────────────────┴─────────────────┴────────────────────────────┘
Total params: 2,508,502,016 (2.34 GB)
Trainable params: 1,363,968 (2.60 MB)
Non-trainable params: 2,507,138,048 (2.34 GB)
Let's fine-tune the QLoRA model.
If you are using a device with int8 acceleration support, you should see an improvement in the training speed.
optimizer = keras.optimizers.SGD(learning_rate=1e-4)
gemma_lm.compile(
loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
optimizer=optimizer,
weighted_metrics=[keras.metrics.SparseCategoricalAccuracy()],
)
gemma_lm.fit(train_ds, epochs=1)
You should get a similar output with QLoRA fine-tuning.
prompt = template.format(inputs="Bonjour, je m'appelle Morgane.")
outputs = gemma_lm.generate(prompt, max_length=256)
print("Translation:\n", outputs.replace(prompt, ""))
Translation:
Hello, my name is Morgane.
And we're all done!
Note that for demonstration purposes, this example fine-tunes the model on a small subset of the dataset for just one epoch and with a low LoRA rank value. To get better responses from the fine-tuned model, you can experiment with:
- Increasing the size of the fine-tuning dataset.
- Training for more steps (epochs).
- Setting a higher LoRA rank.
- Modifying the hyperparameter values such as
learning_rateandweight_decay.