badam 1.4

Package which implements the algorithm proposed by "BAdam: A Memory Efficient Full Parameter Training Method for Large Language Models".

BAdam

The implementation for BAdam: A Memory Efficient Full Parameter Optimization Method for Large Language Models. This paper presents an algorithm named BAdam, which finetunes Llama 2-7b and Llama 3-8B using a single RTX3090 with Adam's update rule and mixed precision training. The core idea of BAdam is to sequentially solve block coordinate optimization sub-problems. From the implementation perspective, the algorithm runs Adam's update on a small portition (usually one single transformer layer) of the parameters, thereby requires much less memory in comparison to full parameter Adam finetuning. Using BAdam only requires one line modification of the original code.

Method	Minimum Memory	Actual Memory Cost (Llama 3-8B)	Actual Memory Cost (Llama 2-7B)
Adam	$18M$	144 GB+	122.8 GB+
BAdam	$2M + \frac{16M}{D}$	23.5 GB	21.8 GB

Table 1: Comparison of Methods. $M$ stands for the number of model's parameters in billion and $D$ is the number of blocks used in BAdam. See Table 2 in paper for detailed analysis on memory consumption.

Method	Llama 3-8b	Llama 2-7b
Pretrained model	5.46	3.93
LoRA	6.41	4.86
BAdam	6.67	5.21

Table 2: MT bench score. The model is instruction finetuned on Alpaca-GPT4 dataset using a single RTX3090. BAdam consistently outperforms LoRA in MT bench under various evaluation models.

One can also apply BAdam for larger models with size such as 13B, 22B, 30B, and 70B. The memory consumption can be estimated to be $2M + \frac{16M}{D}$ (GB), plus some additional memory consumption for gradient checkpointed activations and system use like PyTorch's pre-allocation, etc (minor part). When using model parallelism with $N$ GPUs, the memory cost can be estimated by $\frac{2M + 16M/D}{N}$ (GB), plus the additional communication buffers.

Change log

[24/09/26] BAdam has been accepted by NeurIPS, 2024!

[24/06/16] We support model parallel using Deepspeed ZeRO-3 now!

[24/04/16] Our algorithm has been added to LLaMA-Factory. We would like to express our gratitude to their efforts on integrating BAdam!

[24/04/12] Add LoRA module detection. Make BlockOptimizer compatible with lr scheduler.

Table of Contents

• Environment Setup
• Usage of BAdam
  • Partition by Module (A Single GPU)
  • Partition by Module (Model Parallel)
  • Partition by Parameter Ratio
  • Hyperparameter Suggestion
• Run Paper Experiment

Setup

To install BAdam from Pypi, one can run:

pip install badam

One may also choose to build from source by the following steps:

git clone [e-mail]:Ledzy/BAdam.git
cd BAdam
pip install -e .

For those who are interested in reproducing the results in paper, please follow the steps below to setup environment:

conda create -n badam python=3.10
conda activate badam
pip install -r requirements.txt

Usage of BAdam

Partition by Module (A Single GPU)

BAdam uses mixed-precision training, make sure that the model is loaded in float16 precision for memory saving. One can simply add one line of code that wraps the original optimizer to use BAdam.

from badam import BlockOptimizer

# before training, add this line to wrap the original optimizer
optimizer = BlockOptimizer(
    base_optimizer=original_optimizer, # can be any torch.Optimizer
    named_parameters_list=list(model.named_parameters()), 
    switch_block_every=100, # switch to the new block every 50 updates, the $K$ Adam steps in paper. It can be set adaptively by $K = n/(BD)$, where $n$ is the number of training data points, $B$ is the batch size, and $D$ is the number of blocks in BAdam; see "Hyperparameter Suggestion" section for a detailed explaination about setting this hyperparameter. 
    switch_mode="random", # update order of blocks, one can choose "random" (random reshuffling update order), "ascending" (update from input layer to output layer), or "descending" (update from output layer to input layer). The default is "random".
    verbose=2 # information level, will print trainable parameters when setting to 2
)

The above code automatically creates a block partition according to model.named_parameters. Specifically, it treates each transformer layer module as a single block. For instance, for Llama 3-8B, the block partition ($D = 32$) will be

block 1: model.layers.0.
block 2: model.layers.1.
...
block 32: model.layers.31.

By default, the embedding layer and language modeling head is not included in the training blocks. One can add them as two additional blocks by setting include_embedding=True, include_lm_head=True.

Click to see more partition strategies and example code

One can also specify their own block list for the block optimizer. This can be achieved by adjusting the block_prefix_list argument. For instance, the following code snippets creat block partitions by self_attn and mlp modules (i.e., D = 32 * 2 = 64 for Llama 3-8B), and matrix modules (i.e., D = 32 * 7=224 for Llama 3-8B), respectively, which helps further reduce the memory cost:

# block partition by self_attn and mlp modules
block_prefix_list = []
for i in range(32):
    layer_prefix = [
        [f"model.layers.{i}.self_attn."],
        [f"model.layers.{i}.mlp."],
    ]
    block_prefix_list.extend(layer_prefix)

optimizer = BlockOptimizer(
    base_optimizer=original_optimizer,
    named_parameters_list=list(model.named_parameters()), 
    switch_block_every=100,
    switch_mode="random",
    verbose=2,
    block_prefix_list=block_prefix_list # set the block list
)

#block partition by matrix modules
block_prefix_list = []
for i in range(32):
    layer_prefix = [
        [f"model.layers.{i}.self_attn.q_proj."],
        [f"model.layers.{i}.self_attn.k_proj."],
        [f"model.layers.{i}.self_attn.v_proj."],
        [f"model.layers.{i}.self_attn.o_proj."],
        [f"model.layers.{i}.mlp.gate_proj."],
        [f"model.layers.{i}.mlp.up_proj."],
        [f"model.layers.{i}.mlp.down_proj."],
    ]
    block_prefix_list.extend(layer_prefix)

optimizer = BlockOptimizer(
    base_optimizer=original_optimizer,
    named_parameters_list=list(model.named_parameters_list), 
    switch_block_every=100,
    switch_mode="random",
    verbose=2,
    block_prefix_list=block_prefix_list # set the block list
)

We have tested that block partition by self_attn and mlp modules achieves a MT-bench score 6.65 for finetuning Llama 3-8B. This score matches that (6.67) achieved by block partition by transformer layer modules, while further reduces the memory cost.

Important Notes:

• When setting block partition, one should be careful with the downstream task. Some tasks has randomly initialized classification layers, such as the SuperGLUE where the task_dict and pooler layers are randomly initialized. In this case, make sure to train these layers first, or set it to be trainable through the whole time. To set modules to be trainable through the whole training process, one can use active_modules argument, e.g., set active_modules=["model.task_dict.", "model.pooler."] when create the BlockOptimizer. Note that randomly initialized layers are usually the last layer, so updating these layers will only introduce negligible additional BP time. We thus suggest to always set the last classification layer to be trainable when the memory is permitted, if it is randomly initialized.
• The parameters that are not included in block_prefix_list will be inactive (freezed) through the whole training procedure.
• When setting prefix, it is suggested to include a . at the end. For example, it is preferred to use model.layers.1. instead of model.layers.1, as the later one includes the layer 10, 11, ..., 19 as well (since they have the same prefix).

Partition by Module (Model Parallel)

We support the model parallel offered by deepspeed ZeRO-3. It partitions the model, gradient, and optimizer states across different GPUs so that one can train large models (e.g., Llama 3-70B) that cannot be fit into a single GPU. Given $N$ GPUs, the per GPU memory cost can be estimated by $\frac{2M + 16M/D}{N}$ (GB), plus the additional cost for communication buffer and temporary parameter gathering buffer arised during forward/backward. These buffer sizes can be configurated manually and determines the efficieny of the communication system.

Click to see instructions for model parallelism

To use ZeRO-3, one needs to set ds_zero3_enabled=True when initializing the BlockOptimizer. Then, set block_optimizer.ds_optimizer = ds_optimizer after calling deepspeed.initialize.

from badam import BlockOptimizer

optimizer = BlockOptimizer(
    ...,
    ds_zero3_enabled=True # set it to True
)

model, ds_optimizer = deepspeed.initialize(model=model, optimizer=optimizer, ...)

# create the reference to the ds_optimizer, for the purpose of setup ZeRO-3's environment
optimizer.ds_optimizer = ds_optimizer

When using huggingface Trainer to control the workflow, accessing ds_optimizer is not direct. One can add the BAdamCallback which automatically handles the reference to ds_optimizer:

from badam.utils import BAdamCallback

callbacks = original_callbacks.append(BAdamCallback) # add the callback
trainer = YourTrainerClass(
    ...,
    callbacks=callbacks
)

The model parallelism results in noticable overhead due to the communication cost. In particular, we empirically observe about 3 times overhead when training Llama 3-8B with 4 RTX3090 GPUs (without NVLink) using ZeRO-3, in comparison to using a single GPU, under the same per_device_batch_size. Fortunately, one may use a larger per_device_batch_size to accelerate the training speed as ZeRO-3 greatly reduces the per GPU memory cost.

Make sure to use accelerate config to configurate the distributed training and then use proper command to launch your script in a distributed way, such as accelerate launch and deepspeed.

Partition by Parameter Ratio

Instead of partitioning block by the model's parameter, an alternative choice is to train all the parameters simultaneously with a fixed ratio. For instance, we can train 5% parameters of every transformer layer. Namely, each active block contains 5% parameters from every transformer layer. In this sense, the feature extractor of every layer are jointly trained, which may be preferred in certain scenarios. However, training a block consisting of parameters coming from all the transformer layers may lose partly the benefit of BP time saving of BAdam.

Click to see example code and instructions

from badam import BlockOptimizerRatio

optimizer = BlockOptimizerRatio(
    param_groups=param_groups, # param_group of torch.Optimizer, the same as the original optimizer
    named_parameters_list=list(self.model.named_parameters()),
    switch_every=100, # switch to the new block every 100 updates
    update_ratio=0.1, # ratio of trainable weight for each parameter
    mask_mode = "adjacent", # choices: ["adjacent", "scatter"], see Note below for more explanation
    lr=1e-6,
    betas=(0.9, 0.999), # betas for Adam update
    eps=1e-8, # eps of Adam update
)

Currently, the BlockOptimizerRatio only supports the Adam update. The repository is still under active development.

Notes:

• The mask_mode indicates how should the trainable parameter distribute across a parameter. mask_mode=adjacent indicates that the trainable parameters are adjacent to each other, while mask_mode=scatter indicates that trainable parameters are randomly choosed from the weight. For instance, considering optimizing a $10 \times 10$ matrix with update_ratio=0.1, setting mask_mode=adjacent will let parameters of the same row be the same block, and mask_mode=scatter means randomly choose 10 trainable parameters from the matrix.
• By default, BlockOptimizerRatio does not update embedding layer, since in principle the embedding vectors of the tokens that are included in the training samples should be updated, while randomly freeze embedding parameters makes the update imbalanced. One can set include_embedding=True to include it for experimental purpose.
• For BlockOptimizerRatio, we notice that setting mask_mode = "adjacent" usually performs the best; we leave the study of mask_mode as a future work. The convergence speed is highly positively related to the update_ratio, so we suggest to choose it as high as possible when the memory is permitted.
• The gradient and optimizer states are stored in sparse tensor format. The update rule is exactly the same as the BlockOptimizer: run Adam update on current active block for switch_every steps, and then switch to next block.
• Currently, the operation of sparsifing the gradient causes noticable overhead, which inevitably slow down the training. We leave the acceleration as a future work.

Hyperparameter Suggestion

• Choice of the switch_block_every. Compared to Adam, our BAdam only introduces one additional hyperparameter, i.e., the switch_block_every (the K Adam steps in paper). It determines how many Adam steps we perform for each active block before switching to the next one. Fortunately, this hyperparameter can be set adaptively. Ideally, we expect to balance the data usage for each block in every epoch. This gives a natural choice of switch_block_every = $\frac{n}{BD}$ (rounding to the nearest integer if it is a fractional), where $n$ is the number of training data points, $B$ is the effective batch size, and $D$ is the number of blocks in BAdam. Using such a setting ensures that after one block-epoch, all the training data points are equally distributed to the $D$ blocks for training. Meanwhile, to achieve sufficient decrease for each block coordinate descent subproblem and fully utilize the advantage of mixed precision training for reducing rounding error, the switch frequency should not be too small. Additionally, too large switch frequency may over-optimize one block before moving to others. We notice that setting switch_block_every = $\min(\max(\frac{n}{BD}, 50),100)$ usually yields fast convergence speed on both training loss and validation loss.

Run Paper Experiment

Llama 3-8B and Llama 2-7B on Alpaca-GPT4

Our implementation of finetuning Llama 3 and Llama 2 is based on Llama Factory. This repository mainly serves as the purpose for reproducing our paper's results. For better support on advanced algorithmic features, we suggest to use the latest version of Llama Factory.

For the experiment of finetuning Llama-2 7b on Alpaca-GPT4 dataset, change the working directory to llama:

cd llama-alpaca

Here is a sample command for running the code:

CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
    --stage sft \
    --model_name_or_path meta-llama/Llama-2-7b-hf \
    --do_train \
    --dataset alpaca_gpt4_en \
    --template default \
    --finetuning_type block \
    --output_dir ./outputs/llama2-7b \
    --overwrite_cache \
    --per_device_train_batch_size 2 \
    --per_device_eval_batch_size 2 \
    --gradient_accumulation_steps 8 \
    --lr_scheduler_type cosine \
    --logging_steps 1 \
    --save_steps 1000 \
    --val_size 500 \
    --eval_steps 20 \
    --evaluation_strategy steps \
    --learning_rate 1e-6 \
    --num_train_epochs 3 \
    --overwrite_output_dir \
    --plot_loss \
    --switch_block_every 100 \
    --switch_mode random \
    --bf16 True

To finetune Llama 3-8B, one can set --model_name_or_path meta-llama/Meta-Llama-3-8B. We use learning rate 1e-6 for Llama 3-8B and learning rate 1e-5 for Llama 2-7B, respectively. It is important to note that the favorable learning rate may vary for different models and datasets.

Notes on arguments:

• --stage: Currently we only implement the sft.
• --finetuning_type: Options: (block, full, lora, sparse)
• --switch_mode: How to order the block update. Options: (random, ascending, descending).
• --switch_block_every: Switch block frequency; see "Hyperparameter Suggestion" for how to set this hyperparamter.
• The above sample command is different from the hyperparameters settings in paper, while this version is more efficient. We will update our paper later.

RoBERTa-large on SuperGLUE

Our implementation for finetuning RoBERTa-large on superGLUE is based on jiant. To run the code, go to directory roberta-superglue first:

cd roberta-superglue

Before training the model, download the dataset using the following bash script. Adjust the script to download the required dataset.

EXP_DIR=./content/exp

python jiant/scripts/download_data/runscript.py \
    download \
    --tasks copa \
    --output_path ${EXP_DIR}/tasks

The finetuning command has the following form:

CUDA_VISIBLE_DEVICES=0 python badam_ft.py \
    --task_name boolq \
    --num_train_epochs 32 \
    --eval_every_steps 100 \
    --use_block_optim \
    --switch_every 100 \
    --switch_mode ascending \
    --train_batch_size 16 \
    --train_last_layer \
    --hf_pretrained_model_name FacebookAI/roberta-large

Notes on arguments:

• --task_name: Options: boolq, wic, wsc, rte, multirc, copa
• --use_block_optim: Whether to use BlockOptimizer or not. Remove this argument leads to full parameter Adam update. Change to --use_sparse_optim: to use BlockOptimizerRatio.
• --train_last_layer: Whether to train the last layer through the finetuning. For the superGLUE task, the last layer is randomly initialized and thereby needs to be trained first or being trainable through the whole training.

Apache License Version 2.0, January 2004 http://www.apache.org/licenses/

TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION

1. Definitions.

   "License" shall mean the terms and conditions for use, reproduction, and distribution as defined by Sections 1 through 9 of this document.

   "Licensor" shall mean the copyright owner or entity authorized by the copyright owner that is granting the License.

   "Legal Entity" shall mean the union of the acting entity and all other entities that control, are controlled by, or are under common control with that entity. For the purposes of this definition, "control" means (i) the power, direct or indirect, to cause the direction or management of such entity, whether by contract or otherwise, or (ii) ownership of fifty percent (50%) or more of the outstanding shares, or (iii) beneficial ownership of such entity.

   "You" (or "Your") shall mean an individual or Legal Entity exercising permissions granted by this License.

   "Source" form shall mean the preferred form for making modifications, including but not limited to software source code, documentation source, and configuration files.

   "Object" form shall mean any form resulting from mechanical transformation or translation of a Source form, including but not limited to compiled object code, generated documentation, and conversions to other media types.

   "Work" shall mean the work of authorship, whether in Source or Object form, made available under the License, as indicated by a copyright notice that is included in or attached to the work (an example is provided in the Appendix below).

   "Derivative Works" shall mean any work, whether in Source or Object form, that is based on (or derived from) the Work and for which the editorial revisions, annotations, elaborations, or other modifications represent, as a whole, an original work of authorship. For the purposes of this License, Derivative Works shall not include works that remain separable from, or merely link (or bind by name) to the interfaces of, the Work and Derivative Works thereof.

   "Contribution" shall mean any work of authorship, including the original version of the Work and any modifications or additions to that Work or Derivative Works thereof, that is intentionally submitted to Licensor for inclusion in the Work by the copyright owner or by an individual or Legal Entity authorized to submit on behalf of the copyright owner. For the purposes of this definition, "submitted" means any form of electronic, verbal, or written communication sent to the Licensor or its representatives, including but not limited to communication on electronic mailing lists, source code control systems, and issue tracking systems that are managed by, or on behalf of, the Licensor for the purpose of discussing and improving the Work, but excluding communication that is conspicuously marked or otherwise designated in writing by the copyright owner as "Not a Contribution."

   "Contributor" shall mean Licensor and any individual or Legal Entity on behalf of whom a Contribution has been received by Licensor and subsequently incorporated within the Work.

2. Grant of Copyright License. Subject to the terms and conditions of this License, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare Derivative Works of, publicly display, publicly perform, sublicense, and distribute the Work and such Derivative Works in Source or Object form.

3. Grant of Patent License. Subject to the terms and conditions of this License, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable (except as stated in this section) patent license to make, have made, use, offer to sell, sell, import, and otherwise transfer the Work, where such license applies only to those patent claims licensable by such Contributor that are necessarily infringed by their Contribution(s) alone or by combination of their Contribution(s) with the Work to which such Contribution(s) was submitted. If You institute patent litigation against any entity (including a cross-claim or counterclaim in a lawsuit) alleging that the Work or a Contribution incorporated within the Work constitutes direct or contributory patent infringement, then any patent licenses granted to You under this License for that Work shall terminate as of the date such litigation is filed.

4. Redistribution. You may reproduce and distribute copies of the Work or Derivative Works thereof in any medium, with or without modifications, and in Source or Object form, provided that You meet the following conditions:

   (a) You must give any other recipients of the Work or Derivative Works a copy of this License; and

   (b) You must cause any modified files to carry prominent notices stating that You changed the files; and

   (c) You must retain, in the Source form of any Derivative Works that You distribute, all copyright, patent, trademark, and attribution notices from the Source form of the Work, excluding those notices that do not pertain to any part of the Derivative Works; and

   (d) If the Work includes a "NOTICE" text file as part of its distribution, then any Derivative Works that You distribute must include a readable copy of the attribution notices contained within such NOTICE file, excluding those notices that do not pertain to any part of the Derivative Works, in at least one of the following places: within a NOTICE text file distributed as part of the Derivative Works; within the Source form or documentation, if provided along with the Derivative Works; or, within a display generated by the Derivative Works, if and wherever such third-party notices normally appear. The contents of the NOTICE file are for informational purposes only and do not modify the License. You may add Your own attribution notices within Derivative Works that You distribute, alongside or as an addendum to the NOTICE text from the Work, provided that such additional attribution notices cannot be construed as modifying the License.

   You may add Your own copyright statement to Your modifications and may provide additional or different license terms and conditions for use, reproduction, or distribution of Your modifications, or for any such Derivative Works as a whole, provided Your use, reproduction, and distribution of the Work otherwise complies with the conditions stated in this License.

5. Submission of Contributions. Unless You explicitly state otherwise, any Contribution intentionally submitted for inclusion in the Work by You to the Licensor shall be under the terms and conditions of this License, without any additional terms or conditions. Notwithstanding the above, nothing herein shall supersede or modify the terms of any separate license agreement you may have executed with Licensor regarding such Contributions.

6. Trademarks. This License does not grant permission to use the trade names, trademarks, service marks, or product names of the Licensor, except as required for reasonable and customary use in describing the origin of the Work and reproducing the content of the NOTICE file.

7. Disclaimer of Warranty. Unless required by applicable law or agreed to in writing, Licensor provides the Work (and each Contributor provides its Contributions) on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied, including, without limitation, any warranties or conditions of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A PARTICULAR PURPOSE. You are solely responsible for determining the appropriateness of using or redistributing the Work and assume any risks associated with Your exercise of permissions under this License.

8. Limitation of Liability. In no event and under no legal theory, whether in tort (including negligence), contract, or otherwise, unless required by applicable law (such as deliberate and grossly negligent acts) or agreed to in writing, shall any Contributor be liable to You for damages, including any direct, indirect, special, incidental, or consequential damages of any character arising as a result of this License or out of the use or inability to use the Work (including but not limited to damages for loss of goodwill, work stoppage, computer failure or malfunction, or any and all other commercial damages or losses), even if such Contributor has been advised of the possibility of such damages.

9. Accepting Warranty or Additional Liability. While redistributing the Work or Derivative Works thereof, You may choose to offer, and charge a fee for, acceptance of support, warranty, indemnity, or other liability obligations and/or rights consistent with this License. However, in accepting such obligations, You may act only on Your own behalf and on Your sole responsibility, not on behalf of any other Contributor, and only if You agree to indemnify, defend, and hold each Contributor harmless for any liability incurred by, or claims asserted against, such Contributor by reason of your accepting any such warranty or additional liability.

END OF TERMS AND CONDITIONS

APPENDIX: How to apply the Apache License to your work.

  To apply the Apache License to your work, attach the following
  boilerplate notice, with the fields enclosed by brackets "[]"
  replaced with your own identifying information. (Don't include
  the brackets!)  The text should be enclosed in the appropriate
  comment syntax for the file format. We also recommend that a
  file or class name and description of purpose be included on the
  same "printed page" as the copyright notice for easier
  identification within third-party archives.

Copyright [2024] [Qijun Luo]

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.