Training And Fine Tuning A Large Language Model LLM
Estimated time: 30 mins to 1 hour
This tutorial teaches you how to use the Cerebras Wafer-Scale cluster to train, fine-tune, and evaluate a Large Language Model (LLM) and visualize the results using TensorBoard. The model you will use is the GPT-3 111M with the WikiText-2 dataset. You need access to Cerebras’s Wafer-Scale cluster to train the model and fine-tune it from an existing checkpoint. Later, you will evaluate the models and visualize the results using TensorBoard. Finally, you will learn how to port the models to Hugging Face to generate outputs.
If you are interested in using a different dataset, you can still use the create HDF5 dataset tool from Cerebras Model Zoo used in step 2. In addition, if you are using a larger GPT model, change the YAML configuration used in steps 4 and 5. Here are some examples of GPT3 model configurations from Cerebras Model Zoo GPT and other GPT-3 configurations.
2. Clone the Cerebras Model Zoo inside the demo folder, a repository with reference models and tools to run on the Cerebras Wafer-Scale cluster.
Output
At this point, you should have a folder called
1. Download the WikiText-2 raw dataset by first creating a
Output
Decompress the zip file.
Output
Remove the zip file.
2. Inside the
Output
The
The
Note that now you should be in the
Output
Output
2. Now that your
Output
The
Table 2 HDF5 dataset flags#
After running the Python script, you should have an output directory:
Output
3. Prepare the data in the validation set by changing the
You should have an output directory:
Output
2. Set the absolute path directory for
3. With the absolute path set, modify the
4. Modify the
5. To shorten the runtime, modify the number of
With the new YAML configuration, you are ready to launch the training job.
Alternately, you can use screen, a terminal multiplex, so that your training does not stop if you lose access to the terminal. For this, start a new screen session called
7. Run the
The following list describes the flags and their description:
Table 3 GPT-3 model run.py training flags#
8. Detach the screen using
Output
where
Output
If
Output
Once you have the checkpoint, you are ready to launch the training job. For this, you can start a new screen session or reuse the screen session used during training from scratch. To reattached the
You will now run the same
Table 4 GPT-3 model run.py finetuning flags#
Once the training job is over, you will find inside the
Output
The
Output
Table 5 GPT-3 model run.py evaluation flags#
Given that you have already trained the model in this demo, you will use the
1. To launch evaluation, reattach the
2. Inside this screen session, if not already active, activate the Cerebras virtual environment
After the validation job is complete, you will find additional files inside the
After the validation job is complete, you will find additional files inside the
After launching TensorBoard, you will get a URL link to access the results in your local network.
Fig. 8 Metrics logged in GPT implementation in Cerebras Model Zoo#
Fig. 9 Value of training and validation loss after 10 steps for model train from scratch (
Prerequisites
You have access to the user node in the Cerebras Wafer-Scale cluster. Contact your sys admin if you face any issues in the system configuration.Setting up the Cerebras Model Zoo
For convenience, you will set up the data and code inside a parent folder calleddemo - a directory that you will have to create yourself. In addition, you will define an environment variable called PARENT_CS to return to this parent directory at the anytime.
1. Create a parent directory, demo, to include all the data, code, and checkpoints. Export an environment variable PARENT_CS with the full path to the parent directory. This environment variable will be helpful when pointing to the absolute path during the execution.
modelzoo inside the demo parent directory.
Preparing data for the Cerebras Wafer-Scale cluster
Cerebras Model Zoo contains customized scripts to download and prepare OWT and Pile datasets. If interested, refer to them when training GPT-style models with OWT or Pile datasets. Hierarchical Data Format, Version 5(HDF5) files provide an efficient way of loading data in the Cerebras’s Wafer-Scale cluster. The Cerebras Model Zoo contains a
create_hdf5_dataset.py script that tokenizes the input documents and creates token IDs, labels, and attention masks. For more information, refer to hdf5_dataset_gpt.data folder, and in the folder, download the dataset’s zip file containing the dataset.
wikitext-2-raw-v1 folder, you will find three files corresponding to train, validation, and test splits. Explore these files to understand the raw dataset format. The data is in txt format.
create_hdf5_dataset.py script expects all the files in the same folder to belong to the same split. Therefore, create three folders - train, test, and valid, and move the corresponding files to each folder. In addition, add a .txt file extension to each file since each file is in a raw format.
create_hdf5_dataset.py script also parses file types .json, .jsonl.zst, and .jsonl.zst.tar. For more information on the input format, refer to hdf5_dataset_gpt
Preparing the data in HDF5 format
1. Before you prepare the data in an HDF5 format, set up a Python virtual environment with all the dependencies needed for preprocessing. Create this virtual environment in the parent directorydemo. For more informaion on HDF5 format, click here.
(venv_cerebras_pt) environment.
venv_cerebras_pt virtual environment has been set up and activated, launch the create_hdf5_dataset.py script. This script resides in the modelzoo folder in the Cerebras Model Zoo repository. Launch this script for every data split you intend to use. In this case, you will preprocess the train and valid splits and save the preprocessed data in an HDF5 format in the data/wikitext-2-hdf5/ folder.
Prepare the data for the train set:
create_hdf5_dataset.py is instrumented with multiple flags. Our version of the Python script uses the following flags:
| Flag | Description |
|---|---|
LMData | For processing language modelling datasets that are in .jsonl or .txt format. |
--params ./configs/autoregressive_lm_preprocessing.yaml | Path to YAML config file for setting dataset preprocessing parameters. Optional alternative for providing command line arguments. |
--input_dir data/wikitext-2-raw/train/ | Folder where all the .txt files are located. In this case, you only have one txt file |
--output_dir data/wikitext-2-hdf5/train/ | Destination folder |
--files_per_record 100 | Files per record is set to 100 in comparison to the default setting, which is 50,000. Given that there are 1180 training samples, the number of files per record should be smaller than the total number of samples. After preprocessing, the total number of samples will be floor(#samples/files_per_record) * files_per_record |
data/wikitext-2-hdf5/train/. The following list shows the content of the output directory:
--input_dir to data/wikitext-2-raw/valid/ and --output_dir to data/wikitext-2-hdf5/valid/:
data/wikitext-2-hdf5/valid/. The following list shows the content of the output directory:
NoteThe
create_hdf5_dataset.py script only tokenizes your dataset. Perform data cleaning and shuffling before preparing the data in the HDF5 format, as it depends on the quality of your dataset. Additional resources available in Cerebras Model Zoo can be found in Data processing and dataloaders.Training GPT-3 111M model from scratch
1. Create a copy of the Cerebras-GPT 111M YAML configuration file to point to the WikiText-2 dataset preprocessed in step 2. Make sure that thevenv_cerebras_pt is active.
data/wikitext-2-hdf5/train. You can obtain the absolute path using realpath, or appending data/wikitext-2-hdf5/train/ to the absolute path in $PARENT_CS.
custom_config_GPT111M.yaml file.
data_dir flag for training and evaluation inputs:
max_steps for training, the eval_steps for evaluation, and the checkpoint frequency.
train_wsc.
screen -S train_wsc
6. Inside this screen session, if not already active, activate the Cerebras virtual environment venv_cerebras_pt.
run.py script file associated with the GPT-3 models in Cerebras Model Zoo.
Each model in Cerebras Model Zoo contains a run.py script instrumented to easily launch training and evaluation in the Cerebras Wafer-Scale cluster and other AI accelerators. To learn more on how to launch a training, visit Launch your job, and to view the run.py code for the GPT-3 model, visit here.
| Flag | Description |
|---|---|
--params custom_config_GPT111M.yaml | Points to the YAML configuration file you customed with the training data paths |
--num_csx=1 | Number of CS-X systems used in the run |
--model_dir train_from_scratch_GPT111M | New directory containing all the checkpoints and logging information |
--mode train | Specifies that you are training the model as opposed to evaluation |
--mount_dirs $PARENT_CS $PARENT_CS/modelzoo | Mounts directories to the Cerebras Wafer-Scale cluster. In this case, all the data and code is in the parent directory demo (with absolute path $PARENT_CS) |
--python_paths $PARENT_CS/modelzoo | Enables adding the Cerebras Model Zoo to the list of paths to be exported to PYTHONPATH |
--job_labels model=GPT111M | A list of equal-sign-separated key value pairs served as job label, to query using csctl. |
NoteOnce you have submitted your job to execute in the Cerebras Wafer-Scale cluster, you can track the progress or kill your job using the csctl tool. You can also monitor the performance using a Grafana dashboard
CTRL+A D and reattach the screen using screen -r train_wsc.
Once the training job is complete, you will find inside the train_from_scratch_GPT111M folder:
latest_compile and cerebras_logs contain Cerebras-specific information collected while compiling and executing the model, checkpoint_10.mdl is the checkpoint saved after ten steps, train contains the metrics logged during execution, and the YAML configuration file, and run_<data>_######.log contains the command line output during execution.
Validating during Training
To enable periodic validation during training, set the--mode flag to train_and_eval and use the following additional flags in the runconfig to configure the run:
| Runconfig Param | Description |
|---|---|
eval_frequency | Frequency at which the evaluation loop runs during training |
eval_steps | Number of steps to run in each evaluation loop |
eval_steps isn’t specified, it will loop over the entire evaluation dataloader if the dataloader has a length. Otherwise, it will error out.
Important considerations
1. Checkpoint and Evaluation Frequency: When usingtrain_and_eval mode, make sure that checkpoint_steps is either a multiple of eval_frequency or vice versa, so that the frequency at which evaluations occurs lines up with the frequency at which checkpoints are taken.
-
Example 1:
eval_frequency=5,checkpoint_steps=5Take a checkpoint every 5 steps and run an evaluation every 5 steps. -
Example 2:
eval_frequency=5,checkpoint_steps=10Take a checkpoint every 10 steps, but perform an evaluation every 5 steps. Note that at steps 5, 15, 25, … there are no checkpoints available, but evaluation results are available. -
Example 3:
eval_frequency=10,checkpoint_steps=5Take a checkpoint every 5 steps, but perform an evaluation every 10 steps.
checkpoint_steps is set to a positive number, regardless of whether the num_steps is a multiple of checkpoint_steps. Additionally, in the “Train and Eval” mode, an evaluation will always run at the final step, regardless of whether the num_steps is a multiple of the eval_frequency.
2. Evaluation Code in the Model:
Ensure that evaluation-specific code(such as eval metrics) is guarded by the PyTorch module’s training flag to distinguish between training and evaluation.
For example:
Limitations of Train and Eval mode
1. Single Dataloader: Only a single training dataloader and a single evaluation dataloader can be used in the same configuration file. To evaluate with multiple datasets, they must be concatenated into a single dataloader. 2. Eval Harness: Running the Eval Harness as part oftrain_and_eval mode is not yet supported.
3. Runconfig Limitations: Settings in runconfig (such as precision_opt_level or POL) affect both training and evaluation graphs without nested configurations. This means that the settings specified in runconfig apply to both training and evaluation (e.g., POL). Note, however, that micro_batch_size is part of train_input / eval_input. Therefore, it is possible to specify different micro-batch size settings for the train and eval mode provided that the micro_batch_size for evaluation is less than or equal to the micro_batch_size for training.
4. Cluster Configuration: It’s not currently possible to run training and evaluation on separate portions of a cluster. Separate run.py jobs must be manually executed for this.
Fine-tuning using checkpoints from Cerebras-GPT model
1. Download the Cerebras-GPT 111M checkpoint compatible with the Cerebras Wafer-Scale Cluster from Cerebras Model Zoo. You can find all the Cerebras-GPT checkpoints in the Cerebras-GPT site in Cerebras Model Zoo.train_wsc screen, use
screen -r train_wsc
2. Inside this screen session, if not already active, you will activate the Cerebras virtual environment venv_cerebras_pt.
run.py script used during the training from scratch with new flags.
3. Change the following flags:
| Flag | Description |
|---|---|
--model_dir finetune_GPT111M | Specifies a different model directory to save checkpoints and logging information |
--checkpoint_path Cerebras-checkpoint/cerebras-gpt-dense-111m -sp-checkpoint_final.mdl | Specifies the path where the Cerebras-GPT 111M model is. You will be using this checkpoint to initialize the model weights |
--load_checkpoint_states="model" | Flag used in conjunction with checkpoint_path, to enforce resetting of optimizer states and training steps after loading a given checkpoint. By setting this flag, all the model weights are initialized from checkpoint provided by checkpoint_path, training starts from step 0, and optimizer states present in the checkpoint are ignored. Useful for fine-tuning runs on different tasks (e.g., classification, Q&A, etc.) where weights from a pre-trained model trained on language modeling (LM) tasks are loaded or fine-tuning on a different dataset on the same LM task. |
finetune_GPT111M folder:
run_<data>_######.log contains the command line output generated during execution.
Evaluating the trained models and visualize using TensorBoard
For all models, you can use therun.py script found in the Cerebras Model Zoo for evaluation purposes (i.e., only forward pass). GPT style models use the data specified in eval_input.data_dir, which you had set up in step 2. The run.py script provides three types of evaluation with the --mode flag:
| Flag | Description |
|---|---|
eval | Evaluates a specific checkpoint. The latest checkpoint will be used if you don’t provide the --checkpoint_path flag |
eval_all | Evaluates all the checkpoints inside a model directory once the model has been trained |
train | Evaluates the model periodically during the training process. |
train_and_eval | Evaluates a model at a fixed frequency during training. This is convenient for identifying issues early in long training runs |
NoteThe
train and eval modes require different fabric programming in the CS-X system. Therefore, using train_and_eval mode in the Cerebras Wafer-Scale cluster results in additional overheads any time training is stopped to perform evaluation. When possible, we recommend using the eval_all mode instead.eval_all mode. To learn more about the different types of evaluation, visit eval.
NoteWhen evaluating a model with
run.py, the latest saved checkpoint will be used by default. If no checkpoint exists, then weights will be initialized as stated in the YAML file, and the model will be evaluated using these weights. If you want to evaluate a previously trained model, make sure that the checkpoints are available in the model_dir or provide the --checkpoint_path flag.train_wsc screen.
venv_cerebras_pt.
Evaluating the model trained from scratch
To evaluate the model trained from scratch, run therun.py script associated with the GPT-3 models in the Cerebras Model Zoo. This is the same run.py script used during training from scratch, but you will change the following flag:
--mode eval_all | Specifies the evaluation mode |
train_from_scratch_GPT111M folder, including an eval folder that contains the metrics logged during the model evaluation and a new run_<data>_######.log with the command line output during model evaluation. Here is an example of the command line output.
Evaluating the model fine-tuned from Cerebras-GPT
To evaluate the fine-tuned model, run therun.py script associated with the GPT-3 models in the Cerebras Model Zoo. This is the same run.py script used during evaluating model trained from scratch, but you will change the following flag:
--model_dir finetune_GPT111M | Specifies the model directory that contains the checkpoints from the fine-tuned model |
finetune_GPT_111M folder, including an eval folder that contains the metrics logged during the model evaluation and a new run_<data>_######.log with the command line output during model evaluation. Here is an example of the command line output.
Visualizing the results using TensorBoard#
Launch TensorBoard to visualize the metrics recorded during training and evaluation of the models. Make sure that the Cerebras virtual environment is active.NoteWhen hosting multiple TensorBoard instances concurrently, you may see the error
Port 6006 is in use by another program. Either identify and stop that program, or start the server with a different port.. As a work around, you can launch TensorBoard by specifying a different port with the --port flag.train_from_scratch_GPT111M) and fine-tuning from Cerebras-GPT checkpoint (finetune_GPT111M)#