Llama2 - Optimisations
These configuration options offer various techniques and optimisations to improve the training process
Gradient checkpointing and Flash Attention focus on memory efficiency and computational speed, while early stopping and resuming from checkpoints are useful for preventing overfitting and managing the training workflow.
Distributed training with local_rank enables parallel processing across multiple devices, and adjusting the logging frequency with logging_steps helps in monitoring the training progress.
We will use the generic configurations provided by Axolotl.
Below is an analysis of each of the configurations
gradient_checkpointing: true
Gradient checkpointing is a technique used to reduce the memory usage during training by trading off computation time.
Instead of storing all the intermediate activations for the backward pass, gradient checkpointing selectively stores a subset of activations and recomputes the others when needed.
Setting
gradient_checkpointingtotrueenables this feature, which can be particularly beneficial when training large models with limited memory.By enabling gradient checkpointing, you can potentially train larger models or use larger batch sizes without running out of memory.
However, it's important to note that gradient checkpointing introduces additional computational overhead, as the model needs to recompute the activations during the backward pass.
The decision to use gradient checkpointing depends on the available memory and the trade-off between memory usage and training speed.
gradient_checkpointing_kwargs:
This configuration allows you to specify additional keyword arguments for gradient checkpointing.
In the provided example,
use_reentrant: Trueis specified as a keyword argument.The
use_reentrantflag is related to the implementation of gradient checkpointing in PyTorch.When set to
True, it enables the use of reentrant autograd functions, which can provide additional memory savings during gradient checkpointing.However, the specific behavior and impact of this flag may depend on the PyTorch version and the model architecture.
early_stopping_patience:
Early stopping is a technique used to prevent overfitting and improve generalization performance.
It monitors a validation metric (e.g., validation loss or accuracy) during training and stops the training process if the metric does not improve for a specified number of iterations (patience).
The
early_stopping_patienceconfiguration allows you to set the number of iterations to wait before early stopping is triggered.For example, if
early_stopping_patienceis set to 3, training will stop if the validation metric does not improve for 3 consecutive iterations.Early stopping helps to avoid wasting computational resources on training iterations that do not lead to further improvements and can help prevent the model from overfitting to the training data.
resume_from_checkpoint:
This configuration allows you to resume training from a specific checkpoint.
By specifying a checkpoint directory or file path, you can load the model state, optimizer state, and other necessary information to continue training from where it left off.
Resuming from a checkpoint can be useful in various scenarios, such as when training is interrupted due to system failures, when you want to fine-tune a pre-trained model, or when you want to experiment with different hyperparameters starting from a previously trained model.
It saves time and resources by avoiding the need to start training from scratch.
local_rank:
The
local_rankconfiguration is related to distributed training, specifically when using techniques like Data Parallel or Distributed Data Parallel.In distributed training, multiple GPUs or machines are used to parallelise the training process and speed up computations.
The
local_rankrepresents the unique identifier of a process within a distributed training setup.It is typically used to determine the device placement and communication patterns among the processes.
When using distributed training frameworks like PyTorch's
DistributedDataParallel, thelocal_rankis automatically set by the framework.
logging_steps: 1
The
logging_stepsconfiguration determines the frequency at which training logs and metrics are recorded.In this case, setting
logging_stepsto 1 means that logs will be generated after every training step.Logging can include information such as the current training loss, learning rate, elapsed time, and other relevant metrics.
More frequent logging can be useful for monitoring the training progress and identifying any potential issues early on.
However, generating logs after every step can also introduce overhead and slow down the training process, especially for large datasets or long training runs.
The logging frequency should be adjusted based on the specific needs and the scale of the training task.
xformers_attention
The
xformers_attentionconfiguration is related to the use of the XFormers library, which provides optimized attention implementations for transformers.XFormers offers various attention mechanisms, such as memory-efficient attention, that can speed up training and reduce memory usage compared to the standard attention implementation in PyTorch.
Setting
xformers_attentionto a specific value (not provided in the given configuration) would enable the use of XFormers attention in the model.The specific attention mechanism and its parameters would depend on the value provided for
xformers_attention.Using XFormers attention can be beneficial for training large models or when dealing with long sequences, as it can provide computational and memory efficiency improvements.
flash_attention: true
Flash Attention is a highly optimised attention implementation that can significantly speed up the training of transformers.
It is designed to be memory-efficient and can handle large sequence lengths and batch sizes.
Setting
flash_attentiontotrueenables the use of Flash Attention in the model.Flash Attention can provide substantial performance improvements, especially for models with a large number of attention heads and long sequences.
It achieves this by using techniques like kernel fusion, memory optimization, and efficient parallelization.
Enabling Flash Attention can help reduce training time and allow for training larger models or using larger batch sizes.
Last updated
