Component to finetune Hugging Face pretrained models for chat completion task. The component supports optimizations such as LoRA, Deepspeed and ONNXRuntime for performance enhancement. See docs to learn more.

Version: 0.0.65

View in Studio: https://ml.azure.com/registries/azureml/components/chat_completion_finetune/version/0.0.65

Lora parameters

LoRA reduces the number of trainable parameters by learning pairs of rank-decompostion matrices while freezing the original weights. This vastly reduces the storage requirement for large language models adapted to specific tasks and enables efficient task-switching during deployment all without introducing inference latency. LoRA also outperforms several other adaptation methods including adapter, prefix-tuning, and fine-tuning. Currently, LoRA is supported for gpt2, bert, roberta, deberta, distilbert, t5, bart, mbart and camembert model families

Name	Description	Type	Default	Optional	Enum
apply_lora	If "true" enables lora.	string	false	True	['true', 'false']
merge_lora_weights	If "true", the lora weights are merged with the base Hugging Face model weights before saving.	string	true	True	['true', 'false']
lora_alpha	alpha attention parameter for lora.	integer	128	True
lora_r	lora dimension	integer	8	True
lora_dropout	lora dropout value	number	0.0	True

Training parameters

Name	Description	Type	Default	Optional	Enum
num_train_epochs	Number of epochs to run for finetune.	integer	1	True
max_steps	If set to a positive number, the total number of training steps to perform. Overrides 'epochs'. In case of using a finite iterable dataset the training may stop before reaching the set number of steps when all data is exhausted.	integer	-1	True
per_device_train_batch_size	Per gpu batch size used for training. The effective training batch size is per_device_train_batch_size * num_gpus * num_nodes.	integer	1	True
per_device_eval_batch_size	Per gpu batch size used for validation. The default value is 1. The effective validation batch size is per_device_eval_batch_size * num_gpus * num_nodes.	integer	1	True
auto_find_batch_size	If set to "true" and if the provided 'per_device_train_batch_size' goes into Out Of Memory (OOM) auto_find_batch_size will find the correct batch size by iteratively reducing batch size by a factor of 2 till the OOM is fixed	string	false	True	['true', 'false']
optim	Optimizer to be used while training	string	adamw_hf	True	['adamw_hf', 'adamw_torch', 'adafactor']
learning_rate	Start learning rate used for training.	number	2e-05	True
warmup_steps	Number of steps for the learning rate scheduler warmup phase.	integer	0	True
weight_decay	Weight decay to apply (if not zero) to all layers except all bias and LayerNorm weights in AdamW optimizer	number	0.0	True
adam_beta1	beta1 hyperparameter for the AdamW optimizer	number	0.9	True
adam_beta2	beta2 hyperparameter for the AdamW optimizer	number	0.999	True
adam_epsilon	epsilon hyperparameter for the AdamW optimizer	number	1e-08	True
gradient_accumulation_steps	Number of updates steps to accumulate the gradients for, before performing a backward/update pass	integer	1	True
eval_accumulation_steps	Number of predictions steps to accumulate before moving the tensors to the CPU, will be passed as None if set to -1	integer	-1	True
lr_scheduler_type	learning rate scheduler to use.	string	linear	True	['linear', 'cosine', 'cosine_with_restarts', 'polynomial', 'constant', 'constant_with_warmup']
precision	Apply mixed precision training. This can reduce memory footprint by performing operations in half-precision.	string	32	True	['32', '16']
seed	Random seed that will be set at the beginning of training	integer	42	True
enable_full_determinism	Ensure reproducible behavior during distributed training. Check this link https://pytorch.org/docs/stable/notes/randomness.html for more details.	string	false	True	['true', 'false']
dataloader_num_workers	Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process.	integer	0	True
ignore_mismatched_sizes	Not setting this flag will raise an error if some of the weights from the checkpoint do not have the same size as the weights of the model.	string	false	True	['true', 'false']
max_grad_norm	Maximum gradient norm (for gradient clipping)	number	1.0	True
evaluation_strategy	The evaluation strategy to adopt during training. If set to "steps", either the evaluation_steps_interval or eval_steps needs to be specified, which helps to determine the step at which the model evaluation needs to be computed else evaluation happens at end of each epoch.	string	epoch	True	['epoch', 'steps']
evaluation_steps_interval	The evaluation steps in fraction of an epoch steps to adopt during training. Overwrites eval_steps if not 0.	number	0.0	True
eval_steps	Number of update steps between two evals if evaluation_strategy='steps'	integer	500	True
logging_strategy	The logging strategy to adopt during training. If set to "steps", the logging_steps will decide the frequency of logging else logging happens at the end of epoch.	string	steps	True	['epoch', 'steps']
logging_steps	Number of update steps between two logs if logging_strategy='steps'	integer	10	True
metric_for_best_model	metric to use to compare two different model checkpoints	string	loss	True	['loss', 'f1', 'exact']
resume_from_checkpoint	If set to "true", resumes the training from last saved checkpoint. Along with loading the saved weights, saved optimizer, scheduler and random states will be loaded if exist. The default value is "false"	string	false	True	['true', 'false']
save_total_limit	If a positive value is passed, it will limit the total number of checkpoints saved. The value of -1 saves all the checkpoints, otherwise if the number of checkpoints exceed the save_total_limit, the older checkpoints gets deleted.	integer	-1	True

Early Stopping Parameters

Name	Description	Type	Default	Optional	Enum
apply_early_stopping	If set to "true", early stopping is enabled.	string	false	True	['true', 'false']
early_stopping_patience	Stop training when the metric specified through metric_for_best_model worsens for early_stopping_patience evaluation calls.This value is only valid if apply_early_stopping is set to true.	integer	1	True
early_stopping_threshold	Denotes how much the specified metric must improve to satisfy early stopping conditions. This value is only valid if apply_early_stopping is set to true.	number	0.0	True

Deepspeed Parameters Deepspeed config is a JSON file that can be used to configure optimizer, scheduler, batch size and other training related parameters. A default deepspeed config is used when apply_deepspeed is set to true. Alternatively, you can pass your custom deepspeed config. Please follow the deepspeed docs to create the custom config. Please note that to enable deepspeed, apply_deepspeed must be set to true, only passing the deepspeed input will not suffice

Name	Description	Type	Default	Optional	Enum
apply_deepspeed	If set to true, will enable deepspeed for training	string	false	True	['true', 'false']
deepspeed	Deepspeed config to be used for finetuning	uri_file		True
deepspeed_stage	This parameter configures which DEFAULT deepspeed config to be used - stage2 or stage3. The default choice is stage2. Note that, this parameter is ONLY applicable when user doesn't pass any config information via deepspeed port.	string	2	True	['2', '3']

ORT Parameters ONNX Runtime is a cross-platform machine-learning model accelerator, with a flexible interface to integrate hardware-specific libraries.

Name	Description	Type	Default	Optional	Enum
apply_ort	If set to true, will use the ONNXRunTime training	string	false	True	['true', 'false']

Data and Model inputs

Name	Description	Type	Default	Optional	Enum
preprocess_output	output folder of preprocess component containing encoded train, valid and test data. The tokenizer is also saved as part of preprocess output	uri_folder		False
model_selector_output	output folder of model import component containing model artifacts and a metadata file.	uri_folder		False

Name	Description	Type
pytorch_model_folder	output folder containing best model as defined by metric_for_best_model. Along with the best model, output folder contains checkpoints saved after every evaluation which is defined by the evaluation_strategy. Each checkpoint contains the model weight(s), config, tokenizer, optimzer, scheduler and random number states.	uri_folder

azureml://registries/azureml/environments/acft-hf-nlp-gpu/versions/81