--task translation_dat_task # Task for DA-Transformer
--upsample-base predict # Possible values are: ["predict", "source", "source_old"].
# If set to "predict", the DAG size will be determined by the golden target length during training and the predicted length during inference.
# Note that --length-loss-factor must be greater than 0 during training.
# If set to "source", the DAG size will be determined by the source length during both training and inference. You can disable the length
# predictor during training by setting --length-loss-factor to 0.
# If set to "source_old", the DAG size is determined similarly to "source" but several token longer. This option is only used for
# compatibility with the upsampling method in version 1.0.
--upsample-scale 4~8 # Specifies the upsample scale for the decoder input length in training.
# For instance, "4~8" indicates that the upsampling scale will be uniformly sampled from the range [4, 8];
# "4" indicates fixed upsampling scale.
--seg-tokens 32 # This parameter specifies the number of special tokens that will be used for segment id.
# If you are using pre-trained checkpoints, please set this value to 32.
--filter-max-length 512:128 # Filters samples that exceed the maximum lengths. For example, "128:256" indicates a maximum source length of 128 and a maximum target length of 256.
# The default value of None filters according to max-source-positions and max-target-positions.
--filter-ratio 8 # Filters out samples that do not satisfy the specified len(target)/len(source) ratio constraints.
# For example, if the ratio is set to "8", samples where len(target)/len(source) > 8 or len(target)/len(source) < 1/8 will be removed.
# If set to "0.5~2", samples where len(target)/len(source) < 0.5 or len(target)/len(source) > 2 will be removed.
# Default: None (disabled).
--do-not-load-task-args # Do not load task arguments from the restored checkpoints.--arch glat_decomposed_link_base # The Model Architecture. You can use "ls_glat_decomposed_link_base" to enable LightSeq's Transformer
--links-feature feature:position # Specifies the features used to predict transitions, separated by a colon.
# For example, "feature:position" represents the concatenation of decoder features and learnable positional embeddings.
--segment-embedding # Adds an additional embedding represented segment id for the decoder input.
--max-source-positions 128 # Max length of encoder.
--max-target-positions 1024 # Max length of decoder. If the length of a sample exceeds this limit after up-sampling, the sample will be discarded.
--load-pretrained-model ${data_dir} # Path to a file containing a pre-trained model. It also support a checkpoint file and will automatically convert between lightseq and fairseq architecture.This configs used in the validation. See more configs related to decoding here.
--decode-strategy lookahead # Decoding Strategy. Possible values: greedy, lookahead, beamsearch.
--decode-upsample-scale 8 # Upsampling scale to determine the DAG size during inference.
# If --upsample-scale used in training is a fixed number, this parameter should be the same value.
# If --upsample-scale used in training is a range, this parameter can be the average of the range and optionally tuned after the training."--criterion nat_dag_loss # The Criterion for DA-Transformer.
--length-loss-factor 0.1 # Weights on the length prediction loss. Required if --upsample_base "predict" is set.
--max-transition-length 99999 # Specifies the maximum transition distance. A value of -1 indicates no limit, but this cannot be used with CUDA custom operations.
# To use CUDA operations with no limit, specify a very large number such as 99999.
--glat-p 0.5:0.1@200k # Set the glancing probability and its annealing schedule. For example, '0.5:0.1@200k' indicates annealing probability from 0.5 to 0.1 in 200k steps.
--glance-strategy number-random # Set the glancing strategy. Possible values: "number-random" or "None" or "CMLM"
--use-pretrain-loss # If true, use the pre-training loss, i.e. the position of segment id will be fixed.
--no-force-emit # If true, the position of glanced tokens in the second forward pass will not be fixed.
--torch-dag-loss # Use torch native implementation for logsoftmax-gather. It may be slower and consume more GPU memory.
--torch-dag-best-alignment # Use torch native implementation for dag-best-alignment. It may be slower and consume more GPU memory.
--torch-dag-logsoftmax-gather # Use torch native implementation for dag-loss. It may be slower and consume more GPU memory.--optimizer adam # The optimizer. You can use "ls_adam" instead to enable LightSeq's Optimizer--max-tokens 4096 # Specifies the maximum number of tokens (either source or target) allowed in a single batch during training.
# This number does not include any padding tokens.
--max-tokens-valid 4096 # Specifies the maximum number of tokens (either source or target) allowed in a single batch during validation.
# This number does not include any padding tokens.
--update-freq 2 # Specifies the number of steps of gradient accumulation before updating the model.
# The actual batch size is: GPU number * max_tokens * update_freq.
--max-tokens-after-upsample # If enabled, the maximum number of tokens (--max-tokens) considered during generation
# will take into account the upsampling ratio. In other words, the length of the generated sequence will be capped at
# max(source_length, decoder_length * upsample_scale). Default: False.
--batch-split-by-src 32767 # If this value is greater than 0, it splits a batch into multiple smaller batches.
# The split is based on the number of source tokens in each batch (considering padding tokens),
# ensuring that no batch has more source tokens than the specified value.
# This is different from --update-freq because it works on each GPU separately. It's useful when out-of-memory (OOM) errors occur rarely
# and you do not want to set a smaller batch size.
--max-encoder-batch-tokens 20000 # Specifies the maximum number of tokens for the encoder input to avoid running out of memory. The default value of None indicates no limit.
--max-decoder-batch-tokens 20000 # Specifies the maximum number of tokens for the decoder input to avoid running out of memory. The default value of None indicates no limit.--eval-bleu # Evaluate BLEU scores during validation
--eval-bleu-detok space # Detokenizer used in BLEU evaluation
--eval-bleu-remove-bpe # Whether remove bpe in BLEU evaluation
--eval-bleu-print-samples # Print several samples in BLEU evaluation
--eval-bleu-order 4 # The order of n-gram in BLEU evaluation