src/fairseq/fairseq/modules/transformer_layer.py (2 lines):
	- line 128: # TODO: to formally solve this problem, we need to change fairseq's
	- line 203: # TODO  remove this once we update apex with the fix


src/fairseq/fairseq/tasks/multilingual_translation.py (2 lines):
	- line 276: # TODO make summing of the sample sizes configurable
	- line 293: # TODO make summing of the sample sizes configurable


src/fairseq/fairseq/iterative_refinement_generator.py (2 lines):
	- line 110: # TODO: iterative refinement generator does not support ensemble for now.
	- line 127: # TODO: better encoder inputs?


src/fairseq/fairseq/tasks/translation.py (1 line):
	- line 197: help='split the input based on this index to get general text and graph text') #TODO: fix this


src/fairseq/fairseq/criterions/legacy_masked_lm.py (1 line):
	- line 96: # TODO: Remove this after refactor of BERTModel


src/fairseq/fairseq/data/round_robin_zip_datasets.py (1 line):
	- line 75: # TODO make it configurable whether to use max() or sum() here


src/prepare_data.py (1 line):
	- line 363: #TODO: right now the joining of docs is bad as src tokenize gets wrong


src/fairseq/fairseq/modules/longformer_multihead_attention.py (1 line):
	- line 304: # TODO: remove the redundant computation


src/fairseq/fairseq/models/nat/insertion_transformer.py (1 line):
	- line 179: # TODO: decoding for InsertionTransformer


src/fairseq/fairseq/data/legacy/block_pair_dataset.py (1 line):
	- line 219: TODO: ids in skip_ids should be consecutive, we can extend it to more generic version later


src/fairseq/fairseq/modules/positional_embedding.py (1 line):
	- line 20: # TODO: The right place for this offset would be inside


src/fairseq/fairseq/tasks/semisupervised_translation.py (1 line):
	- line 343: # TODO make summing of the sample sizes configurable


src/fairseq/fairseq/trainer.py (1 line):
	- line 847: # TODO: make 1e-6 a configurable value


src/fairseq/fairseq/models/nat/nonautoregressive_transformer.py (1 line):
	- line 362: # TODO: implementing length-beam


src/fairseq/fairseq/models/lstm.py (1 line):
	- line 397: # TODO make bias configurable


src/fairseq/fairseq/data/legacy/masked_lm_dataset.py (1 line):
	- line 208: # TODO: Can we add deteminism without this constraint?


src/fairseq/fairseq/utils.py (1 line):
	- line 167: # TODO: Very rare cases where the replacement is '<eos>' should be handled gracefully


src/fairseq/fairseq/data/transform_eos_lang_pair_dataset.py (1 line):
	- line 65: # TODO: support different padding direction on target side


src/fairseq/fairseq/data/noising.py (1 line):
	- line 113: # TODO: speed up the following loop


src/graph_construction.py (1 line):
	- line 199: for n in graph.neighbors(s): # TODO: can make better with order or neighbors based on their weight