MCPcopy Create free account
hub / github.com/EverMind-AI/MSA / get_sequence_parallel_preprocess

Function get_sequence_parallel_preprocess

src/utils/data_utils.py:237–246  ·  view source on GitHub ↗
(stage, tokenizer, cutoff_len=None, sequence_parallel_size=1, sequence_parallel_mode="ulysses")

Source from the content-addressed store, hash-verified

235 return examples
236
237def get_sequence_parallel_preprocess(stage, tokenizer, cutoff_len=None, sequence_parallel_size=1, sequence_parallel_mode="ulysses"):
238 if stage == "pad":
239 assert cutoff_len is not None
240 preprocess_func = partial(pad_sequence, cutoff_len=cutoff_len, tokenizer=tokenizer)
241 elif stage == "split":
242 preprocess_func = partial(sp_split, sequence_parallel_size=sequence_parallel_size, sequence_parallel_mode=sequence_parallel_mode)
243 else:
244 raise NotImplementedError(f"Unexpected stage in sequence_parallel_preprocess: {stage}")
245
246 return preprocess_func
247
248def _get_sequence_parallel_dataset(dataset, num_works, tokenizer=None, cutoff_len=10000,
249 sequence_parallel_size=1, sequence_parallel_mode="ulysses",

Callers 1

Calls

no outgoing calls

Tested by

no test coverage detected