SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

Replace accelerator.use_fp16 in examples (#33513)

Browse Source 
* Replace `accelerator.use_fp16` in examples

* pad_to_multiple_of=16 for fp8
This commit is contained in:
hlky
2024-09-17 03:13:06 +01:00
committed by GitHub
parent ba1f1dc132
commit 9f196ef2e0
10 changed files with 79 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
examples/pytorch/multiple-choice/run_swag_no_trainer.py
View File

@@ -473,9 +473,14 @@ def main():
# Otherwise, `DataCollatorWithPadding` will apply dynamic padding for us (by padding to the maximum length of
# the samples passed). When using mixed precision, we add `pad_to_multiple_of=8` to pad all tensors to multiple
# of 8s, which will enable the use of Tensor Cores on NVIDIA hardware with compute capability >= 7.5 (Volta).
data_collator = DataCollatorForMultipleChoice(
tokenizer, pad_to_multiple_of=(8 if accelerator.use_fp16 else None)
)
# For fp8, we pad to multiple of 16.
if accelerator.mixed_precision == "fp8":
pad_to_multiple_of = 16
elif accelerator.mixed_precision != "no":
pad_to_multiple_of = 8
else:
pad_to_multiple_of = None
data_collator = DataCollatorForMultipleChoice(tokenizer, pad_to_multiple_of=pad_to_multiple_of)
train_dataloader = DataLoader(
train_dataset, shuffle=True, collate_fn=data_collator, batch_size=args.per_device_train_batch_size

9
examples/pytorch/question-answering/run_qa_beam_search_no_trainer.py
View File

@@ -670,7 +670,14 @@ def main():
# Otherwise, `DataCollatorWithPadding` will apply dynamic padding for us (by padding to the maximum length of
# the samples passed). When using mixed precision, we add `pad_to_multiple_of=8` to pad all tensors to multiple
# of 8s, which will enable the use of Tensor Cores on NVIDIA hardware with compute capability >= 7.5 (Volta).
data_collator = DataCollatorWithPadding(tokenizer, pad_to_multiple_of=(8 if accelerator.use_fp16 else None))
# For fp8, we pad to multiple of 16.
if accelerator.mixed_precision == "fp8":
pad_to_multiple_of = 16
elif accelerator.mixed_precision != "no":
pad_to_multiple_of = 8
else:
pad_to_multiple_of = None
data_collator = DataCollatorWithPadding(tokenizer, pad_to_multiple_of=pad_to_multiple_of)
train_dataloader = DataLoader(
train_dataset, shuffle=True, collate_fn=data_collator, batch_size=args.per_device_train_batch_size

9
examples/pytorch/question-answering/run_qa_no_trainer.py
View File

@@ -685,7 +685,14 @@ def main():
# Otherwise, `DataCollatorWithPadding` will apply dynamic padding for us (by padding to the maximum length of
# the samples passed). When using mixed precision, we add `pad_to_multiple_of=8` to pad all tensors to multiple
# of 8s, which will enable the use of Tensor Cores on NVIDIA hardware with compute capability >= 7.5 (Volta).
data_collator = DataCollatorWithPadding(tokenizer, pad_to_multiple_of=(8 if accelerator.use_fp16 else None))
# For fp8, we pad to multiple of 16.
if accelerator.mixed_precision == "fp8":
pad_to_multiple_of = 16
elif accelerator.mixed_precision != "no":
pad_to_multiple_of = 8
else:
pad_to_multiple_of = None
data_collator = DataCollatorWithPadding(tokenizer, pad_to_multiple_of=pad_to_multiple_of)
train_dataloader = DataLoader(
train_dataset, shuffle=True, collate_fn=data_collator, batch_size=args.per_device_train_batch_size

8
examples/pytorch/summarization/run_summarization_no_trainer.py
View File

@@ -534,11 +534,17 @@ def main():
logger.info(f"Sample {index} of the training set: {train_dataset[index]}.")
label_pad_token_id = -100 if args.ignore_pad_token_for_loss else tokenizer.pad_token_id
if accelerator.mixed_precision == "fp8":
pad_to_multiple_of = 16
elif accelerator.mixed_precision != "no":
pad_to_multiple_of = 8
else:
pad_to_multiple_of = None
data_collator = DataCollatorForSeq2Seq(
tokenizer,
model=model,
label_pad_token_id=label_pad_token_id,
pad_to_multiple_of=8 if accelerator.use_fp16 else None,
pad_to_multiple_of=pad_to_multiple_of,
)
def postprocess_text(preds, labels):

9
examples/pytorch/text-classification/run_glue_no_trainer.py
View File

@@ -426,7 +426,14 @@ def main():
# Otherwise, `DataCollatorWithPadding` will apply dynamic padding for us (by padding to the maximum length of
# the samples passed). When using mixed precision, we add `pad_to_multiple_of=8` to pad all tensors to multiple
# of 8s, which will enable the use of Tensor Cores on NVIDIA hardware with compute capability >= 7.5 (Volta).
data_collator = DataCollatorWithPadding(tokenizer, pad_to_multiple_of=(8 if accelerator.use_fp16 else None))
# For fp8, we pad to multiple of 16.
if accelerator.mixed_precision == "fp8":
pad_to_multiple_of = 16
elif accelerator.mixed_precision != "no":
pad_to_multiple_of = 8
else:
pad_to_multiple_of = None
data_collator = DataCollatorWithPadding(tokenizer, pad_to_multiple_of=pad_to_multiple_of)
train_dataloader = DataLoader(
train_dataset, shuffle=True, collate_fn=data_collator, batch_size=args.per_device_train_batch_size

11
examples/pytorch/token-classification/run_ner_no_trainer.py
View File

@@ -541,9 +541,14 @@ def main():
# Otherwise, `DataCollatorForTokenClassification` will apply dynamic padding for us (by padding to the maximum length of
# the samples passed). When using mixed precision, we add `pad_to_multiple_of=8` to pad all tensors to multiple
# of 8s, which will enable the use of Tensor Cores on NVIDIA hardware with compute capability >= 7.5 (Volta).
data_collator = DataCollatorForTokenClassification(
tokenizer, pad_to_multiple_of=(8 if accelerator.use_fp16 else None)
)
# For fp8, we pad to multiple of 16.
if accelerator.mixed_precision == "fp8":
pad_to_multiple_of = 16
elif accelerator.mixed_precision != "no":
pad_to_multiple_of = 8
else:
pad_to_multiple_of = None
data_collator = DataCollatorForTokenClassification(tokenizer, pad_to_multiple_of=pad_to_multiple_of)
train_dataloader = DataLoader(
train_dataset, shuffle=True, collate_fn=data_collator, batch_size=args.per_device_train_batch_size

9
examples/pytorch/translation/run_translation_no_trainer.py
View File

@@ -517,11 +517,18 @@ def main():
# Otherwise, `DataCollatorWithPadding` will apply dynamic padding for us (by padding to the maximum length of
# the samples passed). When using mixed precision, we add `pad_to_multiple_of=8` to pad all tensors to multiple
# of 8s, which will enable the use of Tensor Cores on NVIDIA hardware with compute capability >= 7.5 (Volta).
# For fp8, we pad to multiple of 16.
if accelerator.mixed_precision == "fp8":
pad_to_multiple_of = 16
elif accelerator.mixed_precision != "no":
pad_to_multiple_of = 8
else:
pad_to_multiple_of = None
data_collator = DataCollatorForSeq2Seq(
tokenizer,
model=model,
label_pad_token_id=label_pad_token_id,
pad_to_multiple_of=8 if accelerator.use_fp16 else None,
pad_to_multiple_of=pad_to_multiple_of,
)
train_dataloader = DataLoader(

11
examples/research_projects/luke/run_luke_ner_no_trainer.py
View File

@@ -542,9 +542,14 @@ def main():
# Otherwise, `DataCollatorForTokenClassification` will apply dynamic padding for us (by padding to the maximum length of
# the samples passed). When using mixed precision, we add `pad_to_multiple_of=8` to pad all tensors to multiple
# of 8s, which will enable the use of Tensor Cores on NVIDIA hardware with compute capability >= 7.5 (Volta).
data_collator = DataCollatorForLukeTokenClassification(
tokenizer, pad_to_multiple_of=(8 if accelerator.use_fp16 else None)
)
# For fp8, we pad to multiple of 16.
if accelerator.mixed_precision == "fp8":
pad_to_multiple_of = 16
elif accelerator.mixed_precision != "no":
pad_to_multiple_of = 8
else:
pad_to_multiple_of = None
data_collator = DataCollatorForLukeTokenClassification(tokenizer, pad_to_multiple_of=pad_to_multiple_of)
train_dataloader = DataLoader(
train_dataset, shuffle=True, collate_fn=data_collator, batch_size=args.per_device_train_batch_size

9
examples/research_projects/self-training-text-classification/finetuning.py
View File

@@ -704,7 +704,14 @@ def finetune(accelerator, model_name_or_path, train_file, output_dir, **kwargs):
# precision, we add `pad_to_multiple_of=8` to pad all tensors to multiple of
# 8s, which will enable the use of Tensor Cores on NVIDIA hardware with
# compute capability >= 7.5 (Volta).
data_collator = DataCollatorWithPadding(tokenizer, pad_to_multiple_of=(8 if accelerator.use_fp16 else None))
# For fp8, we pad to multiple of 16.
if accelerator.mixed_precision == "fp8":
pad_to_multiple_of = 16
elif accelerator.mixed_precision != "no":
pad_to_multiple_of = 8
else:
pad_to_multiple_of = None
data_collator = DataCollatorWithPadding(tokenizer, pad_to_multiple_of=pad_to_multiple_of)
train_dataloader = DataLoader(
train_dataset,