dnrops
/
transformers
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Fix sagemaker DP/MP (#23681) 

* Check for use_sagemaker_dp

* Add a check for is_sagemaker_mp when setting _n_gpu again. Should be last broken thing

* Try explicit check?

* Quality
This commit is contained in:
Zachary Mueller 2023-05-24 15:51:09 -04:00 committed by GitHub
parent 89159651ba
commit 75bbf20bce
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
2 changed files with 9 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
src/transformers/trainer.py
View File

@ -3398,7 +3398,9 @@ class Trainer:
tensors = nested_xla_mesh_reduce(tensors, name)
elif is_sagemaker_mp_enabled():
tensors = smp_gather(tensors)
elif self.args.parallel_mode == ParallelMode.DISTRIBUTED:
elif (self.args.distributed_state is not None and self.args.distributed_state.distributed_type != "NO") or (
self.args.distributed_state is None and self.local_rank != -1
):
tensors = distributed_concat(tensors)
return tensors

8
src/transformers/training_args.py
View File

@ -1629,6 +1629,9 @@ class TrainingArguments:
device = torch.device("cuda", local_rank)
self._n_gpu = 1
torch.cuda.set_device(device)
elif is_sagemaker_dp_enabled():
self.distributed_state = PartialState(_use_sagemaker_dp=True)
self._n_gpu = 1
elif self.deepspeed:
# Need to do similar for Accelerator init
os.environ["ACCELERATE_USE_DEEPSPEED"] = "true"
@ -1653,8 +1656,9 @@ class TrainingArguments:
if is_torch_tpu_available():
device = self.distributed_state.device
self._n_gpu = 0
elif is_sagemaker_dp_enabled():
self._n_gpu = 1
elif is_sagemaker_dp_enabled() or is_sagemaker_mp_enabled():
# Already set _n_gpu
pass
elif self.distributed_state.distributed_type == DistributedType.NO:
if self.use_mps_device:
if not torch.backends.mps.is_available():