Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[ZeRO Infinity] Allow Init to take a dict for the deepspeed config #983

Merged
merged 6 commits into from
Apr 20, 2021
Merged

[ZeRO Infinity] Allow Init to take a dict for the deepspeed config #983

Changes from 4 commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
bc98d26
Add check to see if json file is already loaded
Apr 20, 2021
4b31cc2
Update doc
Apr 20, 2021
077d4e9
Address review
Apr 20, 2021
e6a6f4e
Remove doc comment
Apr 20, 2021
ec5ead5
Merge branch 'master' into fix/config
tjruwase Apr 20, 2021
bc340e0
Merge branch 'master' into fix/config
tjruwase Apr 20, 2021
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
11 changes: 7 additions & 4 deletions deepspeed/runtime/zero/partition_parameters.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -271,6 +271,7 @@ def __init__(self,
remote_device=None,
pin_memory=False,
deepspeed_config=None,
param_dict=None,
enabled=True):
"""A context to enable massive model construction for training with
ZeRO-3. Models are automatically partitioned (or, sharded) across the
Expand All @@ -293,6 +294,8 @@ def __init__(self,
``"cpu"``. Defaults to ``False``.
deepspeed_config (``json file``, optional): If provided, provides configuration
for swapping fp16 params to NVMe.
param_dict (dict, optional): Instead of requiring a deepspeed_config you can pass your deepspeed config
as a dictionary instead for swapping fp16 params to NVMe.
enabled (bool, optional): If ``False``, this context has no
effect. Defaults to ``True``.

Expand Down Expand Up @@ -382,7 +385,7 @@ def get_model():
#It is the device where parameters are fully instantiated using allgather
self.local_device = torch.device('cuda:{}'.format(os.environ["LOCAL_RANK"]))

self._validate_remote_device(remote_device, deepspeed_config)
self._validate_remote_device(remote_device, deepspeed_config, param_dict)

#Remote device is the device where parameter partiitons are stored
#It can be same as local_device or it could be CPU or NVMe.
Expand All @@ -392,7 +395,7 @@ def get_model():

# Enable fp16 param swapping to NVMe
if self.remote_device == OFFLOAD_NVME_DEVICE:
_ds_config = DeepSpeedConfig(deepspeed_config)
_ds_config = DeepSpeedConfig(deepspeed_config, param_dict=param_dict)
self.param_swapper = AsyncPartitionedParameterSwapper(_ds_config)
else:
self.param_swapper = None
Expand All @@ -406,9 +409,9 @@ def get_model():
self._convert_to_deepspeed_param(param)
param.partition()

def _validate_remote_device(self, remote_device, ds_config):
def _validate_remote_device(self, remote_device, ds_config, param_dict):
if ds_config is not None:
_ds_config = DeepSpeedConfig(ds_config)
_ds_config = DeepSpeedConfig(ds_config, param_dict=param_dict)
if remote_device in [None, OFFLOAD_CPU_DEVICE]:
if _ds_config.zero_config.offload_param is not None:
offload_param_device = _ds_config.zero_config.offload_param[
Expand Down