[Enhance] Update fsdp vit-huge and vit-large config (#1675)
* Update fsdp vit-huge and vit-large config * Update fsdp vit-huge and vit-large config * renamepull/1689/head
parent
658db80089
commit
8afad77a35
|
@ -18,7 +18,7 @@ strategy = dict(
|
|||
),
|
||||
inputs_to_half=['inputs'],
|
||||
zero_optimization=dict(
|
||||
stage=3,
|
||||
stage=1,
|
||||
allgather_partitions=True,
|
||||
reduce_scatter=True,
|
||||
allgather_bucket_size=50000000,
|
|
@ -0,0 +1,13 @@
|
|||
_base_ = ['./vit-huge-p14_8xb128-coslr-50e_in1k.py']
|
||||
|
||||
strategy = dict(
|
||||
type='FSDPStrategy',
|
||||
model_wrapper=dict(
|
||||
auto_wrap_policy=dict(
|
||||
type='torch.distributed.fsdp.wrap.size_based_auto_wrap_policy',
|
||||
min_num_params=1e7)))
|
||||
|
||||
optim_wrapper = dict(type='AmpOptimWrapper')
|
||||
|
||||
# runner which supports strategies
|
||||
runner_type = 'FlexibleRunner'
|
|
@ -18,7 +18,7 @@ strategy = dict(
|
|||
),
|
||||
inputs_to_half=['inputs'],
|
||||
zero_optimization=dict(
|
||||
stage=3,
|
||||
stage=1,
|
||||
allgather_partitions=True,
|
||||
reduce_scatter=True,
|
||||
allgather_bucket_size=50000000,
|
|
@ -0,0 +1,13 @@
|
|||
_base_ = ['./vit-large-p16_8xb128-coslr-50e_in1k.py']
|
||||
|
||||
strategy = dict(
|
||||
type='FSDPStrategy',
|
||||
model_wrapper=dict(
|
||||
auto_wrap_policy=dict(
|
||||
type='torch.distributed.fsdp.wrap.size_based_auto_wrap_policy',
|
||||
min_num_params=1e7)))
|
||||
|
||||
optim_wrapper = dict(type='AmpOptimWrapper')
|
||||
|
||||
# runner which supports strategies
|
||||
runner_type = 'FlexibleRunner'
|
Loading…
Reference in New Issue