muellerzr's picture
muellerzr HF staff
updating fsdp docs as per resolution to issue https://github.com/huggingface/accelerate/issues/1054 (#3)
52eaca3
raw
history blame
2.79 kB
##
Below is an example yaml for BF16 mixed-precision training using PyTorch Fully Sharded Data Parallism (FSDP) with CPU offloading on 8 GPUs.
<pre>
compute_environment: LOCAL_MACHINE
deepspeed_config: {}
+distributed_type: FSDP
downcast_bf16: 'no'
dynamo_backend: 'NO'
+fsdp_config:
+ fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+ fsdp_backward_prefetch_policy: BACKWARD_PRE
+ fsdp_offload_params: true
+ fsdp_sharding_strategy: 1
+ fsdp_state_dict_type: FULL_STATE_DICT
+ fsdp_transformer_layer_cls_to_wrap: T5Block
machine_rank: 0
main_training_function: main
megatron_lm_config: {}
mixed_precision: bf16
num_machines: 1
+num_processes: 8
rdzv_backend: static
same_network: true
use_cpu: false
</pre>
##
<pre>
from accelerate import Accelerator
def main():
accelerator = Accelerator()
- model, optimizer, dataloader, scheduler = accelerator.prepare(
- model, optimizer, dataloader, scheduler
- )
+ model = accelerator.prepare(model)
+ # Optimizer can be any PyTorch optimizer class
+ optimizer = torch.optim.AdamW(params=model.parameters(), lr=lr)
+ optimizer, dataloader, scheduler = accelerator.prepare(
+ optimizer, dataloader, scheduler
+ )
...
accelerator.unwrap_model(model).save_pretrained(
args.output_dir,
is_main_process=accelerator.is_main_process,
save_function=accelerator.save,
+ state_dict=accelerator.get_state_dict(model)
)
...
</pre>
##
If the YAML was generated through the `accelerate config` command:
```
accelerate launch {script_name.py} {--arg1} {--arg2} ...
```
If the YAML is saved to a `~/config.yaml` file:
```
accelerate launch --config_file ~/config.yaml {script_name.py} {--arg1} {--arg2} ...
```
Or you can use `accelerate launch` with right configuration parameters and have no `config.yaml` file:
```
accelerate launch \
--use_fsdp \
--num_processes=8 \
--mixed_precision=bf16 \
--fsdp_sharding_strategy=1 \
--fsdp_auto_wrap_policy=TRANSFORMER_BASED_WRAP \
--fsdp_transformer_layer_cls_to_wrap=T5Block \
--fsdp_offload_params=true \
{script_name.py} {--arg1} {--arg2} ...
```
##
For PyTorch FDSP, you need to prepare the model first **before** preparing the optimizer since FSDP will shard parameters in-place and this will break any previously initialized optimizers.
For transformer models, please use `TRANSFORMER_BASED_WRAP` auto wrap policy as shown in the config above.
##
To learn more checkout the related documentation:
- <a href="https://huggingface.co/docs/accelerate/usage_guides/fsdp" target="_blank">How to use Fully Sharded Data Parallelism</a>
- <a href="https://huggingface.co/blog/pytorch-fsdp" target="_blank">Accelerate Large Model Training using PyTorch Fully Sharded Data Parallel</a>