LLaMA-Factory-Mirror/examples/accelerate/fsdp_config.yaml

26 lines
739 B
YAML
Raw Normal View History

2024-03-21 00:36:06 +08:00
compute_environment: LOCAL_MACHINE
debug: false
distributed_type: FSDP
downcast_bf16: 'no'
fsdp_config:
fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
fsdp_backward_prefetch: BACKWARD_PRE
fsdp_forward_prefetch: false
2024-05-29 18:39:11 +08:00
fsdp_cpu_ram_efficient_loading: true
fsdp_offload_params: true # offload may affect training speed
2024-03-21 00:36:06 +08:00
fsdp_sharding_strategy: FULL_SHARD
2024-03-26 17:53:04 +08:00
fsdp_state_dict_type: FULL_STATE_DICT
2024-03-21 00:36:06 +08:00
fsdp_sync_module_states: true
2024-05-29 18:39:11 +08:00
fsdp_use_orig_params: true
2024-03-21 00:36:06 +08:00
machine_rank: 0
main_training_function: main
2024-05-29 18:39:11 +08:00
mixed_precision: fp16 # or bf16
2024-04-04 14:48:21 +08:00
num_machines: 1 # the number of nodes
num_processes: 2 # the number of GPUs in all nodes
2024-03-21 00:36:06 +08:00
rdzv_backend: static
same_network: true
tpu_env: []
tpu_use_cluster: false
tpu_use_sudo: false
use_cpu: false