deepspeed不起作用

[N-Kingsley]

Reminder

• I have read the README and searched the existing issues.

Reproduction

examples/lora_multi_gpu/ds_zero3.sh
#!/bin/bash

NPROC_PER_NODE=4

CUDA_VISIBLE_DEVICES=0,1,5,6 python -m torch.distributed.run
--nproc_per_node $NPROC_PER_NODE
--nnodes 1
--standalone
src/train.py examples/lora_multi_gpu/baichuan2_lora_sft_ds.yaml

examples/lora_multi_gpu/baichuan2_lora_sft_ds.yaml

model

model_name_or_path: /home/nk/home/nk/eveProject_pytorch/ChatGPT/baichuan-inc/Baichuan2-7B-Chat

method

stage: sft
do_train: true
finetuning_type: lora
lora_target: W_pack

ddp

ddp_timeout: 180000000
deepspeed: examples/deepspeed/ds_z3_config.json

dataset

dataset: identity,alpaca_gpt4_en
template: baichuan2
cutoff_len: 1024
max_samples: 1000
val_size: 0.1
overwrite_cache: true
preprocessing_num_workers: 16

output

output_dir: saves/baichuan2-7b-chat/lora/sft
logging_steps: 10
save_steps: 500
plot_loss: true
overwrite_output_dir: true

train

per_device_train_batch_size: 1
gradient_accumulation_steps: 2
learning_rate: 0.0001
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_steps: 0.1
fp16: true

eval

per_device_eval_batch_size: 1
evaluation_strategy: steps
eval_steps: 500

ds_z3_config.json 没有改变
{
"train_batch_size": "auto",
"train_micro_batch_size_per_gpu": "auto",
"gradient_accumulation_steps": "auto",
"gradient_clipping": "auto",
"zero_allow_untested_optimizer": true,
"fp16": {
"enabled": "auto",
"loss_scale": 0,
"loss_scale_window": 1000,
"initial_scale_power": 16,
"hysteresis": 2,
"min_loss_scale": 1
},
"bf16": {
"enabled": "auto"
},
"zero_optimization": {
"stage": 3,
"overlap_comm": true,
"contiguous_gradients": true,
"sub_group_size": 1e9,
"reduce_bucket_size": "auto",
"stage3_prefetch_bucket_size": "auto",
"stage3_param_persistence_threshold": "auto",
"stage3_max_live_parameters": 1e9,
"stage3_max_reuse_distance": 1e9,
"stage3_gather_16bit_weights_on_model_save": true
}

4*2080ti的卡还是报显存不足，而用pytorch fsdp就能跑起来

Expected behavior

No response

System Info

No response

Others

No response

[N-Kingsley]

lora微调目前支持deepspeed吗？
我用zero3阶段，分多张卡，按理来说，卡越多，每一张卡上占用的显存更少。
但是实际上增加卡的数量，每一张卡的显存占用不变。