search

microsoft / DeepSpeed

DeepSpeed is a deep learning optimization library that makes distributed training and inference easy, efficient, and effective.
https://www.deepspeed.ai/
Apache License 2.0
34.71k stars 4.04k forks source link

[BUG] Fail to Resume From Checkpoint with Different GPU Number(Huggingface Trainer + Deepspeed) #5405

Open Orion-Zheng opened 5 months ago

Orion-Zheng commented 5 months ago

Describe the bug Hello, I encountered a problem when trying to resume from a previous checkpoint when I used Transformers Trainer + ZeRO 3 strategy to train a TinyLlama. My previous run is conducted on *2 nodes 2 A100 40GB on each node**. The structure of the previous checkpoint is showed below.

image

Now I want to resume from this checkpoint with *1 node with 4 A100 40GB**, and the error below occurred. I guess it may related to the different checkpoint format(e.g. the RNG states and model/optim states). Is there any method to consolidate the checkpoint format? Any help will be really appreciated! For us students using HPC clusters, sometimes it's hard for us to always get the same GPU numbers or specifications. Thanks in advance!

    return inner_training_loop(
  File "/home/users/nus/miniconda3/envs/gpu_dev/lib/python3.10/site-packages/transformers/trainer.py", line 1708, in _inner_training_loop
    deepspeed_load_checkpoint(self.model_wrapped, resume_from_checkpoint)
  File "/home/users/nus/miniconda3/envs/gpu_dev/lib/python3.10/site-packages/transformers/integrations/deepspeed.py", line 402, in deepspeed_load_checkpoint
    deepspeed_load_checkpoint(self.model_wrapped, resume_from_checkpoint)
  File "/home/users/nus/miniconda3/envs/gpu_dev/lib/python3.10/site-packages/transformers/integrations/deepspeed.py", line 402, in deepspeed_load_checkpoint
    load_path, _ = deepspeed_engine.load_checkpoint(
  File "/home/users/nus/miniconda3/envs/gpu_dev/lib/python3.10/site-packages/deepspeed/runtime/engine.py", line 2752, in load_checkpoint
    load_path, _ = deepspeed_engine.load_checkpoint(
  File "/home/users/nus/miniconda3/envs/gpu_dev/lib/python3.10/site-packages/deepspeed/runtime/engine.py", line 2752, in load_checkpoint
        load_path, client_states = self._load_checkpoint(load_dir,load_path, client_states = self._load_checkpoint(load_dir,

  File "/home/users/nus/miniconda3/envs/gpu_dev/lib/python3.10/site-packages/deepspeed/runtime/engine.py", line 2804, in _load_checkpoint
  File "/home/users/nus/miniconda3/envs/gpu_dev/lib/python3.10/site-packages/deepspeed/runtime/engine.py", line 2804, in _load_checkpoint
    sd_loader = SDLoaderFactory.get_sd_loader(ckpt_list, checkpoint_engine=self.checkpoint_engine)
  File "/home/users/nus/miniconda3/envs/gpu_dev/lib/python3.10/site-packages/deepspeed/runtime/state_dict_factory.py", line 43, in get_sd_loader
    sd_loader = SDLoaderFactory.get_sd_loader(ckpt_list, checkpoint_engine=self.checkpoint_engine)
  File "/home/users/nus/miniconda3/envs/gpu_dev/lib/python3.10/site-packages/deepspeed/runtime/state_dict_factory.py", line 43, in get_sd_loader
    return MegatronSDLoader(ckpt_list, version, checkpoint_engine)
  File "/home/users/nus/miniconda3/envs/gpu_dev/lib/python3.10/site-packages/deepspeed/runtime/state_dict_factory.py", line 193, in __init__
    super().__init__(ckpt_list, version, checkpoint_engine)
  File "/home/users/nus/miniconda3/envs/gpu_dev/lib/python3.10/site-packages/deepspeed/runtime/state_dict_factory.py", line 55, in __init__
    return MegatronSDLoader(ckpt_list, version, checkpoint_engine)
  File "/home/users/nus/miniconda3/envs/gpu_dev/lib/python3.10/site-packages/deepspeed/runtime/state_dict_factory.py", line 193, in __init__
    self.check_ckpt_list()
  File "/home/users/nus/miniconda3/envs/gpu_dev/lib/python3.10/site-packages/deepspeed/runtime/state_dict_factory.py", line 168, in check_ckpt_list
    super().__init__(ckpt_list, version, checkpoint_engine)
  File "/home/users/nus/miniconda3/envs/gpu_dev/lib/python3.10/site-packages/deepspeed/runtime/state_dict_factory.py", line 55, in __init__
    assert len(self.ckpt_list) > 0
AssertionError

To Reproduce Steps to reproduce the behavior:

  1. Go to '...'
  2. Click on '....'
  3. Scroll down to '....'
  4. See error

Expected behavior A clear and concise description of what you expected to happen.

ds_report output Please run ds_report to give us details about your setup.

Screenshots

System info (please complete the following information):

Launcher context Are you launching your experiment with the deepspeed launcher, MPI, or something else?

Docker context Are you using a specific docker image that you can share?

Additional context Add any other context about the problem here.

tjruwase commented 5 months ago

Now I want to resume from this checkpoint with *1 node with 4 A100 40GB**, and the error below occurred. I guess it may related to the different checkpoint format(e.g. the RNG states and model/optim states). Is there any method to consolidate the checkpoint format?

@Orion-Zheng, thanks for reporting this issue. It is unclear to me that checkpointing consolidation is needed here since it seems you are checkpoint was saved with 4GPUs and resuming with 4GPUs. In other words, no change in number of GPUs between saving and resumption. Is that correct?

Orion-Zheng commented 5 months ago

Now I want to resume from this checkpoint with *1 node with 4 A100 40GB**, and the error below occurred. I guess it may related to the different checkpoint format(e.g. the RNG states and model/optim states). Is there any method to consolidate the checkpoint format?

@Orion-Zheng, thanks for reporting this issue. It is unclear to me that checkpointing consolidation is needed here since it seems you are checkpoint was saved with 4GPUs and resuming with 4GPUs. In other words, no change in number of GPUs between saving and resumption. Is that correct?

Thank you for the timely reply! πŸ˜ƒ I am not very familiar with the components of deepspeed checkpoint. But I think the rank_0 and rank_1 means partitions on different nodeπŸ€”I guess although the total number of gpu is the same(4 GPU), but the checkpoint of 2 node 2 gpu is different from 1 node 4 gpu. Am I correct? If i am correct, can I use the ds_to_universal.py to convert my previous ZeRO3 checkpoint to a universal format so I can resume from it with other GPU settings? I noticed there is an open issue on ds_to_universal.py, https://github.com/microsoft/DeepSpeed/issues/5283 so I am not sure if this issue has been fixed and works for ZeRO 3 checkpoint(previous issue is for ZeRO 2 ckpt). And information will be very appreciated!

tjruwase commented 5 months ago

@Orion-Zheng, deepspeed checkpoint is not aware of node-level information. What matters is parallel dimensions such as data parallel, pipeline parallel, and tensor parallel. So, the checkpoints from 2 nodes 2 GPUs should be the same as 1 node 4 GPUs. I have two questions:

  1. Are you using any form of model parallelism (TP or PP)?
  2. Are the checkpoints of 2 * 2 saved in a global or local storage?
tjruwase commented 5 months ago

@Orion-Zheng, also can you share the log of the run that saved the checkpoint?

Orion-Zheng commented 5 months ago

@tjruwase Thank you for the information!πŸ˜ƒ

  1. I think I didn't use TP and PP (The accelerate and deepspeed configs I used are listed below, please correct me if I am wrong) Accelerate Config 
    compute_environment: LOCAL_MACHINE
debug: false
deepspeed_config:
deepspeed_config_file: config/deepspeed_config/stage3_offload.json
deepspeed_multinode_launcher: standard
zero3_init_flag: false
distributed_type: DEEPSPEED
downcast_bf16: 'no'
machine_rank: 0
main_process_ip: 192.108.1.3
main_process_port: 9999
main_training_function: main
num_machines: 2
num_processes: 4
rdzv_backend: static
same_network: true
tpu_env: []
tpu_use_cluster: false
tpu_use_sudo: false
use_cpu: false

    Deepspeed Config

    {
"bf16": {
    "enabled": "auto"
},
"zero_optimization": {
    "stage": 3,
    "stage3_gather_16bit_weights_on_model_save": true,
    "offload_optimizer": {
        "device": "cpu",
        "pin_memory": true
    },
    "offload_param": {
        "device": "cpu",
        "pin_memory": true
    }
},
"train_batch_size": "auto",
"train_micro_batch_size_per_gpu": "auto",
"gradient_accumulation_steps": "auto",
"zero_allow_untested_optimizer": true,
"gradient_clipping": "auto",
"wall_clock_breakdown": false
}
  2. The checkpoint is saved in global storage.

For the log you mentioned, I am not sure which one you refer to because the Huggingface Trainer seems to only print loss during training, maybe I should set verbose level somewhereπŸ€”I will find it later.

Given that deepspeed checkpoint is not aware of node-level information, I am curious why there are two shards of optim/model/rng states in my checkpoint. Considering I used 2 nodes * 2 gpus, I think the possible number should be 4 shards(equal to gpu number) or 1 shard (totally not aware of node-level and gpu-level information). Is this because stage3_gather_16bit_weights_on_model_save=True in my deepspeed config? Probably the Deepspeed aggregate the optim/model/rng state on each node and save to disk?

rng_state_0.pth
rng_state_1.pth
global_step100/
β€” bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
β€” bf16_zero_P-rank 1 mp_rank_00_optim_states.pt
β€” zero_pp_rank_0_mp_rank_00_model_states .pt
β€” zero_pp_rank_1_mp_rank_00_model_states.pt

By the way, I also hope to know if I want to resume from this ZeRO 3 checkpoint with different GPU numbers, say 3 GPUs. Does ds_to_universal.py help? Is it working well for ZeRO 3 ckpt without bug from https://github.com/microsoft/DeepSpeed/issues/5283 ?πŸ˜ƒ

Orion-Zheng commented 5 months ago

I find the 1 Node 4 GPUs checkpoint structure looks like this, which is different from 2 Node 2 GPU

(gpu_dev) (gpu_dev) bash-4.4$ tree .
.
β”œβ”€β”€ config.json
β”œβ”€β”€ generation_config.json
β”œβ”€β”€ global_step3
β”‚   β”œβ”€β”€ bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
β”‚   β”œβ”€β”€ bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt
β”‚   β”œβ”€β”€ bf16_zero_pp_rank_2_mp_rank_00_optim_states.pt
β”‚   β”œβ”€β”€ bf16_zero_pp_rank_3_mp_rank_00_optim_states.pt
β”‚   β”œβ”€β”€ zero_pp_rank_0_mp_rank_00_model_states.pt
β”‚   β”œβ”€β”€ zero_pp_rank_1_mp_rank_00_model_states.pt
β”‚   β”œβ”€β”€ zero_pp_rank_2_mp_rank_00_model_states.pt
β”‚   └── zero_pp_rank_3_mp_rank_00_model_states.pt
β”œβ”€β”€ latest
β”œβ”€β”€ model.safetensors
β”œβ”€β”€ rng_state_0.pth
β”œβ”€β”€ rng_state_1.pth
β”œβ”€β”€ rng_state_2.pth
β”œβ”€β”€ rng_state_3.pth
β”œβ”€β”€ scheduler.pt
β”œβ”€β”€ trainer_state.json
β”œβ”€β”€ training_args.bin
└── zero_to_fp32.py
Orion-Zheng commented 5 months ago

Hello, I tried to use ds_to_universal.py to convert the deepspeed checkpoint but the error below occurred.

python ds_to_universal.py --input_folder experiment_ckpts/tinyllama_expanded_frez_embed-2024-04-15-011900/checkpoint-3/ \
                                         --output_folder experiment_ckpts/tinyllama_expanded_frez_embed-2024-04-15-011900/checkpoint-3_merge \
                                         --num_extract_workers 10 \
                                         --num_merge_workers 10 \
[2024-04-15 01:43:34,022] [INFO] [real_accelerator.py:203:get_accelerator] Setting ds_accelerator to cuda (auto detect)
 [WARNING]  async_io requires the dev libaio .so object and headers but these were not found.
 [WARNING]  async_io: please install the libaio-devel package with yum
 [WARNING]  If libaio is already installed (perhaps from source), try setting the CFLAGS and LDFLAGS environment variables to where it can be found.
 [WARNING]  Please specify the CUTLASS repo directory as environment variable $CUTLASS_PATH
 [WARNING]  sparse_attn requires a torch version >= 1.5 and < 2.0 but detected 2.2
 [WARNING]  using untested triton version (2.2.0), only 1.0.0 is known to be compatible
args = Namespace(input_folder='experiment_ckpts/tinyllama_expanded_frez_embed-2024-04-15-011900/checkpoint-3', output_folder='experiment_ckpts/tinyllama_expanded_frez_embed-2024-04-15-011900/checkpoint-3_merge', num_extract_workers=10, num_merge_workers=10, keep_temp_folder=False, strict=True)
Convert DeepSpeed Checkpoint to Universal Checkpoint
Converting DeepSpeed checkpoint in experiment_ckpts/tinyllama_expanded_frez_embed-2024-04-15-011900/checkpoint-3 to Universal checkpoint in experiment_ckpts/tinyllama_expanded_frez_embed-2024-04-15-011900/checkpoint-3_merge
Traceback (most recent call last):
  File "/scratch/users/nus/e0792473/EfficientVocabExtend/dist_env_tools/ds_to_universal.py", line 363, in <module>
    main(args)
  File "/scratch/users/nus/e0792473/EfficientVocabExtend/dist_env_tools/ds_to_universal.py", line 319, in main
    ds_checkpoint = DeepSpeedCheckpoint(args.input_folder)
  File "/home/users/nus/e0792473/miniconda3/envs/gpu_dev/lib/python3.10/site-packages/deepspeed/checkpoint/deepspeed_checkpoint.py", line 40, in __init__
    self._validate_folder(dir, pipeline_parallel)
  File "/home/users/nus/e0792473/miniconda3/envs/gpu_dev/lib/python3.10/site-packages/deepspeed/checkpoint/deepspeed_checkpoint.py", line 292, in _validate_folder
    assert len(
AssertionError: experiment_ckpts/tinyllama_expanded_frez_embed-2024-04-15-011900/checkpoint-3 seems a bogus DeepSpeed checkpoint folder: Cannot find mp_rank_* files in there.
tjruwase commented 5 months ago

I find the 1 Node 4 GPUs checkpoint structure looks like this, which is different from 2 Node 2 GPU

(gpu_dev) (gpu_dev) bash-4.4$ tree .
.
β”œβ”€β”€ config.json
β”œβ”€β”€ generation_config.json
β”œβ”€β”€ global_step3
β”‚   β”œβ”€β”€ bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
β”‚   β”œβ”€β”€ bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt
β”‚   β”œβ”€β”€ bf16_zero_pp_rank_2_mp_rank_00_optim_states.pt
β”‚   β”œβ”€β”€ bf16_zero_pp_rank_3_mp_rank_00_optim_states.pt
β”‚   β”œβ”€β”€ zero_pp_rank_0_mp_rank_00_model_states.pt
β”‚   β”œβ”€β”€ zero_pp_rank_1_mp_rank_00_model_states.pt
β”‚   β”œβ”€β”€ zero_pp_rank_2_mp_rank_00_model_states.pt
β”‚   └── zero_pp_rank_3_mp_rank_00_model_states.pt

This confirms that the 2 2 run was saving in local folder not a global (e.g., nfs) folder. With zero3, we have distributed checkpoints where each rank will save a pair of `zero_pp.ptandbf16_.pt` corresponding to its rank. This is why you are seeing half of the files when you inspect the first node of your 2 2 run.

Can you inspect the checkpoint path in both nodes of your 2 * 2 run?

Orion-Zheng commented 5 months ago

OH, I understand. Yes you are right! I just find another directory where the shards from rank 3 and 4 were store in tmp-checkpoint-* directories. I will further investigate why this would happen. Probably because I wrote a custom Callback in Trainer to store the checkpoint. But some mechanisms make the saving not work well.

β”œβ”€β”€ tinyllama_expanded_frez_embed-2024-04-12-221505
β”‚   β”œβ”€β”€ checkpoint-100
β”‚   β”‚   β”œβ”€β”€ config.json
β”‚   β”‚   β”œβ”€β”€ generation_config.json
β”‚   β”‚   β”œβ”€β”€ global_step100
β”‚   β”‚   β”‚   β”œβ”€β”€ bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
β”‚   β”‚   β”‚   β”œβ”€β”€ bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt
β”‚   β”‚   β”‚   β”œβ”€β”€ zero_pp_rank_0_mp_rank_00_model_states.pt
β”‚   β”‚   β”‚   └── zero_pp_rank_1_mp_rank_00_model_states.pt
β”‚   β”‚   β”œβ”€β”€ latest
β”‚   β”‚   β”œβ”€β”€ model.safetensors
β”‚   β”‚   β”œβ”€β”€ rng_state_0.pth
β”‚   β”‚   β”œβ”€β”€ rng_state_1.pth
β”‚   β”‚   β”œβ”€β”€ scheduler.pt
β”‚   β”‚   β”œβ”€β”€ trainer_state.json
β”‚   β”‚   β”œβ”€β”€ training_args.bin
β”‚   β”‚   └── zero_to_fp32.py
β”‚   └── checkpoint-132
β”‚       β”œβ”€β”€ config.json
β”‚       β”œβ”€β”€ generation_config.json
β”‚       β”œβ”€β”€ global_step132
β”‚       β”‚   β”œβ”€β”€ bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
β”‚       β”‚   β”œβ”€β”€ bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt
β”‚       β”‚   β”œβ”€β”€ zero_pp_rank_0_mp_rank_00_model_states.pt
β”‚       β”‚   └── zero_pp_rank_1_mp_rank_00_model_states.pt
β”‚       β”œβ”€β”€ latest
β”‚       β”œβ”€β”€ model.safetensors
β”‚       β”œβ”€β”€ rng_state_0.pth
β”‚       β”œβ”€β”€ rng_state_1.pth
β”‚       β”œβ”€β”€ scheduler.pt
β”‚       β”œβ”€β”€ trainer_state.json
β”‚       β”œβ”€β”€ training_args.bin
β”‚       └── zero_to_fp32.py
β”œβ”€β”€ tinyllama_expanded_frez_embed-2024-04-12-221513
β”‚   β”œβ”€β”€ tmp-checkpoint-100
β”‚   β”‚   β”œβ”€β”€ global_step100
β”‚   β”‚   β”‚   β”œβ”€β”€ bf16_zero_pp_rank_2_mp_rank_00_optim_states.pt
β”‚   β”‚   β”‚   β”œβ”€β”€ bf16_zero_pp_rank_3_mp_rank_00_optim_states.pt
β”‚   β”‚   β”‚   β”œβ”€β”€ zero_pp_rank_2_mp_rank_00_model_states.pt
β”‚   β”‚   β”‚   └── zero_pp_rank_3_mp_rank_00_model_states.pt
β”‚   β”‚   β”œβ”€β”€ rng_state_2.pth
β”‚   β”‚   └── rng_state_3.pth
β”‚   └── tmp-checkpoint-132
β”‚       β”œβ”€β”€ global_step132
β”‚       β”‚   β”œβ”€β”€ bf16_zero_pp_rank_2_mp_rank_00_optim_states.pt
β”‚       β”‚   β”œβ”€β”€ bf16_zero_pp_rank_3_mp_rank_00_optim_states.pt
β”‚       β”‚   β”œβ”€β”€ zero_pp_rank_2_mp_rank_00_model_states.pt
β”‚       β”‚   └── zero_pp_rank_3_mp_rank_00_model_states.pt
β”‚       β”œβ”€β”€ rng_state_2.pth
β”‚       └── rng_state_3.pth
Orion-Zheng commented 5 months ago

@tjruwase Thank you! Now I understand what you mean. May I ask how to convert a 4-GPU checkpoint to a universal checkpoint so I can continue training with other gpu numbers?πŸ˜ƒThis is the Google Drive link to my 4-GPU checkpoint. https://drive.google.com/drive/folders/1efsIy9CcI_dTL41AtxdpcCJnSF1dn_uX?usp=sharing I tried the ds_to_universal.py but it seems doesn't work(xxx seems a bogus DeepSpeed checkpoint folder: Cannot find mp_rank_* files in there.). Could you help me with this? 

python ds_to_universal.py --input_folder my_4_gpu_ckpt \
                                         --output_folder my_4_gpu_ckpt_merge \
                                         --num_extract_workers 10 \
                                         --num_merge_workers 10 \
Orion-Zheng commented 5 months ago

OH, I understand. Yes you are right! I just find another directory where the shards from rank 3 and 4 were store in tmp-checkpoint-* directories. I will further investigate why this would happen. Probably because I wrote a custom Callback in Trainer to store the checkpoint. But some mechanisms make the saving not work well.

β”œβ”€β”€ tinyllama_expanded_frez_embed-2024-04-12-221505
β”‚   β”œβ”€β”€ checkpoint-100
β”‚   β”‚   β”œβ”€β”€ config.json
β”‚   β”‚   β”œβ”€β”€ generation_config.json
β”‚   β”‚   β”œβ”€β”€ global_step100
β”‚   β”‚   β”‚   β”œβ”€β”€ bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
β”‚   β”‚   β”‚   β”œβ”€β”€ bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt
β”‚   β”‚   β”‚   β”œβ”€β”€ zero_pp_rank_0_mp_rank_00_model_states.pt
β”‚   β”‚   β”‚   └── zero_pp_rank_1_mp_rank_00_model_states.pt
β”‚   β”‚   β”œβ”€β”€ latest
β”‚   β”‚   β”œβ”€β”€ model.safetensors
β”‚   β”‚   β”œβ”€β”€ rng_state_0.pth
β”‚   β”‚   β”œβ”€β”€ rng_state_1.pth
β”‚   β”‚   β”œβ”€β”€ scheduler.pt
β”‚   β”‚   β”œβ”€β”€ trainer_state.json
β”‚   β”‚   β”œβ”€β”€ training_args.bin
β”‚   β”‚   └── zero_to_fp32.py
β”‚   └── checkpoint-132
β”‚       β”œβ”€β”€ config.json
β”‚       β”œβ”€β”€ generation_config.json
β”‚       β”œβ”€β”€ global_step132
β”‚       β”‚   β”œβ”€β”€ bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
β”‚       β”‚   β”œβ”€β”€ bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt
β”‚       β”‚   β”œβ”€β”€ zero_pp_rank_0_mp_rank_00_model_states.pt
β”‚       β”‚   └── zero_pp_rank_1_mp_rank_00_model_states.pt
β”‚       β”œβ”€β”€ latest
β”‚       β”œβ”€β”€ model.safetensors
β”‚       β”œβ”€β”€ rng_state_0.pth
β”‚       β”œβ”€β”€ rng_state_1.pth
β”‚       β”œβ”€β”€ scheduler.pt
β”‚       β”œβ”€β”€ trainer_state.json
β”‚       β”œβ”€β”€ training_args.bin
β”‚       └── zero_to_fp32.py
β”œβ”€β”€ tinyllama_expanded_frez_embed-2024-04-12-221513
β”‚   β”œβ”€β”€ tmp-checkpoint-100
β”‚   β”‚   β”œβ”€β”€ global_step100
β”‚   β”‚   β”‚   β”œβ”€β”€ bf16_zero_pp_rank_2_mp_rank_00_optim_states.pt
β”‚   β”‚   β”‚   β”œβ”€β”€ bf16_zero_pp_rank_3_mp_rank_00_optim_states.pt
β”‚   β”‚   β”‚   β”œβ”€β”€ zero_pp_rank_2_mp_rank_00_model_states.pt
β”‚   β”‚   β”‚   └── zero_pp_rank_3_mp_rank_00_model_states.pt
β”‚   β”‚   β”œβ”€β”€ rng_state_2.pth
β”‚   β”‚   └── rng_state_3.pth
β”‚   └── tmp-checkpoint-132
β”‚       β”œβ”€β”€ global_step132
β”‚       β”‚   β”œβ”€β”€ bf16_zero_pp_rank_2_mp_rank_00_optim_states.pt
β”‚       β”‚   β”œβ”€β”€ bf16_zero_pp_rank_3_mp_rank_00_optim_states.pt
β”‚       β”‚   β”œβ”€β”€ zero_pp_rank_2_mp_rank_00_model_states.pt
β”‚       β”‚   └── zero_pp_rank_3_mp_rank_00_model_states.pt
β”‚       β”œβ”€β”€ rng_state_2.pth
β”‚       └── rng_state_3.pth

Oh I think I know the reason. It was my fault. I start distributed training by manually run the same script on two nodes. In my script, the checkpoint path will use the real timestamp as the suffix of the ckpt path. Because two scripts didn't launch exactly at the same time, they save checkpoint to different directories.

tjruwase commented 5 months ago

@Orion-Zheng, I am glad the original mystery is now resolved. For your second question, unfortunately zero3 is not yet supported. Is it possible for you to use zero2? I think zero2 should be adequate for your model size.

@minjiaz, @xylian86, @zhangninja FYI

Orion-Zheng commented 5 months ago

@Orion-Zheng, I am glad the original mystery is now resolved. For your second question, unfortunately zero3 is not yet supported. Is it possible for you to use zero2? I think zero2 should be adequate for your model size.

@minjiaz, @xylian86, @zhangninja FYI

Thank you! Yes, the ZeRO-2 code does run successfully. But in my case, when using a large batch size, I can only use ZeRO-3. Probably I have to wait for the ZeRO-3 supportπŸ˜ƒ By the way, may I ask what's the difference between ZeRO-2 and ZeRO-3 checkpoint? According the paper, the ZeRO-3 further shards model parameters compared to ZeRO-2. But in my ZeRO-3 checkpoint, the model parameters have been merged into one model.safetensors shard (I guess it's due to stage3_gather_16bit_weights_on_model_save = True). So I think my checkpoint should be very similar to ZeRO-2 checkpoint?πŸ€”Please correct me if I am wrong

image
tjruwase commented 5 months ago

@Orion-Zheng, with zero2 checkpoints there is a single mp_rank* checkpoint file containing unshared parameter information (fp16/bf16 and layers) in addition to the 4 bf16_zero* files. Whereas for zero3, there are 4 zero_pp* files corresponding to the sharded parameter information. So, zero2 and zero3 checkpoints are not compatible.

If zero2 is failing on large batch size, a more appropriate solution is gradient checkpointing.

Orion-Zheng commented 5 months ago

Oh yes, good idea! I will try ZeRO-2 + Gradient Checkpointing later :) Thank you for your help!

Orion-Zheng commented 5 months ago

@tjruwase Hi😳sorry to bother again. I tried ZeRO-2 and got a ZeRO-2 checkpoint. But it seems the Accelerate+Deepspeed checkpoint structure is a bit different from Universal Checkpoint examples in Megatron-Deepspeed. So some errors occurred.

args = Namespace(input_folder='experiment_ckpts/tinyllama_expanded_frez_embed-2024-04-16-010251/checkpoint-3', output_folder='experiment_ckpts/tinyllama_expanded_frez_embed-2024-04-16-010251/checkpoint-3_universal', num_extract_workers=10, num_merge_workers=10, keep_temp_folder=False, strict=True)
Convert DeepSpeed Checkpoint to Universal Checkpoint
Converting DeepSpeed checkpoint in experiment_ckpts/tinyllama_expanded_frez_embed-2024-04-16-010251/checkpoint-3 to Universal checkpoint in experiment_ckpts/tinyllama_expanded_frez_embed-2024-04-16-010251/checkpoint-3_universal
Traceback (most recent call last):
  File "dist_env_tools/ds_to_universal.py", line 363, in <module>
    main(args)
  File "dist_env_tools/ds_to_universal.py", line 320, in main
    _check_for_required_state(ds_checkpoint)
  File "dist_env_tools/ds_to_universal.py", line 311, in _check_for_required_state
    assert universal_checkpoint_info is not None, f'Required {UNIVERSAL_CHECKPOINT_INFO} state is missing in checkpoint. Verify that client creates this state.'
AssertionError: Required universal_checkpoint_info state is missing in checkpoint. Verify that client creates this state.

This is my ZeRO-2 checkpoint's structure and this is the Google Drive link.

checkpoint-3/
β”œβ”€β”€ config.json
β”œβ”€β”€ generation_config.json
β”œβ”€β”€ global_step3
β”‚   β”œβ”€β”€ bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
β”‚   β”œβ”€β”€ bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt
β”‚   β”œβ”€β”€ bf16_zero_pp_rank_2_mp_rank_00_optim_states.pt
β”‚   β”œβ”€β”€ bf16_zero_pp_rank_3_mp_rank_00_optim_states.pt
β”‚   └── mp_rank_00_model_states.pt
β”œβ”€β”€ latest
β”œβ”€β”€ model.safetensors
β”œβ”€β”€ rng_state_0.pth
β”œβ”€β”€ rng_state_1.pth
β”œβ”€β”€ rng_state_2.pth
β”œβ”€β”€ rng_state_3.pth
β”œβ”€β”€ scheduler.pt
β”œβ”€β”€ trainer_state.json
β”œβ”€β”€ training_args.bin
└── zero_to_fp32.py

Although I have not idea how to rectify it, I think for you guys it only takes one look to see how to make it compatible with the ZeRO-2 format in my case.πŸ˜ƒ Any help will be very appreciated! Besides, I am also curious about how to deal with multiple rng_state_*.pth if I resume training with different gpus?πŸ€”I think these seem unmergable.

ldh127 commented 4 months ago

@tjruwase Hi😳sorry to bother again. I tried ZeRO-2 and got a ZeRO-2 checkpoint. But it seems the Accelerate+Deepspeed checkpoint structure is a bit different from Universal Checkpoint examples in Megatron-Deepspeed. So some errors occurred.

args = Namespace(input_folder='experiment_ckpts/tinyllama_expanded_frez_embed-2024-04-16-010251/checkpoint-3', output_folder='experiment_ckpts/tinyllama_expanded_frez_embed-2024-04-16-010251/checkpoint-3_universal', num_extract_workers=10, num_merge_workers=10, keep_temp_folder=False, strict=True)
Convert DeepSpeed Checkpoint to Universal Checkpoint
Converting DeepSpeed checkpoint in experiment_ckpts/tinyllama_expanded_frez_embed-2024-04-16-010251/checkpoint-3 to Universal checkpoint in experiment_ckpts/tinyllama_expanded_frez_embed-2024-04-16-010251/checkpoint-3_universal
Traceback (most recent call last):
  File "dist_env_tools/ds_to_universal.py", line 363, in <module>
    main(args)
  File "dist_env_tools/ds_to_universal.py", line 320, in main
    _check_for_required_state(ds_checkpoint)
  File "dist_env_tools/ds_to_universal.py", line 311, in _check_for_required_state
    assert universal_checkpoint_info is not None, f'Required {UNIVERSAL_CHECKPOINT_INFO} state is missing in checkpoint. Verify that client creates this state.'
AssertionError: Required universal_checkpoint_info state is missing in checkpoint. Verify that client creates this state.

This is my ZeRO-2 checkpoint's structure and this is the Google Drive link.

checkpoint-3/
β”œβ”€β”€ config.json
β”œβ”€β”€ generation_config.json
β”œβ”€β”€ global_step3
β”‚   β”œβ”€β”€ bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
β”‚   β”œβ”€β”€ bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt
β”‚   β”œβ”€β”€ bf16_zero_pp_rank_2_mp_rank_00_optim_states.pt
β”‚   β”œβ”€β”€ bf16_zero_pp_rank_3_mp_rank_00_optim_states.pt
β”‚   └── mp_rank_00_model_states.pt
β”œβ”€β”€ latest
β”œβ”€β”€ model.safetensors
β”œβ”€β”€ rng_state_0.pth
β”œβ”€β”€ rng_state_1.pth
β”œβ”€β”€ rng_state_2.pth
β”œβ”€β”€ rng_state_3.pth
β”œβ”€β”€ scheduler.pt
β”œβ”€β”€ trainer_state.json
β”œβ”€β”€ training_args.bin
└── zero_to_fp32.py

Although I have not idea how to rectify it, I think for you guys it only takes one look to see how to make it compatible with the ZeRO-2 format in my case.πŸ˜ƒ Any help will be very appreciated! Besides, I am also curious about how to deal with multiple rng_state_*.pth if I resume training with different gpus?πŸ€”I think these seem unmergable.

yes, i also find this , but i use transformers trainer call deepspeed，and get the same file as you , have you reslove this problem？

tjruwase commented 4 months ago

Besides, I am also curious about how to deal with multiple rng_state_*.pth if I resume training with different gpus?πŸ€”I think these seem unmergable.

Since you are doing data parallel training, those should be duplicates and you need only one. Things will get interesting with model parallel training.

tjruwase commented 4 months ago

@tjruwase Hi😳sorry to bother again. I tried ZeRO-2 and got a ZeRO-2 checkpoint. But it seems the Accelerate+Deepspeed checkpoint structure is a bit different from Universal Checkpoint examples in Megatron-Deepspeed. So some errors occurred.

@Orion-Zheng, you are hitting this error because some work is needed to port Universal Checkpointing to Accelerate. Currently, we have only ported to Megatron-DeepSpeed. However, in this case, can you confirm that you are planning to change the numbers of GPUs in your training?

tjruwase commented 4 months ago

@Orion-Zheng, are you still having this issue?

xiyang-aads-lilly commented 3 weeks ago

@tjruwase is this issue solved? I try to continue training the model switching from 4 nodes to 2 nodes and encounter the same problem using huggingface trainer.

args = Namespace(input_folder='experiment_ckpts/tinyllama_expanded_frez_embed-2024-04-16-010251/checkpoint-3', output_folder='experiment_ckpts/tinyllama_expanded_frez_embed-2024-04-16-010251/checkpoint-3_universal', num_extract_workers=10, num_merge_workers=10, keep_temp_folder=False, strict=True)
Convert DeepSpeed Checkpoint to Universal Checkpoint
Converting DeepSpeed checkpoint in experiment_ckpts/tinyllama_expanded_frez_embed-2024-04-16-010251/checkpoint-3 to Universal checkpoint in experiment_ckpts/tinyllama_expanded_frez_embed-2024-04-16-010251/checkpoint-3_universal
Traceback (most recent call last):
  File "dist_env_tools/ds_to_universal.py", line 363, in <module>
    main(args)
  File "dist_env_tools/ds_to_universal.py", line 320, in main
    _check_for_required_state(ds_checkpoint)
  File "dist_env_tools/ds_to_universal.py", line 311, in _check_for_required_state
    assert universal_checkpoint_info is not None, f'Required {UNIVERSAL_CHECKPOINT_INFO} state is missing in checkpoint. Verify that client creates this state.'
AssertionError: Required universal_checkpoint_info state is missing in checkpoint. Verify that client creates this state.
tjruwase commented 3 weeks ago

@xylian86, can you please help with this?

xylian86 commented 3 weeks ago

@tjruwase Yes, for sure. @xiyang-aads-lilly Could you please try 1. Install the latest DeepSpeed version (v0.14.5 Patch release) 2. Add the argument inject_missing_state when you run the conversion?

ArtificialZeng commented 2 weeks ago

@Orion-Zheng, are you still having this issue?

yes