DeepSpeed Zero3 and Peft LoRA fp16 issue

[conceptofmind]

Hi all,

I am having an issue when running Peft LoRA with DeepSpeed Zero3.

Error:

ValueError: fp16 is enabled but the following parameters have dtype that is not fp16: base_model.model.gpt_neox.layers.0.attention.query_key_value.lora_A.weight, 
base_model.model.gpt_neox.layers.0.attention.query_key_value.lora_B.weight, base_model.model.gpt_neox.layers.1.attention.query_key_value.lora_A.weight, 
base_model.model.gpt_neox.layers.1.attention.query_key_value.lora_B.weight, base_model.model.gpt_neox.layers.2.attention.query_key_value.lora_A.weight, 
base_model.model.gpt_neox.layers.2.attention.query_key_value.lora_B.weight, base_model.model.gpt_neox.layers.3.attention.query_key_value.lora_A.weight,

How to reproduce:

CLI:

deepspeed finetune_pythia.py --per_device_train_batch_size 1 --output_dir /home/training_scripts/pythia-1.4b --fp16 --deepspeed configs/ds_z3_config.json

Code:

from transformers import AutoTokenizer, GPTNeoXForCausalLM
from peft import get_peft_model, LoraConfig, TaskType

peft_config = LoraConfig(
  task_type=TaskType.CAUSAL_LM, 
  inference_mode=False, 
  r=16, 
  lora_alpha=32, 
  lora_dropout=0.1,
)

tokenizer = AutoTokenizer.from_pretrained("EleutherAI/pythia-1.4b")
model = GPTNeoXForCausalLM.from_pretrained("EleutherAI/pythia-1.4b")
model = get_peft_model(model, peft_config)
model.print_trainable_parameters()

DeepSpeed Zero 3 config:

{
    "fp16": {
        "enabled": "auto",
        "loss_scale": 0,
        "loss_scale_window": 1000,
        "initial_scale_power": 16,
        "hysteresis": 2,
        "min_loss_scale": 1
    },

    "optimizer": {
        "type": "AdamW",
        "params": {
            "lr": "auto",
            "betas": "auto",
            "eps": "auto",
            "weight_decay": "auto"
        }
    },

    "scheduler": {
        "type": "WarmupLR",
        "params": {
            "warmup_min_lr": "auto",
            "warmup_max_lr": "auto",
            "warmup_num_steps": "auto"
        }
    },

    "zero_optimization": {
        "stage": 3,
        "offload_optimizer": {
            "device": "cpu",
            "pin_memory": true
        },
        "offload_param": {
            "device": "cpu",
            "pin_memory": true
        },
        "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
    },

    "gradient_accumulation_steps": "auto",
    "gradient_clipping": "auto",
    "steps_per_print": 2000,
    "train_batch_size": "auto",
    "train_micro_batch_size_per_gpu": "auto",
    "wall_clock_breakdown": false
}

Any help would be greatly appreciated.

Thank you,

Enrico

[pacman100]

@conceptofmind, could you please try the latest version of DeepSpeed 0.8.2 along with PEFT from the main branch and let us know if that solves the issue?

[pacman100]

I am unable to reproduce the error, the output that I get when running the minimal example from issue desc:

[2023-03-07 20:34:42,145] [WARNING] [runner.py:186:fetch_hostfile] Unable to find hostfile, will proceed with training with local resources only.
[2023-03-07 20:34:42,185] [INFO] [runner.py:550:main] cmd = /home/sourab/miniconda3/envs/ml/bin/python -u -m deepspeed.launcher.launch --world_info=eyJsb2NhbGhvc3QiOiBbMCwgMSwgMiwgMywgNF19 --master_addr=127.0.0.1 --master_port=29500 --enable_each_rank_log=None issue_138.py --per_device_train_batch_size 1 --output_dir /home/training_scripts/pythia-1.4b --fp16 --deepspeed configs/ds_z3_config.json
[2023-03-07 20:34:43,667] [INFO] [launch.py:142:main] WORLD INFO DICT: {'localhost': [0, 1, 2, 3, 4]}
[2023-03-07 20:34:43,667] [INFO] [launch.py:148:main] nnodes=1, num_local_procs=5, node_rank=0
[2023-03-07 20:34:43,667] [INFO] [launch.py:161:main] global_rank_mapping=defaultdict(<class 'list'>, {'localhost': [0, 1, 2, 3, 4]})
[2023-03-07 20:34:43,667] [INFO] [launch.py:162:main] dist_world_size=5
[2023-03-07 20:34:43,667] [INFO] [launch.py:164:main] Setting CUDA_VISIBLE_DEVICES=0,1,2,3,4

===================================BUG REPORT===================================
Welcome to bitsandbytes. For bug reports, please submit your error trace to: https://github.com/TimDettmers/bitsandbytes/issues
================================================================================

===================================BUG REPORT===================================
Welcome to bitsandbytes. For bug reports, please submit your error trace to: https://github.com/TimDettmers/bitsandbytes/issues
================================================================================

===================================BUG REPORT===================================
Welcome to bitsandbytes. For bug reports, please submit your error trace to: https://github.com/TimDettmers/bitsandbytes/issues
================================================================================

===================================BUG REPORT===================================
Welcome to bitsandbytes. For bug reports, please submit your error trace to: https://github.com/TimDettmers/bitsandbytes/issues
================================================================================

===================================BUG REPORT===================================
Welcome to bitsandbytes. For bug reports, please submit your error trace to: https://github.com/TimDettmers/bitsandbytes/issues
================================================================================
Downloading: 100%|██████████████████████████████████████████████████| 394/394 [00:00<00:00, 847kB/s]
Downloading: 100%|█████████████████████████████████████████████| 2.11M/2.11M [00:00<00:00, 4.47MB/s]
Downloading: 100%|████████████████████████████████████████████████| 99.0/99.0 [00:00<00:00, 108kB/s]
Downloading: 100%|██████████████████████████████████████████████████| 537/537 [00:00<00:00, 409kB/s]
Downloading: 100%|█████████████████████████████████████████████| 2.93G/2.93G [01:14<00:00, 39.2MB/s]
/home/sourab/pet/src/peft/tuners/lora.py:173: UserWarning: fan_in_fan_out is set to True but the target module is not a Conv1D. Setting fan_in_fan_out to False.
  warnings.warn(
/home/sourab/pet/src/peft/tuners/lora.py:173: UserWarning: fan_in_fan_out is set to True but the target module is not a Conv1D. Setting fan_in_fan_out to False.
  warnings.warn(
/home/sourab/pet/src/peft/tuners/lora.py:173: UserWarning: fan_in_fan_out is set to True but the target module is not a Conv1D. Setting fan_in_fan_out to False.
  warnings.warn(
/home/sourab/pet/src/peft/tuners/lora.py:173: UserWarning: fan_in_fan_out is set to True but the target module is not a Conv1D. Setting fan_in_fan_out to False.
  warnings.warn(
/home/sourab/pet/src/peft/tuners/lora.py:173: UserWarning: fan_in_fan_out is set to True but the target module is not a Conv1D. Setting fan_in_fan_out to False.
  warnings.warn(
trainable params: 3145728 || all params: 1417793536 || trainable%: 0.22187490069075896
trainable params: 3145728 || all params: 1417793536 || trainable%: 0.22187490069075896
trainable params: 3145728 || all params: 1417793536 || trainable%: 0.22187490069075896
trainable params: 3145728 || all params: 1417793536 || trainable%: 0.22187490069075896
trainable params: 3145728 || all params: 1417793536 || trainable%: 0.22187490069075896
[2023-03-07 20:36:22,771] [INFO] [launch.py:350:main] Process 2399300 exits successfully.
[2023-03-07 20:36:23,772] [INFO] [launch.py:350:main] Process 2399303 exits successfully.
[2023-03-07 20:36:23,772] [INFO] [launch.py:350:main] Process 2399301 exits successfully.
[2023-03-07 20:36:23,772] [INFO] [launch.py:350:main] Process 2399299 exits successfully.
[2023-03-07 20:36:24,773] [INFO] [launch.py:350:main] Process 2399302 exits successfully.

[conceptofmind]

@conceptofmind, could you please try the latest version of DeepSpeed 0.8.2 along with PEFT from the main branch and let us know if that solves the issue?

I will update everything and rerun the script. Thank you.

[alex-ht]

use model = model.half() to convert dtype

...
model = get_peft_model(model, peft_config)
model = model.half()
...

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.