DeepSpeed运行报错‘DeepSpeedCPUAdam’ object has no attribute ‘ds_opt_adam’由ninja触发 - 解决

kappam

已于 2024-10-02 14:50:45 修改

阅读量280

点赞数 9

文章标签： python llama transformer nlp

于 2024-10-02 13:00:16 首次发布

本文链接：https://blog.csdn.net/weixin_44646237/article/details/142678911

版权

用deepspeed运行python程序时出现报错
AttributeError: ‘DeepSpeedCPUAdam’ object has no attribute ‘ds_opt_adam’ 的一种情况是由ninja的报错导致的:
xxx/.local/bin/ninja", line 5, in
from ninja import ninja
ModuleNotFoundError: No module named ‘ninja’

分析是由于程序识别ninja的位置问题导致的，上述报错说明python去用户本地路径下而不是从conda环境中寻找ninja

测试环境 python 3.12 + CUDA 12.1

accelerate==0.34.2
deepspeed==0.15.1
ninja==1.11.1.1
torch==2.4.0
transformers==4.44.2
trl==0.11.1

解决方法

将用户路径中的ninja进行更名备份:

 mv ~/.local/bin/ninja ~/.local/bin/ninja.bak

使用pip安装ninja

pip install ninja

重新加载conda环境

source ~/.bashrc
conda activate ENV_NAME

问题解决

附上相关代码/指令参考

deepspeed_config: 设置 bf16 为 True以防止报错Current loss scale already at minimum - cannot decrease scale anymore. Exiting run.

deepspeed_config = {
    "fp16": {
        "enabled": False,
        "loss_scale": 0,
        "loss_scale_window": 1000,
        "initial_scale_power": 16,
        "hysteresis": 2,
        "min_loss_scale": 1
    },
	"bf16": {
		"enabled": True,
	},
    "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": 2,
        "offload_optimizer": {
            "device": "cpu",
            "pin_memory": True
        },
        "allgather_partitions": True,
        "allgather_bucket_size": 2e8,
        "overlap_comm": True,
        "reduce_scatter": True,
        "reduce_bucket_size": 2e8,
        "contiguous_gradients": True
    },
    "steps_per_print": 100,
    "gradient_accumulation_steps": "auto",
    "gradient_clipping": "auto",
    "train_batch_size": "auto",
    "train_micro_batch_size_per_gpu": "auto",
    "activation_checkpointing": {
        "partition_activations": True,
        "contiguous_memory_optimization": True
    },
    "wall_clock_breakdown": False
}

finetune.py

base_model = '/path/to/Meta-Llama-3.1-8B-Instruct'
model = AutoModelForCausalLM.from_pretrained(
	base_model,
	device_map=device_map,
)

train_args=transformers.TrainingArguments(
	deepspeed=deepspeed_config,
	num_train_epochs=num_epochs,
	learning_rate=learning_rate,
	fp16=deepspeed_config['fp16']['enabled'],
	bf16=deepspeed_config['bf16']['enabled'],
	save_safetensors=False,
	logging_steps=50,
	logging_dir=os.path.join(output_dir, 'logs'),
	evaluation_strategy="steps" if val_set_size > 0 else "no",
	save_strategy="steps",
	eval_steps=eval_steps if val_set_size > 0 else None,
	save_steps=save_steps,
	output_dir=output_dir,
	save_total_limit=2,
	group_by_length=group_by_length,
	save_only_model=True,
	seed=42,
	data_seed=42,
)

# trainer = transformers.Trainer(
trainer = SFTTrainer(
	model=model,
	train_dataset=train_data,
	eval_dataset=val_data,
	args=train_args,
	data_collator=transformers.DataCollatorForSeq2Seq(
		tokenizer, pad_to_multiple_of=8, return_tensors="pt", padding=True
	),
)

model.config.use_cache = False

trainer.train()

model.save_pretrained(output_dir, safe_serialization=False, max_shard_size='5GB')

运行指令: 使用deepspeed（而不是python）运行; ~~CUDA_VISIBLE_DEVICES~~ 不适用于deepspeed，使用localhost指定GPU。

LR=5e-4
GPUID=2
MODEL_DIR=/path/to/Meta-Llama-3.1-8B-Instruct
deepspeed --include localhost:${GPUID} finetune.py \
    --base_model ${MODEL_DIR} \
    --output_dir ${OUTPUT_DIR} \
    --batch_size 4 \
    --micro_batch_size 4 \
    --num_epochs 20 \
    --learning_rate ${LR} \
    --cutoff_len 150 \
    --eval_steps 2000 \
    --save_steps 2000 \
    --val_set_size 6000 \
    --train_on_inputs false \
    --group_by_length \