examples/train_lora/llama3_lora_sft_ds3.yaml 报错

[JerryZeyu]

Reminder

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

System Info

用ds_z3_config.json的时候就会报错，错误显示：pydantic_core._pydantic_core.ValidationError: 1 validation error for DeepSpeedZeroConfig
[rank3]: stage3_prefetch_bucket_size
[rank3]: Input should be a valid integer, got a number with a fractional part [type=int_from_float, input_value=15099494.4, input_type=float]

请问这是deepspeed的版本问题吗

Reproduction

torch == 2.4.0
deepspeed == 0.15.0
llamafactory-cli train examples/train_lora/llama3_lora_sft_ds3.yaml

Expected behavior

No response

Others

No response

[JerryZeyu]

但是用llamafactory-cli train examples/train_lora/llama3_lora_sft_ds0.yaml的时候就不会报错

[zhangguoxin1]

我也遇到此类问题

[junqi-lu]

同样的错误，将 deepspeed 退回到 0.14.0 版本对我有用。

[sunzhufeng12345]

将 deepspeed 退回到 0.14.0 版本又报与pytroch版本不对应，无法运行，pytroch是根据cuda版本来的

[junqi-lu]

将 deepspeed 退回到 0.14.0 版本又报与pytroch版本不对应，无法运行，pytroch是根据cuda版本来的

所有依赖版本都采用仓库推荐版本呢？

[sunzhufeng12345]

将 deepspeed 退回到 0.14.0 版本又报与pytroch版本不对应，无法运行，pytroch是根据cuda版本来的

所有依赖版本都采用仓库推荐版本呢？
我与楼主的版本一样，也是出现了相同的问题，我不是使用llamafactory训练的，是复现longwrite的时候遇到的这个问题
torch == 2.4.0
deepspeed == 0.15.0

[gannim]

I encountered a similar issue, and it was resolved by using DeepSpeed version 0.14.4. I suspect that the problem arises in later versions of DeepSpeed due to type checking with Pydantic. Specifically, when the stage3_prefetch_bucket_size option is set to auto, Accelerate calculates it based on the model's hidden size. However, I suspect that it might not be properly converted to an integer during this process, leading to the error.

[chenhuiyu]

I encountered a similar issue, and it was resolved by using DeepSpeed version 0.14.4. I suspect that the problem arises in later versions of DeepSpeed due to type checking with Pydantic. Specifically, when the stage3_prefetch_bucket_size option is set to auto, Accelerate calculates it based on the model's hidden size. However, I suspect that it might not be properly converted to an integer during this process, leading to the error.

Thanks! this solution solve my issue!

[HughesZhang2021]

deepspeed==0.14.4 solved