zhengyangyong

[email protected]

@apache

Results 8 comments of


                                            zhengyangyong

sqlite requires System.Data.SQLite, doesn't play nicely with SQLite PCL

yes,also think need support t4 template db first,auto gen poco class

💡 [REQUEST] - 能不能把微调语料的格式改为jsonl（Json Line）

我现在就是这么干的，如果能原生支持更好了

CUDA out of memory when training DPO in parallel on multiple GPUs

V100不支持BF16

CUDA out of memory when training DPO in parallel on multiple GPUs

> 用 lora lora效果不太好感觉，所以...

CUDA out of memory when training DPO in parallel on multiple GPUs

最新进展我使用分布式微调，两台 V100 16G * 8，一样OOM，所以感觉就是batch size=1 都吃不下去

CUDA out of memory when training DPO in parallel on multiple GPUs

> 用 lora 现在更换了lora，如果使用默认的lora_rank=8微调0.5B的模型，调整的参数是： trainable params: 4,399,104 || all params: 498,431,872 || trainable%: 0.8826 请问一般什么比例更好呢？

[BUG] [5.1] : MYSQL : SQL Parsed Error

we need support \uff08 and \uff09

VLLM报错

按文档pip install vllm 0.9.0后执行报错，CUDA 12.2