superhg

论文里面描述的mel 4倍的下采样是在bottle net feature extractor里面做的，然而，代码里面在bnf-prenet里面通过两层卷积(stride=2)也做了4倍下采样，这边不是很理解

@mymusise 帮忙看一下

> prompt_ids[-2] = 150001 > prompt_ids[-1] = 150004 谢谢！！

> I have tested the effect of converting Chinese to Chinese, and it sounds very good, but I have not tested the mixing of Chinese and English ASR模型在中文上建模使用的是音素还是字建模单元？我用的2000小时字建模，转换有很多字发音的Badcase @madosma

met same issue: 蚂蟥 的 “蟥 ” 椪柑树 的 “椪” ，token id会被转换为0

> Can you please share the command you ran? `python -m torch.distributed.launch --nproc_per_node 8 --nnodes 2 --node_rank 0 --master_addr 10.19.102.26 --master_port 29500 train_gpt.py --config ./configs/gpt2_small_zero3_pp1d.py --from_torch $DUMMY_DATA`

跟了这个问题很久，目前能跑起来mixtral 8x7B模型full sft的框架只有LLama2-Accessory这个库

CUDA_VISIBLE_DEVICES=gpu_id python your_script.py

export CUDA_VISIBLE_DEVICES=0 deepspeed --num_gpus=1 xxxx.py