xtuner
xtuner copied to clipboard
InternLM
An efficient, flexible and full-featured toolkit for fine-tuning LLM (InternLM2, Llama3, Phi3, Qwen, Mistral, ...)
``` xtuner pth_to_hf $CONFIG $PTH $SAVE_PATH --save-format [xtuner, official, huggingface] ``` More details can be found on https://github.com/LZHgrla/xtuner/blob/a1a0ce8f7f47d1e1079d9bc17876980225dbb4df/xtuner/configs/llava/llama3_8b_instruct_clip_vit_large_p14_336/README.md#model-conversion
环境：win11，单卡 RTX 4070 (12G) 按照 Tutorial 操作 [Tutorial](https://github.com/InternLM/Tutorial/tree/main)/[xtuner](https://github.com/InternLM/Tutorial/tree/main/xtuner) /README.md 2.3.6 将得到的 PTH 模型转换为 HuggingFace 模型 进行到这一步时，出错如下： ``` (xtuner) λ xtuner convert pth_to_hf ./internlm_chat_7b_qlora_oasst1_e3_copy.py ./work_dirs/internlm_cha t_7b_qlora_oasst1_e3_copy/iter_6501.pth ./hf quantization_config convert to `low_cpu_mem_usage`...
I have never been able to correctly quantify awq for llava-llama3 in the official format of llava。 Can anyone help me?
When could we expect to see support for the InternLM-XComposer family of models? Thanks!
I have successfully SFTed Qwen2-0.5B using XTuner. When I convert the qlora checkpoint to hf, `weight is on the meta device, we need a `value` to put it in 0`...
model: internlm2.5 基于官方文档的多轮对话格式，训练一批次数据，但单个messages，就有上千个json，数据大小达到了60K，两个conversation就能达到129k，有如下几个问题： 1. 基于chat模型训练还是base模型比较好一些？ 2. 如果基于base模型，那么微调训练后，能有chat模型的效果么？是不是还需要再做什么处理？ 3. GPU为96GB的前提下，xtuner能支持两轮129K的conversation训练吗？或者单轮的conversation，可以实现微调训练吗？主要是这个长对话的微调效果。
