Yuxuan Zhang

这个应该是你环境的问题吧，我们不会用到.c文件和python.h 估计是你环境安装的问题

目前的显存用peft调不动VIT混合的部分，显存会炸，所以我们有规划放出SAT的微调版本，能够微调全部参数

因为这个模型的微调脚本时数据并行的，没有办法做zero3或者tensor并行（目前的forward不支持），所以48G就是48G，多少x48G还是48G，因为模型也复制了那么多个载入在不同的卡

You cannot directly use the conversation template; you need to process the input according to our demo. You seem to be missing a step in the dtype conversion process. Please...

ollama应该没有适配，我们和他们官方团队正在联系～

If you are doing full parameter fine-tuning, you can directly follow the default conversion script, here https://github.com/THUDM/CogVideo/blob/main/tools/convert_weight_sat2hf.py

你应该更新现在的最新的vLLM源代码并应用 https://github.com/vllm-project/vllm/pull/16618 丢弃 --model-impl transformers 因为 https://github.com/vllm-project/vllm/pull/16618 已经修复了vLLM支持问题

暂时不确定，这个PR不是由我提起，需要原始作者满足vLLM的合并规范，vLLM官方会进行合并操作～，感谢理解，同时，感谢相关社区作者的热情贡献

I believe this is a good start. If you can provide detailed explanations for each step and make it runnable on T4 (which seems to be free) devices, we would...

This method cannot run on T4; it is designed for H100, and the fP8 format can only run properly on H100. You should try adding these two lines: ``` pipe.enable_sequential_cpu_offload()...