ningshanwutuobang

Here is one output of `llava.py`  The result is not correct, but i think the process is correct.

Here is one output of panda_gpt.py 

Here is one output of `minigpt4.py`  And the full log is [minigpt4.log](https://github.com/ggerganov/llama.cpp/files/11860166/minigpt4.log)

@ggerganov Thanks for your comments. I have update the function of `llama_eval_embd`. Can you check if it is correct? I am not sure if the ftype and the layout tensor...

@monatis Nice work for the clip.cpp. It seems to be a large model, I think it would is better to be a standalone project for the llava.cpp. The main point...

The penalty is used to avoid some bad output such as repeating nonsense. It is not implemented in that function.

I think it is possible to add interface to input `inpL` in https://github.com/ggerganov/llama.cpp/blob/ffb06a345e3a9e30d39aaa5b46a23201a74be6de/llama.cpp#L1255 directly.

@prcamp @AsakusaRinne I have tried to add such interface in https://github.com/ningshanwutuobang/llama.cpp/blob/embd_inp/examples/embd_input/embd_input_test.cpp The process look like follow: 1. create context. 2. ~~quantize~~ use the float tensor to replace `inpL` (which need...

iOS 不是有xp3player吗？我没iOS的开发环境。

这个代码是不全的，只能运行一些简单的。 一些重要的插件是没有的，比如emote的那些。