MaggotHATE / Llama_chat

A chat UI for Llama.cpp
MIT License
10 stars 0 forks source link

Crash upon more than character inserted into `Prefix` field #3

Closed Succubyss closed 1 week ago

Succubyss commented 1 month ago

I'm using Llama_Chat_gguf_VK_vulkan_mini.exe (which I believe is what most NVIDIA GPU users should be using?) and on completely default settings, I'll go into "Text Settings" and type into this field:

image

But upon inserting more than 1 character, the program crashes.

P.S. Please make Shift+Enter always insert a newline irrespective of the toggle status of Send messages by Enter (should probably be changed to "with Enter").

MaggotHATE commented 1 month ago

Hi! Thanks for the report, seems like Prefix is broken again, I'll look into it. Haven't used prefix for a long time.

As for Shift+Enter, I'm not sure it's possible in ImgUI (should be?), but it would be handy indeed.

MaggotHATE commented 1 month ago

Oh, and btw:

which I believe is what most NVIDIA GPU users should be using?

Both Clblast and Vulkan versions are compatible with Nvidia GPUs, just be aware that Clblast doesn't work with MoE models.

Succubyss commented 1 month ago

Oh, and btw:

which I believe is what most NVIDIA GPU users should be using?

Both Clblast and Vulkan versions are compatible with Nvidia GPUs, just be aware that Clblast doesn't work with MoE models.

Which do you think would be faster? Pascal architecture if that makes a difference.

MaggotHATE commented 1 month ago

Pascal architecture if that makes a difference.

I see not much difference on my 1060 3GB (Vulkan is slightly faster, but each layer takes more VRAM), but that amount of VRAM is not enough to test that properly. However, Vulkan takes additional time to compile shaders on each launch, so it's a bit less convenient - and I don't use MoE models for now.

I've fixed the problem you've reported, please check. Shift + Enter should also work.