likelovewant / ROCmLibs-for-gfx1103-AMD780M-APU

ROCm Library Files for gfx1103 and update with others arches based on AMD GPUs for use in Windows.
GNU General Public License v3.0
147 stars 17 forks source link

can it be used to patch rocm hip library 5.7? #2

Closed lcweik closed 7 months ago

lcweik commented 8 months ago

I have tried in llama.cpp with hip 5.7, but i found that gpu mode is much slower than cpu mode, so is it any thing wrong?

这是哪个版本的rocm的补丁,5.7能用吗?

likelovewant commented 8 months ago

it is for 5.7 .The orignal code is here https://github.com/ROCm/rocBLAS , Maybe you could try to use https://lmstudio.ai/rocm for llama

windows 5.7版本, 运算逻辑和gfx1101 一样,只是修改了相关信息,设备号等 方便适配 gfx 1103,可以理解成 这本来就是 官方 未发布的 hip library

lcweik commented 8 months ago

it is for 5.7 .The orignal code is here https://github.com/ROCm/rocBLAS , Maybe you could try to use https://lmstudio.ai/rocm for llama

windows 5.7版本, 运算逻辑和gfx1101 一样,只是修改了相关信息,设备号等 方便适配 gfx 1103,可以理解成 这本来就是 官方 未发布的 hip library

大佬,看一下,我用LM STUDIO加载QWEN最小的模型,出现这个错误啊,我的电脑是yoga air14s 7840s "(Exit code: -36861). Unknown error. Try a different model and/or config."

{ "memory": { "ram_capacity": "27.82 GB", "ram_unused": "19.98 GB" }, "gpu": { "type": "AmdROCm", "vram_recommended_capacity": "42.73 GB", "vram_unused": "14.11 GB" }, "os": { "platform": "win32", "version": "10.0.22631", "supports_avx2": true }, "app": { "version": "0.2.16", "downloadsDir": "D:\code\ai\model" }, "model": {} }

likelovewant commented 8 months ago

我也试过会报错,然后把 rocm library 替换为自己的,roclabs.dll 替换掉。再试试。他这里有个单独的rocm ,和系统那个是独立的。 实在不行,尝试不用rocm 的版本,就是cpu 版,且cpu 版可以顺利加载,也可以使用显卡加速。然后不死心的话再切换回rocm版就可以加载了。不过不是很理想。需要在作者的discord 里面反应一下

likelovewant commented 8 months ago

所有的model 不可以直接扔进去model 文件里面,会不识别,可以试试在models文件下创建TheBloke/ Llama-2-7B-Chat-GGUF/文件夹,然后在这里面放置模型 需要是gguf格式,通义的没有测试过