Bug: Intel Arc - not working at all #9106
Comments
Eugeniusz-Gienek
added
bug-unconfirmed
critical severity
|
Additionally: And |
|
...and just in case: |
|
Run |
|
@Eugeniusz-Giene |
|
@ClarkChin08 please give a review |
ngxson
added
the
SYCL
|
I am having this exact same issue, on Gentoo, with the exact same LLVM spirv library. I built llama.cpp according to the SYCL docs, using the example build script, which builds with FP32 by default. So I have this issue when using FP32. |
Hi， you use '-sm none -mg 0'， This will keep the model running on the GPU-0 and the memory is only 6064M, Please change the command to '-sm layer' instead of '-sm none -mg 0' , this will split the graph to two GPUs. |
|
@Eugeniusz-Gienek |
Oh that's a very important information which I would assume belongs to SYCL installation manual... Thanks for sharing! |
|
Hmm, I have a 10th gen CPU. That might be why mine does not work, either. |
|
To be honest I have bought specifically Intel Arc A380 with intent to use on a HP Prodesk 600 G4 SFF as there are not a lot of options out there for SFF format. If I knew in advance it won't work I wouldn't buy it. Anyways, I've managed to return it to seller and out of curiosity checked if the onboard Intel GPU would work at least somehow. The only way for it to work was to use Vulkan, which on that GPU is TERRIBLY slow. So, to sum up - nope, you probably not want to use any Intel GPU if Your CPU is not Intel 11th gen+. Ah, and actually I've switched to another PC for llama.cpp with Intel 10th gen and NVidia 3060, which works great with GPT4ALL but to my surprise didn't work that good with llama.cpp as it appears to be that if one wants to run a webserver the RAM requrement is also a thing, not only VRam - thus for a 8Gb llama model my 16Gb is .... not enough. Anyways, now it is a waiting game for additional RAM to appear and maybe finally the local GPT webserver will be functional in the network. |
|
It's limited by oneAPI/SYCL support in fact. Some motherboard of 10th Core support Arc GPU by support resize bar feature. |
|
I would propose to include this information in the manual - in order for potential new users to not be frustrated and purchase the hardware which won't fulfill the expectations... What do You think? Does it make sense? |
|
I suggest you might consider using the Vulkan version of llama.cpp, as it is compatible with almost all graphics cards.Intel's GPU software & driver is bad |
This is exactly what I did to be honest but that implementation is extremely slow... |
Yes. resize bar feature could be detected in BIOS. If your PC with 10th Core CPU has resize bar, it could support Arc GPU. |
|
i7 8700K on GIGABYTE Z370d motherboard by enable "Above 4G Decoding" in BOIS. |
|
technically resizable BARs have been supported since PCIe 2.0, you can most likely enable them by modifying the DSDTs/DXE of your UEFI BIOS, there is a framework for that here: https://github.com/xCuri0/ReBarUEFI if you do have Above 4G Decoding as an option, the linux kernel will attempt itself to resize the BARs when loading the drivers, although I'm not sure if the ARC drivers do that. amdgpu drivers do however. |
|
@NeoZhangJianyu The command that works properly: And I got this error after I specified -ngl option: llama.cpp: build: 3848 (c919d5d) My A770 can run ComfyUI flawlessly in the meantime, so it doesn't look like a oneAPI compatibility issue. |
|
@NineMeowICT Additional, does it work well with llama2-7b-q4 as the example in SYCL.md? |
|
@NeoZhangJianyu build cmd: ./examples/sycl/build.sh But it can pass run-llama2.sh test. |
|
@NineMeowICT |
|
@NineMeowICT Seems like #9612 (comment) |
|
@NineMeowICT |
|
@Eugeniusz-Gienek |
|
Thank You, I will test it hopefully today after work :) |
|
@NeoZhangJianyu Thank you. I just built it and tested several models and no errors occurred. |
|
This issue was closed because it has been inactive for 14 days since being marked as stale. |
What happened?
Going through the manual - SYCL I mean. Everything compiles okay. Running it always thows an error. Can't make it work. OS used: Linux Gentoo. P.S. docker doesn't work either. P.P.S. device IS listed in the list.
Name and Version
./build/bin/llama-cli --version
version: 3609 (2f3c146)
built with Intel(R) oneAPI DPC++/C++ Compiler 2024.2.1 (2024.2.1.20240711) for x86_64-unknown-linux-gnu
What operating system are you seeing the problem on?
Linux
Relevant log output
The text was updated successfully, but these errors were encountered: