Unable to load SYCL compiled backend

[mrtopkid]

Description

I tried compiling a sycl backend for use with intel gpus. By itself, it works fine (usual testing via the outputted main.exe of llama.cpp). However when pointing LlamaSharp to the compiled libs (I set shared libs to 'ON'), it fails to load and provides no further information.

I believe this is because with normal, non-LLamaSharp usage, one needs to run it via the intel one-API CMD line or to call the set vars script of the one-API tool. LlamaSharp won't do this. Is there any workaround for this?

[martindevans]

Are you sure you're using the correct version? Each version of LLamaSharp is compatible with one version of llama.cpp - you must use exactly the right version when compiling the binaries!

See the table at the bottom of the readme for the versions: https://github.com/SciSharp/LLamaSharp?tab=readme-ov-file#map-of-llamasharp-and-llamacpp-versions

[mrtopkid]

Thanks for your reply. Yes I've compiled against the correct commit. The error I receive is:

An error occurred during startup
at LLama.Native.NativeApi.llama_max_devices()
at LLama.Abstractions.TensorSplitsCollection..ctor()
at LLama.Common.ModelParams..ctor(String modelPath)
at PSLLM.CompletionEngine.StartMessageProcessor() in my-app\CompletionEngine.cs:line 266
at PSLLM.CompletionEngine.Start(LLMConfig config) in my-app\CompletionEngine.cs:line 40
at Program.

$(String[] args) in my-app\Program.cs:line 66

I've verified that LlamaSharp is pointing to a valid dll by adding a log callback (the 'Path' parameter shown is correct)

[martindevans]

Is there more to that error dump? I'm hoping for a more specific message at the bottom of the stack trace.

[mrtopkid]

Unfortunately not much, the inner exception is:

LLama.Exceptions.RuntimeError: Failed to load the native library [my-app/bin/Debug/net8.0/CustomBackends/IntelARC/llama.dll] you specified.
at LLama.Native.NativeApi.TryLoadLibraries(LibraryName lib)
at LLama.Native.NativeApi.<>c.b__60_0(String name, Assembly _, Nullable`1 _)
at System.Runtime.InteropServices.NativeLibrary.LoadLibraryCallbackStub(String libraryName, Assembly assembly, Boolean hasDllImportSearchPathFlags, UInt32 dllImportSearchPathFlags)
at LLama.Native.NativeApi.llama_max_devices()
at LLama.Native.NativeApi.llama_empty_call()
at LLama.Native.NativeApi..cctor()

[mrtopkid]

I've hidden the path for anonymity, but other than that it's unchanged

[martindevans]

Ah, unfortunately that's a pretty generic error.

I don't really know anything about SYCL, but from a bit of googling I think your guess about the set-vars script looks likely to be the issue. If you run the script before launching your application does that not fix the issue?

[SanftMonster]

Maybe you could try to use dotnet cli in the intel one-API CMD line? Such as dotnet run xxx.csproj

[gfody]

I ran into this too, I think it's more an issue with llama.cpp than with LLamaSharp. you need to copy all the dependencies to wherever llama.dll is being loaded from - it's these files:

ggml_shared.dll
libmmd.dll
llava_shared.dll
mkl_core.2.dll
mkl_sycl_blas.4.dll
mkl_tbb_thread.2.dll
pi_win_proxy_loader.dll
svml_dispmd.dll
sycl7.dll

..which are all included in the sycl package but then that package doesn't include llama.dll 😂 so you have to build that yourself (if you do that then at least you'll already have all these dependencies on your system!). here's an issue to include the shared libs in the sycl package: ggml-org/llama.cpp#7361

I would love a backend nuget for SYCL!

[mrtopkid]

Hi All,

Thank you for all your input, @gfody I think that was it. Got the built dependencies alongside llama.dll and it seemingly works. The backend loads and I'm able to start inference. However, I have noticed a interesting issue:

Inference is wildly unstable. Unlike when using the cpu backend, the sycl based backend just spits out garbage. Here's a copy of what my Llama 3 8b instruct model responded to me with when I prompted it with "Hello, how are you?":

"I gravel ?? enced pm ["$ Equipment secondary initWithStyle O zcze / assistant equality standby Perkins .Is balances opak I or 378 juices Is shel )data"

And another run:

' I """""""""""""""""""""""""""""""""""""""" '

I believe this is something unique to my sycl backend, though I'm not sure what yet.

@AsakusaRinne & @martindevans, I've added a call to Process to call the setvars.bat of the intel oneAPI tool from within my app before the backend is loaded. I originally thought that it wouldn't work since it'd be run against the context of the child process, but it does? I'm not really sure what happens under the hood here so idk. But I will try and do more testing. Would LlamaSharp accept a SYCL nuget package if we / I am able to create one? Apologies, I'm new here (and to collaborative open source dev).

Thank you again

[mrtopkid]

Scratch that, a different model file works. Unsure why that particular model runs fine on cpu / nvidia but fails with sycl.

[SanftMonster]

@gfody @mrtopkid Thank you for your efforts here. :)

I would love a backend nuget for SYCL!
Would LlamaSharp accept a SYCL nuget package if we / I am able to create one?

We certainly would not reject such a PR as long as it could run stably. I'll appreciate it if you'd like to contribute. If you would like to add such a backend, here's some tips for you:

• Please add a new .nuspec file like what was done in Create nuspec for OpenCL #489
• Please include both the windows and linux library files.
• I'm not sure whether the files @gfody mentioned are always requires to be in the same folder with llama.dll, or setting env var is enough. If the latest matches, please split the sycl backend to LLamaSharp.Backend.SYCL and LLamaSharp.Backend.SYCL.Deps because maybe some users only want the sycl-version llama.dll.

the sycl based backend just spits out garbage...

Could you please test the same model with llama.cpp main.exe compiled with SYCL? I'm not sure if its a LLamaSharp issue.
Note that some gguf files on hyuggingface are outdated. For example, a gguf file generated by llama.cpp 3 months ago probably cannot work well with the latest llama.cpp version. LLamaSharp has just completed a binary updating 2 weeks ago. So please avoid using gguf files uploaded several months ago.

[gfody]

Scratch that, a different model file works. Unsure why that particular model runs fine on cpu / nvidia but fails with sycl.

@mrtopkid I've noticed this as well increasing ContextSize and GpuLayerCount helps (I'm using 4096, and 80 for llava) - I think that Intel ARCs are just not getting as much testing considering all the hoops we had to jump through, maybe that will change once there's a LLamaSharp.Backend.SYCL nuget!

@AsakusaRinne I've tested this and you can indeed just put the sycl dependencies in your path, they don't necessarily have to be in the same location as llama.dll. presumably though maybe this could cause some unintended consequences for whatever reason why the oneAPI Toolkit doesn't just put these in our path and why also llama.cpp bundle these with their sycl release package - I don't know

[github-actions]

This issue has been automatically marked as stale due to inactivity. If no further activity occurs, it will be closed in 7 days.