Bug: Models not working any more when using SYCL backend

[michael72]

Name and Version

version: 6717 (b260213)
built with Intel(R) oneAPI DPC++/C++ Compiler 2025.3.1 (2025.3.1.20251023) for x86_64-unknown-linux-gnu

Operating systems

Linux

Which llama.cpp modules do you know to be affected?

No response

Command line

./llama-server -m ${MODEL_FILE}

Problem description & steps to reproduce

Since the following change I have problems running gguf models (any model I tested - mostly qwen models) with SYCL backend.

b260213#diff-3b2a242ed9135b23435e9ac1c2302c321017416eb545c294ca4108ad6dab7079L4389

I have compiled the project using the script examples/sycl/build.sh
I am loading a model with it (e.g. Qwen2.5.1-Coder-7B-Instruct-Q8_0.gguf) and running a request.
Generally the output starts and after a few tokens it either stops or there is garbage added - like "import DebugCall09090909090..." with a lot of repeated characters.

I didn't have those problems before the change above.
SYCL Compiler version: 20250301

First Bad Commit

b260213

Relevant log output

[NeoZhangJianyu]

@michael72
Let me check it!
Could you share the full log including the cmd line?

Thank you for the feedback!

[michael72]

logoutput.txt

the command call was
./llama-server -m /home/michael/Downloads/Qwen2.5.1-Coder-7B-Instruct-Q8_0.gguf --port 8222

OK - I'm pretty sure my request would produce some garbage, but there was only ????????? (and so on) coming out of it.
I actually just downloaded this file:
https://github.com/ggml-org/llama.cpp/blob/master/ggml/src/ggml-sycl/common.cpp
uploaded it to the llama chat interface and gave the following instructions (stripped down version of how I normally would write tests - not in cpp though):

Hey - could you write test for me?

• try to test all code lines in the class under test except code parts excluded from coverage (between $COVERAGE-OFF$ and $COVERAEG-ON$)
• give the tests just as chat output, don't edit any file.
• the name of the test suite is where is the name of the class under test

it starts "Processing" and then the ????????? appear

@NeoZhangJianyu Thanks for looking into this!

[NeoZhangJianyu]

@michael72
Thank you for your sharing!
I think you are using the iGPU in 11th Core CPU. Is it right?
I have one too. :)

I remember b260213#diff-3b2a242ed9135b23435e9ac1c2302c321017416eb545c294ca4108ad6dab7079L4389 will lead to some UT cases are fault on iGPU.
I guess less body use the iGPU. So I ignore it.

Now, I know you are using it.
Let me fix it.

[michael72]

@NeoZhangJianyu OK I thought SYCL backend was the way to go when having iGPU.
Yes - let's say my hardware setup is not a perfect use case - but if you're able to fix this without affecting performance of other devices that would be cool. Otherwise I will keep using the b6715 version until I have a new hardware.

The actual hardware is:
GPU:
Intel Iris Xe Graphics G7 96EUs, Device ID 9A49
CPU:
i7-1260P
on a Dell XPS 13 Plus 9320 laptop

[NeoZhangJianyu]

@michael72
OK, I will try to fix it without side effect to other hardware.

Thank you!