fix(ggml-sycl): add synchronization before exiting argsort kernel

[MengAiDev]

• Add stream->wait() to ensure all kernels finish execution before proceeding
• This resolves potential race conditions in the argsort operation

Fixes: #15580

[simonlui]

@MengAiDev The closing brace for the function is missing so it fails to compile when I tried to check out the branch. I added an extra line to close it with } and it works.

[NeoZhangJianyu]

#15580 support on iGPU.
Could you check if dGPU has this issue?
if no, maybe add the condition to check the iGPU and add wait() for iGPU only.

It could reduce the protentional risk to dGPU.

[simonlui]

@NeoZhangJianyu I have an Intel Arc A770 16GB and can confirm the issue existed on my dGPU too. This is a snippet from the backtrace I posted in the issue.
/home/simonlui/Code_Repositories/llama-cpp-python/vendor/llama.cpp/ggml/src/ggml-sycl/ggml-sycl.cpp:3380: GGML_ASSERT(row_id_i >= 0 && row_id_i < n_as) failed
Same assert error as iGPU.

[NeoZhangJianyu]

@NeoZhangJianyu I have an Intel Arc A770 16GB and can confirm the issue existed on my dGPU too. This is a snippet from the backtrace I posted in the issue. /home/simonlui/Code_Repositories/llama-cpp-python/vendor/llama.cpp/ggml/src/ggml-sycl/ggml-sycl.cpp:3380: GGML_ASSERT(row_id_i >= 0 && row_id_i < n_as) failed Same assert error as iGPU.

OK! Thank you for your feedback!
It's OK to me!

[MengAiDev]

I have fix the }

[NeoZhangJianyu]

@MengAiDev
Sorry for delayed reply!
I have thought other maintainer will merge this PR.
But some maintainers won't focus on SYCL backend now.

I will continue to support SYCL backend.

I test this PR, it's passed.
But as my experiment, adding wait() will reduce a little performance.
And it will break the SYCL graph feature. (This feature is pending for other issue).

I have created a PR to fix argsort OP too: #16521.
Could it fix the issue #15580?

If the merged PR (#16521) could resolved the issue, I suggest not merging this PR.

How do you think?

Thank you!