Docker image update needed to support Kernel 6.18

[plumlis]

After upgrading the host kernel to Linux 6.18, Ollama-based inference using ipex-llm-inference-cpp-xpu docker (oneAPI / SYCL backend) on Intel Xe (Arc) GPUs fails to function correctly.

The same container and model configuration works as expected on an older kernel（6.17.12)

The issue appears to be related to GPU memory detection / allocation via Level Zero after the kernel upgrade.

Additional

1.The issue does not occur on older kernels with the same container image

2.No changes were made to Docker image, Ollama version, or model

3.This suggests a regression or behavior change in kernel 6.18 affecting:

here are error logs from ollama in docker.

get_memory_info: [warning] ext_intel_free_memory is not supported (export/set ZES_ENABLE_SYSMAN=1 to support), use total memory as free memorytime=2025-12-14T00:04:07.339+08:00 level=INFO source=server.go:652 msg="waiting for server to become available" status="llm server loading model"

[GIN] 2025/12/14 - 00:04:06 | 200 |      23.915µs |       127.0.0.1 | HEAD     "/"
[GIN] 2025/12/14 - 00:04:06 | 200 |  125.603561ms |       127.0.0.1 | POST     "/api/show"
Native API failed. Native API returns: 39 (UR_RESULT_ERROR_OUT_OF_DEVICE_MEMORY)
Exception caught at file:/home/runner/_work/llm.cpp/llm.cpp/ollama-llama-cpp/ggml/src/ggml-sycl/ggml-sycl.cpp, line:405, func:operator()
SYCL error: CHECK_TRY_ERROR(ctx->stream->memset( (char *)tensor->data + original_size, 0, padded_size - original_size).wait()): Exception caught in this line of code.
  in function ggml_backend_sycl_buffer_init_tensor at /home/runner/_work/llm.cpp/llm.cpp/ollama-llama-cpp/ggml/src/ggml-sycl/ggml-sycl.cpp:405
/home/runner/_work/llm.cpp/llm.cpp/ollama-llama-cpp/ggml/src/ggml-sycl/../ggml-sycl/common.hpp:115: SYCL error

It appears that GPU memory cannot be correctly queried after upgrading to Linux kernel 6.18, while the same setup works normally on kernel 6.17; upgrading to the latest Level Zero runtime and using newer Ollama versions does not resolve the issue, suggesting a possible regression or change in Intel Xe GPU memory management introduced in kernel 6.18.

[plumlis]

pytorch/pytorch#157989

[plumlis]

I’ve gone through a series of tests, including installing a newer kernel (6.19 RC1), building my own Docker image, and using the latest Intel oneAPI runtimes and related components. So far, the situation remains the same: I’m unable to retrieve the Intel Arc GPU memory information.

My personal guess is that this is very likely related to some changes in the Intel Xe driver introduced around the 6.18 kernel series. After rolling back to kernel 6.17.12, even though there are many warnings and errors when running larger models, it does at least work.

At this point, I’m not sure how this issue should be resolved, or even where it should be reported. I also don’t know whether this is a bug, or whether there were API changes on the Xe side that now require matching updates in the user-space drivers. Due to time constraints, I’ve decided to stick with kernel 6.17.12 for now.

[plumlis]

From intel/compute-runtime#875
It seems docker image need update to new version runtime builtin to support kernel 6.18

[Wildcat1]

@plumlis, seeing the same issue as you. Would be great to get this fixed because it impacts NAS appliances running docker containers as well. In that case you're limited to the kernel that is part of the NAS FW so no opportunity to downgrade the kernel etc.

Hopefully a fix will be available for this soon

[plumlis]

@plumlis, seeing the same issue as you. Would be great to get this fixed because it impacts NAS appliances running docker containers as well. In that case you're limited to the kernel that is part of the NAS FW so no opportunity to downgrade the kernel etc.

Hopefully a fix will be available for this soon

It's a kernel side issue:
intel/compute-runtime#875

Only thing we can do is wait.
Hope intel fix it.

[aaricantto]

I'm getting a very similar error on Windows (portable ollama, on Two ARC B60s, single GPU inference does not throw error "ONEAPI_DEVICE_SELECTOR=level_zero:0")

Now to preface, I can run "ollama run modelX" and i'll see the model load into both GPU's VRAM split evenly (as expected), but as soon as I submit a prompt I get this error below which is very similar to yours.

I'm not great at C, but It seems as if I'm at least able to get past tensor allocation unlike you.

Native API failed. Native API returns: 39 (UR_RESULT_ERROR_OUT_OF_DEVICE_MEMORY)
Exception caught at file:D:\actions-runner\release-cpp-oneapi_2024_2\_work\llm.cpp\llm.cpp\ollama-llama-cpp\ggml\src\ggml-sycl\ggml-sycl.cpp, line:4602, func:operator()
SYCL error: CHECK_TRY_ERROR((stream)->memcpy( data, (const char *)tensor->data + offset, size).wait()): Exception caught in this line of code.
  in function ggml_backend_sycl_get_tensor_async at D:\actions-runner\release-cpp-oneapi_2024_2\_work\llm.cpp\llm.cpp\ollama-llama-cpp\ggml\src\ggml-sycl\ggml-sycl.cpp:4602
D:\actions-runner\release-cpp-oneapi_2024_2\_work\llm.cpp\llm.cpp\ollama-llama-cpp\ggml\src\ggml-sycl\..\ggml-sycl\common.hpp:115: SYCL error
time=2026-01-09T16:26:23.258+10:00 level=ERROR source=server.go:827 msg="post predict" error="Post \"http://127.0.0.1:50346/completion\": read tcp 127.0.0.1:50348->127.0.0.1:50346: wsarecv: An existing connection was forcibly closed by the remote host."

Hope this might help

[plumlis]

https://gitlab.freedesktop.org/drm/xe/kernel/-/issues/6948