After updating from 7.3.4 to 7.4.0, GPU compute on my Intel Arc B580 no longer works in Docker. Two different stacks that ran fine on 7.3.4 now fail with Level Zero allocation errors. The kernel log shows no errors. I am reporting it in case others are affected and to ask whether anything in the 6.18.58 -> 6.18.60 kernel or the 7.4.0 image could explain it.

SYSTEM

- Unraid 7.4.0 (kernel 6.18.60-Unraid, confirmed with uname -r). Previously 7.3.4 (6.18.58-Unraid), updated 2026-10-24.

- GPUs: Intel Arc B580 12 GB (xe driver, 0d:00.0, /dev/dri/renderD128 + card0) used for compute; Intel Arc A310 (i915, 06:00.0) used for transcoding.

- 78.5 GiB RAM (about 47 GiB free at the time of the tests).

- Containers get /dev/dri/renderD128 and /dev/dri/card0. The test containers were created with plain "docker run" (no memory limit).

WORKED ON 7.3.4 (kernel 6.18.58)

1) Ollama 0.9.3 (IPEX-LLM build, uberchuckie/ollama-intel-gpu:stable), qwen3:8b, context 65536:

   4 benchmark runs on 2026-10-25: 55.88, 61.18, 58.29, 58.57 tok/s (average 58.5). About 11.5 GB of the card's 12.8 GB in use at 64K; 9.09 GB at 16K.

2) Upstream llama.cpp SYCL (ghcr.io/ggml-org/llama.cpp:server-intel, oneAPI 2026.1), Qwen3.6-35B-A3B IQ3_XXS:

   -c 65536 -np 1 -ngl 99 --n-cpu-moe 10 -fa on -ctk q8_0 -ctv q8_0

   Started healthy in about 80 s, 11.5 GB card usage at 64K, 28-39 tok/s generation depending on context. Restarted several times on 2026-11-02 to 2026-11-04 without problems.

Context: 7.3.3 (kernel 6.18.54) had made the same workloads about 60% slower (24.3 vs 64.1 tok/s); 7.3.4 recovered them. So this card/stack has been sensitive to 6.18.x point releases before.

ON 7.4.0 (2026-11-14, kernel 6.18.60)

1) Ollama (IPEX) cannot load any model, at any context size (16384, 32768, 49152, 65536 all fail identically). A container restart does not help.

   ggml_backend_sycl_buffer_type_alloc_buffer: can't allocate 850335744 Bytes of memory on device

   alloc_tensor_range: failed to allocate SYCL0 buffer of size 850335744

   llama_model_load: error loading model: unable to allocate SYCL0 buffer

   API reply: {"error":"llama runner process has terminated: error loading model: unable to allocate SYCL0 buffer ..."}

2) llama.cpp (same image and flags as above) loads for about 4 minutes (was about 80 s), then aborts in the first decode:

   terminate called after throwing an instance of 'sycl::_V1::exception'

   what():  level_zero backend failed with error: 38 (UR_RESULT_ERROR_OUT_OF_HOST_MEMORY)

   (raised from a kernel submission, ggml_sycl_op_rms_norm_fused -> sycl::handler::finalize)

CHECKED/RULED OUT

- Not a capacity problem: /sys/kernel/debug/dri/0000:0d:00.0/vram0_mm shows size 12809404416, usage 481116160 (idle), visible_avail 11757MiB; a 16K context fails exactly like 64K.

- Not one runtime: two separate userland stacks (old IPEX Ollama and upstream llama.cpp with oneAPI 2026.1) fail differently but both at allocation level.

- dmesg: no xe error, GuC reset or wedged message after boot. Boot lines look normal (VRAM usable 0x2fb800000, GuC bmg_guc_70.bin 70.72.1, HuC 8.2.10), plus "xe 0000:0d:00.0: [drm] Can't init xe mei late bind missing mei component". I do not know whether that line was also present on 7.3.4 (I did not save that boot log).

- The 7.4.0 release notes list kernel 6.18.60 and no Intel GPU / xe / Docker engine change, and no new known issues.

NOT KNOWN

- Root cause, and whether it is the kernel, something else in the 7.4.0 image, or an interaction with the userland runtimes in the containers.

- Whether the A310 (transcoding) or other GPU containers are affected. I have not tested.

Is anyone else on 7.4.0 with an Arc (Alchemist/Battlemage) card seeing compute failures in containers (Ollama, llama.cpp, OpenVINO, etc.)?

Did anything in the xe driver, its configuration or module parameters change between 6.18.58 and 6.18.60 in the Unraid build?

UPDATE: Rolled back to 7.3.4,

same container, same model. Kernel is back to 6.18.58-Unraid and GPU inference works

again: qwen3:8b at 64K context, 4 runs after rollback: 64.14, 61.49, 63.05, 63.11 tok/s

(my earlier 7.3.4 results were 55.9–61.2 tok/s). On 7.4.0 (kernel 6.18.60) the same

test fails at every context size from 16K to 64K with an allocation error. So the

regression reproduces with the OS version as the only change.

Also checked on 7.4.0 before rolling back: boot files verified against their .sha256

files, xe firmware (bmg_guc_70.bin, bmg_huc.bin) byte-identical to 7.3.4's, no extra

kernel parameters or xe/i915 modprobe overrides.

I'm holding at 7.3.4 until I hear some sort of update regarding this breaking change gets resolved in future release.