The IPEX-LLM packaged Ollama (v2.3.0-nightly build 20250725 for Ubuntu, from ollama-ipex-llm-2.3.0b20250725-ubuntu.tgz) crashes with SIGABRT due to an assertion failure in sdp_xmx_kernel.cpp when attempting to load or run Llama 3.1 models (e.g., llama3.1:8b, llama3.1:8b-instruct-q5_K_M). This occurs on an Intel Arc B50 Pro GPU with current drivers. Other models like gemma2:9b-instruct-q5_K_M work correctly with GPU acceleration on the same setup.

How to reproduce

Assuming a working Ubuntu system with appropriate Intel GPU drivers and the extracted ollama-ipex-llm-2.3.0b20250725-ubuntu package:

Set the required environment variables:

Bash
export OLLAMA_LLM_LIBRARY=$(pwd)/llm_c_intel
export LD_LIBRARY_PATH=$(pwd)/llm_c_intel/lib:${LD_LIBRARY_PATH}
export ZES_ENABLE_SYSMAN=1
Start the Ollama server in the background: ./ollama serve &
Attempt to run a Llama 3.1 model: ./ollama run llama3.1:8b "Test"

Observe the server process crashing with the SIGABRT signal and the assertion failure mentioned above in its logs.

Screenshots
N/A - Relevant log output below.

Environment information

GPU: Intel Arc B50 Pro
OS: Ubuntu 24.04.3 LTS (Noble Numbat)
Kernel: 6.14.0-33-generic #33 24.04.1-Ubuntu
GPU Drivers (from ppa:kobuk-team/intel-graphics):
intel-opencl-icd: 25.35.35096.9-124.04ppa3
libze-intel-gpu1: 25.35.35096.9-124.04ppa3
libze1: 1.24.1-124.04ppa1

IPEX-LLM Ollama Version: v2.3.0-nightly (Build 20250725 from ollama-ipex-llm-2.3.0b20250725-ubuntu.tgz)

Additional context
The model gemma2:9b-instruct-q5_K_M works correctly.

Key Log Output during Crash:

[...]
ollama-bin: /home/runner/_work/llm.cpp/llm.cpp/llm.cpp/bigdl-core-xe/llama_backend/sdp_xmx_kernel.cpp:439: auto ggml_sycl_op_sdp_xmx_casual(...)::(anonymous class)::operator()() const: Assertion `false' failed.
SIGABRT: abort
PC=0x742c8f49eb2c m=3 sigcode=18446744073709551610
signal arrived during cgo execution
[...] (Goroutine stack trace follows)