ollama ipex-llm crashes when running tinyllama llama 3.1b and qwen 2.5

[funky-gibbon]

Describe the bug
A clear and concise description of what the bug or error is.

How to reproduce
Steps to reproduce the error:

1. ...
2. ...
3. ...
4. ...

Screenshots
If applicable, add screenshots to help explain the problem

Environment information
If possible, please attach the output of the environment check script, using:

• https://github.com/intel/ipex-llm/blob/main/python/llm/scripts/env-check.bat, or
• https://github.com/intel/ipex-llm/blob/main/python/llm/scripts/env-check.sh

Additional context
Add any other context about the problem here.

[funky-gibbon]

1. Problem Description:

When attempting to run any model (e.g., tinyllama, llama3.1, qwen2.5:7b) using the ollama-ipex-llm optimized builds on an Intel Arc A770 discrete GPU, the llama runner process consistently terminates with a SYCL error during the ggml_sycl_op_mul_mat operation. intel_gpu_top shows blitter activity, indicating data transfer to the GPU, but no significant "Render/3D" or "Compute" utilization occurs before the crash. This issue persists across different IPEX-LLM build versions and various models.

2. Steps to Reproduce:

Operating System: Clean installation of Ubuntu 24.04 LTS (Noble Numbat).

Kernel Version: 6.11.0-26-generic.

Intel GPU Driver Installation:

intel-opencl-icd was installed via sudo apt install intel-opencl-icd.
Latest discrete GPU drivers (25.18.33578.6) were manually downloaded from Intel's website and installed.
IPEX-LLM Ollama Setup (Primary Test: Build 2.3.0b20250612):

Downloaded ollama-ipex-llm-2.3.0b20250612-ubuntu.tgz.
Extracted to ~/ollama_ipex/ollama-ipex-llm-2.3.0b20250612-ubuntu/.
Created /usr/local/bin/start_ollama_ipex.sh wrapper script and /etc/systemd/system/ollama.service systemd unit file.
The start_ollama_ipex.sh script included all export variables from the IPEX-LLM's start-ollama.sh (e.g., OLLAMA_NUM_GPU=999, ZES_ENABLE_SYSMAN=1, SYCL_CACHE_PERSISTENT=1, OLLAMA_KEEP_ALIVE=10m).
SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1 was tested (uncommented, then re-commented), but showed no change in behavior.
The systemd service was configured to run as user michael-fothergill and point to /usr/local/bin/start_ollama_ipex.sh.
Model Run Attempts (Build 2.3.0b20250612):

ollama pull tinyllama
ollama run tinyllama "Hello, what is your purpose?"
Result: SYCL error termination.
ollama pull llama3.1
ollama run llama3.1 "Write a short, evocative poem..."
Result: SYCL error termination.
ollama pull qwen2.5:7b
ollama run qwen2.5:7b "What is the capital of France?"
Result: SYCL error termination.
IPEX-LLM Ollama Setup (Secondary Test: Older Build 2.3.0b20250429):

Downloaded ollama-ipex-llm-2.3.0b20250429-ubuntu.tgz.
Extracted to ~/ollama_ipex_older/ollama-ipex-llm-2.3.0b20250429-ubuntu/.
Updated start_ollama_ipex.sh to point to the ollama binary in this older directory.
SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1 was kept commented out for this test.
sudo systemctl daemon-reload && sudo systemctl restart ollama.
Model Run Attempt (Older Build 2.3.0b20250429):

ollama run tinyllama "Hello, what is your purpose?"
Result: SYCL error termination (same as newer build).
3. Expected Behavior:

The IPEX-LLM optimized Ollama should leverage the Intel Arc A770 discrete GPU for inference, resulting in high "Render/3D" or "Compute" utilization in intel_gpu_top and successful text generation.

4. Actual Behavior:

The llama runner process consistently terminates with the following SYCL error during matrix multiplication (ggml_sycl_op_mul_mat), regardless of the model size (tinyllama, llama3.1, qwen2.5:7b) or IPEX-LLM build version tested:

Error: llama runner process has terminated: error:CHECK_TRY_ERROR(op(ctx, src0, src1, dst, src0_dd_i, src1_ddf_i, src1_ddq_i, dst_dd_i, dev[i].row_low, dev[i].row_high, src1_ncols, src1_padded_col_size, stream)): Meet error in this line code!
in function ggml_sycl_op_mul_mat at /home/runner/_work/llm.cpp/llm.cpp/ollama-llama-cpp/ggml/src/ggml-sycl/ggml-sycl.cpp:3260
/home/runner/_work/llm.cpp/llm.cpp/ollama-llama-cpp/ggml/src/ggml-sycl/../ggml-sycl/common.hpp:117: SYCL error
(Note: The line number may vary slightly, e.g., 3277 vs 3260, but the core error remains the same.)

intel_gpu_top shows blitter activity (data transfer) but no significant "Render/3D" or "Compute" engine utilization before the crash.

5. System Information:

GPU: Intel Arc A770 (Desktop)
Operating System: Ubuntu 24.04 LTS (Noble Numbat)
Kernel Version: 6.11.0-26-generic
Intel GPU Driver Version: 25.18.33578.6 (manually installed from Intel's website). Confirmed intel-opencl-icd also installed.
sycl-ls Output:
[level_zero:gpu][level_zero:0] Intel(R) oneAPI Unified Runtime over Level-Zero, Intel(R) Arc(TM) A770 Graphics 12.55.8 [1.6.33578.600000]
[opencl:cpu][opencl:0] Intel(R) OpenCL, AMD Ryzen 7 8700F 8-Core Processor OpenCL 3.0 (Build 0) [2025.19.4.0.18_160000.xmain-hotfix]
IPEX-LLM Ollama Builds Tested:
ollama-ipex-llm-2.3.0b20250612-ubuntu.tgz
ollama-ipex-llm-2.3.0b20250429-ubuntu.tgz

Relevant Environment Variables set in /usr/local/bin/start_ollama_ipex.sh (consistent across tests):
Bash

export OLLAMA_NUM_GPU=999
export no_proxy=localhost,127.0.0.1
export ZES_ENABLE_SYSMAN=1
export SYCL_CACHE_PERSISTENT=1
export OLLAMA_
...

I am trying to get the intel arc a 770 gpu to increase its utilisation with an llm load under ollama. Eventually working with an LLM I got hold of the IPEX version of ollama and tried it - blitter worked a little bit before the crash. So there is hope that it can be made to work. Previously it had been sending all the load to the CPU but it was not recognising the GPU which it now does. I would try this with two gpus eventually but I have to learn to walk before I can run. Cheers Regards Michael Fothergill

[funky-gibbon]

I forgot to add that the machine I have has an amd 8700f cpu with 32GB RAM and a b650 Auros Eagle AX V2 motherboard and an intel A 770 16GB graphics card. Cheers MF