Intel Arc B70 / BMG G31: IPEX-Ollama oneAPI runner crashes on qwen2.5:0.5b, CPU path works

[AnonyymiIsmo]

Summary

I am testing IPEX-LLM / IPEX-Ollama on Intel Arc Pro B70 / BMG G31 on Ubuntu 24.04.4.

The Intel GPU is detected correctly through the oneAPI backend, and the model is offloaded to the GPU, but generation crashes. The same Qwen2.5 0.5B model works with the same IPEX-Ollama build when running in CPU mode.

Hardware / OS

• GPU: Intel Graphics [0xe223], BMG G31 / Arc Pro B70
• VRAM reported by IPEX-Ollama: 31.9 GiB total, 29.8 GiB available
• Kernel driver: xe
• Resizable BAR: BAR 2 current size 32GB
• OS: Ubuntu 24.04.4
• Kernel: 6.17.0-29-generic

IPEX / Ollama environment

• ipex-llm: 2.3.0b20251110
• IPEX-Ollama reports version: 0.9.3
• OLLAMA_INTEL_GPU=true
• OLLAMA_NUM_GPU=999
• OLLAMA_NUM_PARALLEL=1
• OLLAMA_CONTEXT_LENGTH=1024
• ONEAPI_DEVICE_SELECTOR=level_zero:0
• SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
• ZES_ENABLE_SYSMAN=1

What works

• Host OpenCL sees Intel Graphics [0xe223].
• VAAPI works.
• PyTorch XPU matmul works.
• IPEX-Ollama server starts.
• IPEX-Ollama detects the Intel GPU:
  • library=oneapi
  • name="Intel(R) Graphics [0xe223]"
  • total="31.9 GiB"
• qwen2.5:0.5b works on CPU control port with:
  • OLLAMA_INTEL_GPU=false
  • OLLAMA_NUM_GPU=0

What fails

qwen2.5:0.5b on Intel GPU / oneAPI

The model loads/offloads to the Intel GPU, but generation fails with:

HTTP 500:
model runner has unexpectedly stopped

Reducing:

• OLLAMA_NUM_PARALLEL=1
• OLLAMA_CONTEXT_LENGTH=1024
• num_batch=128
• num_batch=64

does not fix it.

gemma4:e4b

gemma4:e4b cannot be tested with this build because the loader reports:

unknown model architecture: 'gemma4'

This appears to be due to IPEX-Ollama 0.9.3 not supporting the newer gemma4 GGUF architecture.

Expected behavior

qwen2.5:0.5b should generate normally on Intel Arc B70 through oneAPI/SYCL.

Actual behavior

qwen2.5:0.5b works on CPU, but the oneAPI GPU runner crashes during generation.

Attachment

I attached a small bugreport tar.gz with:

• start scripts
• GPU-mode log
• CPU-control log
• conda list
• pip freeze
• system/GPU/package information

bugreport-b70-ipex-ollama-2026-05-21-193230.tar.gz

[AnonyymiIsmo]

Additional note:

I opened this here because intel/ipex-llm appears to be archived/read-only. Please redirect me if there is a better current place for IPEX-Ollama / Intel Arc B70 oneAPI runner issues.

The most relevant crash signature from the GPU-mode log is:

library=oneapi
name="Intel(R) Graphics [0xe223]"
load_tensors: offloaded 25/25 layers to GPU

sdp_xmx_kernel.cpp:439: Assertion `false' failed.
SIGABRT: abort
HTTP 500: model runner has unexpectedly stopped

The CPU-control run with the same qwen2.5:0.5b model returns HTTP 200 and generates OK, so this seems isolated to the oneAPI/SYCL GPU runner path rather than the model store or Ollama API in general.

[edisonchan]

The issue seems related to IPEX-LLM with llama.cpp/Ollama, but IPEX has been archived and is no longer maintained. Since llm-scaler mainly focuses on vLLM and ComfyUI, this might be better addressed in the respective project’s issue tracker.

[AnonyymiIsmo]

The issue seems related to IPEX-LLM with llama.cpp/Ollama, but IPEX has been archived and is no longer maintained. Since llm-scaler mainly focuses on vLLM and ComfyUI, this might be better addressed in the respective project’s issue tracker.

Thank you for the clarification.

That makes sense. I opened the issue here mainly because intel/ipex-llm is archived/read-only and I was not sure where Intel Arc B70 / oneAPI LLM runtime issues should currently be reported.

Since this appears to be from the old IPEX-LLM Ollama/llama.cpp path, I will treat that path as unmaintained and not production-ready.

Could you clarify the currently recommended Intel-supported path for Arc Pro B70 LLM inference testing?

For example, should I move testing to:

• llm-scaler-vLLM
• upstream vLLM
• upstream llama.cpp
• upstream Ollama
• or another Intel GPU / oneAPI issue tracker?

I can retest with the currently recommended stack if that is the supported direction.

[AnonyymiIsmo]

I accidentally closed this issue. Reopening it, since I would still appreciate guidance on the currently recommended Intel-supported path for Arc Pro B70 LLM inference testing.

[edisonchan]

llama.cpp built with xpu support is a better choice for you, it does support almost all model, for example Qwen 3.5/Qwen 3.6.

git clone https://github.com/ggml-org/llama.cpp
cd llama.cpp
source ~/youraccount/oneapi/setvars.sh 
cmake -B build -DGGML_SYCL=ON   -DCMAKE_C_COMPILER=icx   -DCMAKE_CXX_COMPILER=icpx   -DCMAKE_ASM_COMPILER=/home/youraccount/intel/oneapi/compiler/2026.0/bin/icx
cmake --build build -j$(nproc)

build/bin/llama-server \
  -m ~/LLM/gguf/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf \
  --host 0.0.0.0 \
  --port 8000 \
  --ctx-size 16384 \
  --n-gpu-layers -1 \
  --parallel 1 \
  --jinja \
  --chat-template-kwargs '{"preserve_thinking": true}' \
  --cache-type-k q8_0 \
  --cache-type-v q8_0 \
  --reasoning on \
  --temp 0.6 \
  --top-p 0.95 \
  --top-k 20 \
  --min-p 0.0 \
  --presence-penalty 0.0 \
  --repeat-penalty 1.0 \
  --batch-size 2048 \
  --ubatch-size 512 \
  -sm layer \
  --perf \
  --metrics