Could not create a primitive descriptor for a matmul primitive during qwen3 32b inference with portable llama.cpp

[XCRobert]

Describe the bug

1. Install latest driver and install latest windows version llama.cpp form https://github.com/ipex-llm/ipex-llm/releases/tag/v2.3.0-nightly
2. download Qwen3-32B-Q8_0.gguf model
3. prepare a 64GB machine with intel igpus
4. Modify regkey:
   [HKEY_LOCAL_MACHINE\System\CurrentControlSet\Control\GraphicsDrivers\MemoryManager] "SystemPartitionCommitLimitPercentage"
   set it to 0x4b ,which means use 75% memory as shared gpu memory
5. llama-cli.exe -m ..\models\Qwen3\Qwen3-32B-Q8_0.gguf -p "how to become an expert on GPU driver" -n 2048 -e -ngl 999 --color -c 2500 --temp 0 -no-cnv

Then will meet issue:
llama_context: SYCL_Host output buffer size = 0.58 MiB
llama_kv_cache_unified: SYCL0 KV buffer size = 632.00 MiB
llama_kv_cache_unified: size = 632.00 MiB ( 2528 cells, 64 layers, 1 seqs), K (f16): 316.00 MiB, V (f16): 316.00 MiB
llama_context: SYCL0 compute buffer size = 1497.80 MiB
llama_context: SYCL_Host compute buffer size = 73.54 MiB
llama_context: graph nodes = 2054
llama_context: graph splits = 2
common_init_from_params: setting dry_penalty_last_n to ctx_size = 2528
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
could not create a primitive descriptor for a matmul primitive
Exception caught at file:D:\actions-runner\release-cpp-oneapi_2024_2_work\llm.cpp\llm.cpp\llama-cpp-bigdl\ggml\src\ggml-sycl\ggml-sycl.cpp, line:3552, func:operator()
SYCL error: CHECK_TRY_ERROR(op(ctx, src0, src1, dst, src0_dd_i, src1_ddf_i, src1_ddq_i, dst_dd_i, dev[i].split_dim_low, dev[i].split_dim_high, src1_ncols, src1_padded_col_size, stream)): Exception caught in this line of code.
in function ggml_sycl_op_mul_mat at D:\actions-runner\release-cpp-oneapi_2024_2_work\llm.cpp\llm.cpp\llama-cpp-bigdl\ggml\src\ggml-sycl\ggml-sycl.cpp:3552
D:\actions-runner\release-cpp-oneapi_2024_2_work\llm.cpp\llm.cpp\llama-cpp-bigdl\ggml\src\ggml-sycl..\ggml-sycl\common.hpp:127: SYCL error

[qiyuangong]

Hi @XCRobert .

You command is correct. This error maybe related to oneAPI verson conflict.

Note that this llama.cpp portable package is built with oneAPI 2025.0 (dependencies already in package). It may conflict with 2024.2 in your env.

[bibekyess]

Hello @qiyuangong
One of my clients also faced the same issue when using portable ollama. I am wondering since portable executable is already self contained with oneAPI 2025.0, then why is the PC trying to load 2024.2, can you let us know how to fix it?
"""
source=server.go:657 msg="llama runner started in 3.03 seconds"
could not create a primitive descriptor for a matmul primitive
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:2876, func:operator()
SYCL 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)): Exception caught in this line of code.
in function ggml_sycl_op_mul_mat at D:\actions-runner\release-cpp-oneapi_2024_2_work\llm.cpp\llm.cpp\ollama-llama-cpp\ggml\src\ggml-sycl\ggml-sycl.cpp:2876
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=2026-01-02T14:49:23.360Z level=ERROR source=server.go:827 msg="post predict" error="Post "http://127.0.0.1:59410/completion\": read tcp 127.0.0.1:59412->127.0.0.1:59410: wsarecv: An existing connection was forcibly closed by the remote host."
time=2026-01-02T14:49:23.384Z level=ERROR source=server.go:484 msg="llama runner terminated" error="exit status 0xc0000409"
"""

After some experiments, it doesn't look like oneAPI version conflicts:
Notebook Specs (Panther Lake)
OS: win32 (10.0.26200) - x64
CPU: Genuine Intel(R) 0000
Total RAM: 23.57 GB
Free RAM: 11.47 GB
GPU: Intel(R) Graphics

We tried with these combinations but still facing the same error.
GPU DRIVER version: 32.0.101.8074 or 32.0.101.8188
OneAPI: 2024.2 or 2025.3