22 KiB
Run llama.cpp with IPEX-LLM on Intel GPU
< English | 中文 >
ggerganov/llama.cpp provides fast LLM inference in pure C++ across a variety of hardware; you can now use the C++ interface of ipex-llm as an accelerated backend for llama.cpp running on Intel GPU (e.g., local PC with iGPU, discrete GPU such as Arc, Flex and Max).
Important
You may use llama.cpp Portable Zip to directly run llama.cpp on Intel GPU with ipex-llm (without the need of manual installations).
Note
For installation on Intel Arc B-Series GPU (such as B580), please refer to this guide.
Note
Our latest version is consistent with 4ad2436 of llama.cpp.
ipex-llm[cpp]==2.2.0b20250629is consistent with d7cfe1f of llama.cpp.
See the demo of running LLaMA2-7B on Intel Arc GPU below.
 |
| You could also click here to watch the demo video. |
Table of Contents
- Prerequisites
- Install IPEX-LLM for llama.cpp
- Setup for running llama.cpp
- Example: Running community GGUF models with IPEX-LLM
- Troubleshooting
Quick Start
This quickstart guide walks you through installing and running llama.cpp with ipex-llm.
0 Prerequisites
IPEX-LLM's support for llama.cpp now is available for Linux system and Windows system.
Linux
For Linux system, we recommend Ubuntu 20.04 or later (Ubuntu 22.04 is preferred).
Visit the Install IPEX-LLM on Linux with Intel GPU, follow Install Intel GPU Driver to install GPU driver for Ubuntu 22.04 (or follow the Intel client GPU driver installation guide for a higher version of Ubuntu). And follow the guide here to install Intel® oneAPI Base Toolkit 2025.0.
Windows (Optional)
Please make sure your GPU driver version is equal or newer than 31.0.101.5522. If it is not, follow the instructions in this section to update your GPU driver; otherwise, you might encounter gibberish output.
1. Install IPEX-LLM for llama.cpp
To use llama.cpp with IPEX-LLM, first ensure that ipex-llm[cpp] is installed.
-
For Linux users:
conda create -n llm-cpp python=3.11 conda activate llm-cpp pip install --pre --upgrade ipex-llm[cpp] -
For Windows users:
Please run the following command in Miniforge Prompt.
conda create -n llm-cpp python=3.11 conda activate llm-cpp pip install --pre --upgrade ipex-llm[cpp]
After the installation, you should have created a conda environment, named llm-cpp for instance, for running llama.cpp commands with IPEX-LLM.
2. Setup for running llama.cpp
First you should create a directory to use llama.cpp, for instance, use following command to create a llama-cpp directory and enter it.
mkdir llama-cpp
cd llama-cpp
Initialize llama.cpp with IPEX-LLM
Then you can use following command to initialize llama.cpp with IPEX-LLM:
-
For Linux users:
init-llama-cppAfter
init-llama-cpp, you should see many soft links ofllama.cpp's executable files and aconvert.pyin current directory.
-
For Windows users:
Please run the following command with administrator privilege in Miniforge Prompt.
init-llama-cpp.batAfter
init-llama-cpp.bat, you should see many soft links ofllama.cpp's executable files and aconvert.pyin current directory.
Tip
init-llama-cppwill create soft links of llama.cpp's executable files to current directory, if you want to use these executable files in other places, don't forget to run above commands again.
Note
If you have installed higher version
ipex-llm[cpp]and want to upgrade your binary file, don't forget to remove old binary files first and initialize again withinit-llama-cpporinit-llama-cpp.bat.
Now you can use these executable files by standard llama.cpp's usage.
Runtime Configuration
To use GPU acceleration, several environment variables are required or recommended before running llama.cpp.
-
For Linux users:
source /opt/intel/oneapi/setvars.sh export SYCL_CACHE_PERSISTENT=1 # [optional] under most circumstances, the following environment variable may improve performance, but sometimes this may also cause performance degradation export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1 # [optional] if you want to run on single GPU, use below command to limit GPU may improve performance export ONEAPI_DEVICE_SELECTOR=level_zero:0 -
For Windows users:
Please run the following command in Miniforge Prompt.
set SYCL_CACHE_PERSISTENT=1 rem under most circumstances, the following environment variable may improve performance, but sometimes this may also cause performance degradation set SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
Tip
When your machine has multi GPUs and you want to run on one of them, you need to set
ONEAPI_DEVICE_SELECTOR=level_zero:[gpu_id], here[gpu_id]varies based on your requirement. For more details, you can refer to this section.
Note
The environment variable
SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTSdetermines the usage of immediate command lists for task submission to the GPU. While this mode typically enhances performance, exceptions may occur. Please consider experimenting with and without this environment variable for best performance. For more details, you can refer to this article.
3. Example: Running community GGUF models with IPEX-LLM
Here we provide a simple example to show how to run a community GGUF model with IPEX-LLM.
Model Download
Before running, you should download or copy community GGUF model to your current directory. For instance, mistral-7b-instruct-v0.1.Q4_K_M.gguf of Mistral-7B-Instruct-v0.1-GGUF.
Run the quantized model
-
For Linux users:
./llama-cli -m mistral-7b-instruct-v0.1.Q4_K_M.gguf -n 32 --prompt "Once upon a time, there existed a little girl who liked to have adventures. She wanted to go to places and meet new people, and have fun" -c 1024 -t 8 -e -ngl 99 --color -no-cnvNote
:
For more details about meaning of each parameter, you can use
./llama-cli -h. -
For Windows users:
Please run the following command in Miniforge Prompt.
llama-cli -m mistral-7b-instruct-v0.1.Q4_K_M.gguf -n 32 --prompt "Once upon a time, there existed a little girl who liked to have adventures. She wanted to go to places and meet new people, and have fun" -c 1024 -t 8 -e -ngl 99 --color -no-cnvNote
:
For more details about meaning of each parameter, you can use
./llama-cli -h.
Sample Output
main: llama backend init
main: load the model and apply lora adapter, if any
llama_model_load_from_file_impl: using device SYCL0 (Intel(R) Arc(TM) A770 Graphics) - 15473 MiB free
llama_model_loader: loaded meta data with 20 key-value pairs and 291 tensors from /home/arda/ruonan/mistral-7b-instruct-v0.1.Q4_K_M.gguf (version GGUF V2)
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv 0: general.architecture str = llama
llama_model_loader: - kv 1: general.name str = mistralai_mistral-7b-instruct-v0.1
llama_model_loader: - kv 2: llama.context_length u32 = 32768
llama_model_loader: - kv 3: llama.embedding_length u32 = 4096
llama_model_loader: - kv 4: llama.block_count u32 = 32
llama_model_loader: - kv 5: llama.feed_forward_length u32 = 14336
llama_model_loader: - kv 6: llama.rope.dimension_count u32 = 128
llama_model_loader: - kv 7: llama.attention.head_count u32 = 32
llama_model_loader: - kv 8: llama.attention.head_count_kv u32 = 8
llama_model_loader: - kv 9: llama.attention.layer_norm_rms_epsilon f32 = 0.000010
llama_model_loader: - kv 10: llama.rope.freq_base f32 = 10000.000000
llama_model_loader: - kv 11: general.file_type u32 = 15
llama_model_loader: - kv 12: tokenizer.ggml.model str = llama
llama_model_loader: - kv 13: tokenizer.ggml.tokens arr[str,32000] = ["<unk>", "<s>", "</s>", "<0x00>", "<...
llama_model_loader: - kv 14: tokenizer.ggml.scores arr[f32,32000] = [0.000000, 0.000000, 0.000000, 0.0000...
llama_model_loader: - kv 15: tokenizer.ggml.token_type arr[i32,32000] = [2, 3, 3, 6, 6, 6, 6, 6, 6, 6, 6, 6, ...
llama_model_loader: - kv 16: tokenizer.ggml.bos_token_id u32 = 1
llama_model_loader: - kv 17: tokenizer.ggml.eos_token_id u32 = 2
llama_model_loader: - kv 18: tokenizer.ggml.unknown_token_id u32 = 0
llama_model_loader: - kv 19: general.quantization_version u32 = 2
llama_model_loader: - type f32: 65 tensors
llama_model_loader: - type q4_K: 193 tensors
llama_model_loader: - type q6_K: 33 tensors
print_info: file format = GGUF V2
print_info: file type = Q4_K - Medium
print_info: file size = 4.07 GiB (4.83 BPW)
load: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
load: special tokens cache size = 3
load: token to piece cache size = 0.1637 MB
print_info: arch = llama
print_info: vocab_only = 0
print_info: n_ctx_train = 32768
print_info: n_embd = 4096
print_info: n_layer = 32
print_info: n_head = 32
print_info: n_head_kv = 8
print_info: n_rot = 128
print_info: n_swa = 0
print_info: n_embd_head_k = 128
print_info: n_embd_head_v = 128
print_info: n_gqa = 4
print_info: n_embd_k_gqa = 1024
print_info: n_embd_v_gqa = 1024
print_info: f_norm_eps = 0.0e+00
print_info: f_norm_rms_eps = 1.0e-05
print_info: f_clamp_kqv = 0.0e+00
print_info: f_max_alibi_bias = 0.0e+00
print_info: f_logit_scale = 0.0e+00
print_info: n_ff = 14336
print_info: n_expert = 0
print_info: n_expert_used = 0
print_info: causal attn = 1
print_info: pooling type = 0
print_info: rope type = 0
print_info: rope scaling = linear
print_info: freq_base_train = 10000.0
print_info: freq_scale_train = 1
print_info: n_ctx_orig_yarn = 32768
print_info: rope_finetuned = unknown
print_info: ssm_d_conv = 0
print_info: ssm_d_inner = 0
print_info: ssm_d_state = 0
print_info: ssm_dt_rank = 0
print_info: ssm_dt_b_c_rms = 0
print_info: model type = 7B
print_info: model params = 7.24 B
print_info: general.name = mistralai_mistral-7b-instruct-v0.1
print_info: vocab type = SPM
print_info: n_vocab = 32000
print_info: n_merges = 0
print_info: BOS token = 1 '<s>'
print_info: EOS token = 2 '</s>'
print_info: UNK token = 0 '<unk>'
print_info: LF token = 13 '<0x0A>'
print_info: EOG token = 2 '</s>'
print_info: max token length = 48
load_tensors: loading model tensors, this can take a while... (mmap = true)
load_tensors: offloading 32 repeating layers to GPU
load_tensors: offloading output layer to GPU
load_tensors: offloaded 33/33 layers to GPU
load_tensors: CPU_Mapped model buffer size = 70.31 MiB
load_tensors: SYCL0 model buffer size = 4095.05 MiB
.................................................................................................
llama_init_from_model: n_seq_max = 1
llama_init_from_model: n_ctx = 1024
llama_init_from_model: n_ctx_per_seq = 1024
llama_init_from_model: n_batch = 1024
llama_init_from_model: n_ubatch = 1024
llama_init_from_model: flash_attn = 0
llama_init_from_model: freq_base = 10000.0
llama_init_from_model: freq_scale = 1
llama_init_from_model: n_ctx_per_seq (1024) < n_ctx_train (32768) -- the full capacity of the model will not be utilized
Running with Environment Variables:
GGML_SYCL_DEBUG: 0
GGML_SYCL_DISABLE_OPT: 1
Build with Macros:
GGML_SYCL_FORCE_MMQ: no
GGML_SYCL_F16: no
Found 1 SYCL devices:
| | | | |Max | |Max |Global | |
| | | | |compute|Max work|sub |mem | |
|ID| Device Type| Name|Version|units |group |group|size | Driver version|
|--|-------------------|---------------------------------------|-------|-------|--------|-----|-------|---------------------|
| 0| [level_zero:gpu:0]| Intel Arc A770 Graphics| 12.55| 512| 1024| 32| 16225M| 1.6.31294.120000|
SYCL Optimization Feature:
|ID| Device Type|Reorder|
|--|-------------------|-------|
| 0| [level_zero:gpu:0]| Y|
llama_kv_cache_init: kv_size = 1024, offload = 1, type_k = 'f16', type_v = 'f16', n_layer = 32, can_shift = 1
llama_kv_cache_init: SYCL0 KV buffer size = 128.00 MiB
llama_init_from_model: KV self size = 128.00 MiB, K (f16): 64.00 MiB, V (f16): 64.00 MiB
llama_init_from_model: SYCL_Host output buffer size = 0.12 MiB
llama_init_from_model: SYCL0 compute buffer size = 164.01 MiB
llama_init_from_model: SYCL_Host compute buffer size = 20.01 MiB
llama_init_from_model: graph nodes = 902
llama_init_from_model: graph splits = 2
common_init_from_params: setting dry_penalty_last_n to ctx_size = 1024
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
main: llama threadpool init, n_threads = 8
system_info: n_threads = 8 (n_threads_batch = 8) / 32 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | LLAMAFILE = 1 | OPENMP = 1 | AARCH64_REPACK = 1 |
sampler seed: 403565315
sampler params:
repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000
dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 1024
top_k = 40, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.800
mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000
sampler chain: logits -> logit-bias -> penalties -> dry -> top-k -> typical -> top-p -> min-p -> xtc -> temp-ext -> dist
generate: n_ctx = 1024, n_batch = 4096, n_predict = 32, n_keep = 1
Once upon a time, there existed a little girl who liked to have adventures. She wanted to go to places and meet new people, and have fun exploring the world. But sometimes, she found it hard to find friends who shared her interests.
One day, she decided to take matters into her own
llama_perf_sampler_print: sampling time = x.xx ms / 63 runs ( x.xx ms per token, xx.xx tokens per second)
llama_perf_context_print: load time = xx.xx ms
llama_perf_context_print: prompt eval time = xx.xx ms / 31 tokens ( xx.xx ms per token, xx.xx tokens per second)
llama_perf_context_print: eval time = xx.xx ms / 31 runs ( xx.xx ms per token, xx.xx tokens per second)
llama_perf_context_print: total time = xx.xx ms / 62 tokens
Troubleshooting
1. Unable to run the initialization script
If you are unable to run init-llama-cpp.bat, please make sure you have installed ipex-llm[cpp] in your conda environment. If you have installed it, please check if you have activated the correct conda environment. Also, if you are using Windows, please make sure you have run the script with administrator privilege in prompt terminal.
2. DeviceList is empty. -30 (PI_ERROR_INVALID_VALUE) error
On Linux, this error happens when devices starting with [ext_oneapi_level_zero] are not found. Please make sure you have installed level-zero, and have sourced /opt/intel/oneapi/setvars.sh before running the command.
3. Prompt is too long error
If you encounter main: prompt is too long (xxx tokens, max xxx), please increase the -c parameter to set a larger size of context.
4. gemm: cannot allocate memory on host error / could not create an engine error
If you meet oneapi::mkl::oneapi::mkl::blas::gemm: cannot allocate memory on host error, or could not create an engine on Linux, this is probably caused by pip installed OneAPI dependencies. You should prevent installing like pip install dpcpp-cpp-rt==2024.0.2 mkl-dpcpp==2024.0.0 onednn==2024.0.0, and instead use apt to install on Linux. Please refer to this guide for more details.
5. Fail to quantize model
If you encounter main: failed to quantize model from xxx, please make sure you have created related output directory.
6. Program hang during model loading
If your program hang after llm_load_tensors: SYCL_Host buffer size = xx.xx MiB, you can add --no-mmap in your command.
7. How to set -ngl parameter
-ngl means the number of layers to store in VRAM. If your VRAM is enough, we recommend putting all layers on GPU, you can just set -ngl to a large number like 999 to achieve this goal.
If -ngl is set to 0, it means that the entire model will run on CPU. If -ngl is set to greater than 0 and less than model layers, then it's mixed GPU + CPU scenario.
8. How to specificy GPU
If your machine has multi GPUs, llama.cpp will default use all GPUs which may slow down your inference for model which can run on single GPU. You can add -sm none in your command to use one GPU only.
Also, you can use ONEAPI_DEVICE_SELECTOR=level_zero:[gpu_id] to select device before excuting your command, more details can refer to here.
9. Program crash with Chinese prompt
If you run the llama.cpp program on Windows and find that your program crashes or outputs abnormally when accepting Chinese prompts, you can search for region in the Windows search bar and go to Region->Administrative->Change System locale.., tick Beta: Use Unicode UTF-8 for worldwide language support option and then restart your computer.
For detailed instructions on how to do this, see this issue.
10. sycl7.dll not found error
If you meet System Error: sycl7.dll not found on Windows or you meet similar error on Linux, please check:
- if you have installed conda and if you are in the right conda environment which has pip installed oneapi dependencies on Windows
- if you have executed
source /opt/intel/oneapi/setvars.shon Linux
11. Check driver first when you meet garbage output on Windows
If you meet garbage output on Windows, please check if your GPU driver version is >= 31.0.101.5522. If not, please follow the instructions in this section to update your GPU driver.
12. Why my program can't find sycl device
If you meet GGML_ASSERT: C:/Users/Administrator/actions-runner/cpp-release/_work/llm.cpp/llm.cpp/llama-cpp-bigdl/ggml-sycl.cpp:18283: main_gpu_id<g_all_sycl_device_count error or similar error, and you find nothing is output when using ls-sycl-device, this is because llama.cpp cannot find the sycl device. On some laptops, the installation of the ARC driver may lead to a forced installation of OpenCL, OpenGL, and Vulkan Compatibility Pack by Microsoft, which inadvertently blocks the system from locating sycl devices. This issue can be resolved by manually uninstalling it in Microsoft store.
13. Core dump when having both integrated and dedicated graphics
If you have both integrated and dedicated graphics displayed in your llama.cpp's device log and don't specify which device to use, it will cause a core dump. In such case, you may need to specify export ONEAPI_DEVICE_SELECTOR=level_zero:0 before running llama-cli.
14. Native API failed error
On latest version of ipex-llm, you might come across native API failed error with certain models without the -c parameter. Simply adding -c xx would resolve this problem.
15. signal: bus error (core dumped) error
If you meet this error, please check your Linux kernel version first. You may encounter this issue on higher kernel versions (like kernel 6.15). You can also refer to this issue to see if it helps.
16. backend buffer base cannot be NULL error
If you meet ggml-backend.c:96: GGML_ASSERT(base != NULL && "backend buffer base cannot be NULL") failed, simply adding -c xx parameter during inference, for example -c 1024 would resolve this problem.