Public » Linux

OneAPI i software¶

1. Instalacja bibliotek PokażUkryj

   mkdir /usr/src/neo/; \
   cd /usr/src/neo; \
   wget https://github.com/intel/intel-graphics-compiler/releases/download/v2.34.4/intel-igc-core-2_2.34.4+21428_amd64.deb; \
   wget https://github.com/intel/intel-graphics-compiler/releases/download/v2.34.4/intel-igc-opencl-2_2.34.4+21428_amd64.deb; \
   wget https://github.com/intel/compute-runtime/releases/download/26.18.38308.1/intel-ocloc-dbgsym_26.18.38308.1-0_amd64.ddeb; \
   wget https://github.com/intel/compute-runtime/releases/download/26.18.38308.1/intel-ocloc_26.18.38308.1-0_amd64.deb; \
   wget https://github.com/intel/compute-runtime/releases/download/26.18.38308.1/intel-opencl-icd-dbgsym_26.18.38308.1-0_amd64.ddeb; \
   wget https://github.com/intel/compute-runtime/releases/download/26.18.38308.1/intel-opencl-icd_26.18.38308.1-0_amd64.deb; \
   wget https://github.com/intel/compute-runtime/releases/download/26.18.38308.1/libigdgmm12_22.10.0_amd64.deb; \
   wget https://github.com/intel/compute-runtime/releases/download/26.18.38308.1/libze-intel-gpu1-dbgsym_26.18.38308.1-0_amd64.ddeb; \
   wget https://github.com/intel/compute-runtime/releases/download/26.18.38308.1/libze-intel-gpu1_26.18.38308.1-0_amd64.deb; \
    \
   apt-get install ./*; \
   apt-get install intel-oneapi-dnnl-devel-2026.0 intel-oneapi-dnnl-2026.0'
   

Intel Deep Learning Essentials¶

1. Instalacja PokażUkryj

   apt-get install -y gpg; \
   wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor | tee /usr/share/keyrings/oneapi-archive-keyring.gpg > /dev/null; \
   echo "deb [signed-by=/usr/share/keyrings/oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main" | tee /etc/apt/sources.list.d/oneAPI.list; \
   apt-get update; \
   apt-get install intel-deep-learning-essentials;
   

2. Inicjalizacja PokażUkryj

   source /opt/intel/oneapi/setvars.sh;
   

llama.cpp + SYCL¶

1. Instalacja zależności(cmake itd.) PokażUkryj

   apt-get install -y cmake ccache libssl-dev;
   

2. Kompilacja llama.cpp PokażUkryj

   git clone https://github.com/ggml-org/llama.cpp `date +%Y%m%d-intel`; \
   cd `date +%Y%m%d-intel`; \
   \
   cmake -B build \
     -DGGML_SYCL=ON \
     -DGGML_SYCL_F16=ON \
     -DGGML_SYCL_TARGET=INTEL \
     -DGGML_SYCL_FA_ONEDNN=1 \
     -DGGML_SYCL_ENABLE_MKL_FA=1 \
     -DGGML_SYCL_DEVICE_ARCH=xe2 \
     -DGGML_SYCL_DNN=ON \
     -DCMAKE_C_COMPILER=icx \
     -DCMAKE_CXX_COMPILER=icpx \
     -DBUILD_SHARED_LIBS=0 \
     -DLLAMA_BUILD_UI=OFF -DLLAMA_USE_PREBUILT_UI=OFF \
     -DCMAKE_C_FLAGS="-O3 -march=native -mtune=native -pthread" \
     -DCMAKE_CXX_FLAGS="-O3 -march=native -mtune=native -pthread"; \
   \
   cmake --build build --config Release -j$(nproc);
   

   Uwagi:

   • -DGGML_SYCL_DEVICE_ARCH=bmg_g21 używamy bmg_g21 bo na ten moment nie ma innego przełącznika, Intel B70 wpisuje sie w tą architekturę

Pobranie modeli i uruchomienie¶

1. Pobranie modeli
   1. Instalacja zależności PokażUkryj

      apt-get install -y pipx; \
       \
      pipx install huggingface_hub; \
      pipx ensurepath; \
      source ~/.bashrc; \
       \
      export HF_XET_HIGH_PERFORMANCE=1;
      

   2. Qwen3.8-27B + Vision PokażUkryj

      # 32GB VRAM GPU z 8GB RAM Cache
      hf download unsloth/Qwen3.8-27B-GGUF \
        Qwen3.8-27B-UD-Q4_K_XL.gguf \
        --cache-dir ./models/qwen38-27b
      
      # 32GB VRAM GPU z 16GB RAM Cache
      hf download unsloth/Qwen3.8-27B-GGUF \
        Qwen3.8-27B-UD-Q6_K_XL.gguf \
        --cache-dir ./models/qwen38-27b
      
      # Model dla 2x32GB VRAM GPU
      hf download unsloth/Qwen3.8-27B-GGUF \
        Qwen3.8-27B-UD-Q8_K_XL.gguf \
        --cache-dir ./models/qwen38-27b
      
      # Model dla 1x32GB VRAM GPU, może nawet dla 16GB
      hf download ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF \
        Qwen3.8-27B-GSQ-RCO-IQ3_S-mtp.gguf \
        --cache-dir ./models/qwen38-27b
      
      # Vision mmproj 
      hf download unsloth/Qwen3.8-27B-GGUF \
        mmproj-F16.gguf \
        --cache-dir ./models/qwen38-27b;
      
      # DFlash2
      hf download incoai/Qwen3.8-27B-DFlash2-GGUF \
        Qwen3.8-27B-DFlash2-Q8_0.gguf \
        --cache-dir ./models/qwen38-27b;
      
      

   3. Qwen3.8-27B-Flash-Next PokażUkryj

      # 96GB VRAM GPU z 64GB RAM Cache
      hf download unsloth/Qwen3.8-Flash-Next-GGUF \
        --include "*UD-Q2_K_XL*" \
        --cache-dir ./models/qwen38-flash-next
      

      # 96GB VRAM GPU z 64GB RAM Cache
      hf download unsloth/Qwen3.8-Flash-Next-GGUF \
        --include "*UD-IQ3_XXS*" \
        --cache-dir ./models/qwen38-flash-next
      

      # 96GB VRAM GPU z 64GB RAM Cache
      hf download unsloth/Qwen3.8-Flash-Next-GGUF \
        --include "MTP/mtp-Qwen3.8-Flash-Next-shared-Q4_K_M.gguf" \
        --cache-dir ./models/qwen38-flash-next
      

   4. Qwen3.6-35B-A3B-MTP PokażUkryj

      # 32GB VRAM GPU z 64GB RAM Cache
      hf download unsloth/Qwen3.6-35B-A3B-MTP-GGUF \
        Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf \
        --cache-dir ./models/qwen3.6-35B-A3B
      

   5. VSCode FIM/NextLine/Embedding/ReRanker Auto Complete dla IDE PokażUkryj

      # FIM/NextLine dla Auto Complete IDE
      hf download unsloth/Qwen2.5-Coder-3B-Instruct-128K-GGUF \
        Qwen2.5-Coder-3B-Instruct-Q8_0.gguf \
        --cache-dir ./models/Qwen2.5-Coder-3B
      
      # Embedding
      hf download unsloth/embeddinggemma-300m-GGUF \
        embeddinggemma-300M-Q8_0.gguf \
        --cache-dir ./models/embeddinggemma-300M
      
      # ReRanker
      hf download sinjab/jina-reranker-v2-base-multilingual-Q4_K_M-GGUF \
        jina-reranker-v2-base-multilingual-Q4_K_M.gguf \
        --cache-dir ./models/jina-reranker-v2-base-multilingual
      

   6. Vision Qwen3.5-9B PokażUkryj

      Mały model kodujący z Vision który można uruchomić na osobnej karcie GPU i wykorzystywać do analizy przez zewnętrzne oprogramowanie
      

      hf download unsloth/Qwen3.5-9B-GGUF \
        Qwen3.5-9B-UD-Q4_K_XL.gguf \
        --cache-dir ./models/Qwen3.5-9B;
      
      # Vision mmproj 
      hf download unsloth/Qwen3.5-9B-GGUF \
        mmproj-BF16.gguf \
        --cache-dir ./models/Qwen3.5-9B;
      

   7. Audio STT Qwen3-ASR-1.7B PokażUkryj

      Mały model STT który można uruchomić na starszym GPU aby obsługiwał STT i wykorzystywać do analizy przez zewnętrzne oprogramowanie
      

      # Qwen3-ASR Audio
      hf download unslothai/Qwen3-ASR-1.7B-GGUF \
        Qwen3-ASR-1.7B-Q8_0.gguf \
        --cache-dir ./models/Qwen3-ASR-1.7B;
      
      # Qwen3-ASR Audio mmproj
      hf download unslothai/Qwen3-ASR-1.7B-GGUF \
        mmproj-Qwen3-ASR-1.7B-Q8_0.gguf \
        --cache-dir ./models/Qwen3-ASR-1.7B;
      

   8. Audio TTS llama-tts PokażUkryj
      1. OuteAI/Llama-OuteTTS-1.0-1B-GGUF
         hf download Serveurperso/Qwen3-TTS-GGUF \
         qwen-talker-1.7b-customvoice-Q8_0.gguf \
         --cache-dir ./models/qwen3-tts;
   9. Archiwum
      Modele zastąpione nowszymi wersjami
      1. Qwen3.6-27B-MTP PokażUkryj

         hf download unsloth/Qwen3.6-27B-MTP-GGUF \
           Qwen3.6-27B-UD-Q4_K_XL.gguf \
           --cache-dir ./models/qwen36-27b-MTP;
         
         # Model dla GPU z 16GB VRAM
         hf download unsloth/Qwen3.6-27B-MTP-GGUF \
           Qwen3.6-27B-UD-IQ3_XXS.gguf \
           --cache-dir ./models/qwen36-27b-MTP;
         

      2. Qwen2.5-Coder-7B-Ins by AceCoder PokażUkryj

         Model ma mniejszy context tylko 32K, zjada więcej RAM niż modele unsloth/Qwen2.5-Coder-3B-Instruct-128K-GGUF Qwen2.5-Coder-3B-Instruct-Q8_0.gguf
         

         hf download mradermacher/AceCoder-Qwen2.5-Coder-7B-Ins-V1.1-i1-GGUF \
           AceCoder-Qwen2.5-Coder-7B-Ins-V1.1.i1-Q4_K_M.gguf \
           --cache-dir ./models/Qwen2.5-Coder-7B
         

2. Uruchomienie
   • Model kodujacy PokażUkryj

     TREE=20260723
     MODEL=/home/storage/stripe/llm/models/Qwen2.5-Coder-3B/models--unsloth--Qwen2.5-Coder-3B-Instruct-128K-GGUF/snapshots/5326551926d06f7f9cab53c9b9b552e3bedfe8ba/Qwen2.5-Coder-3B-Instruct-Q8_0.gguf
     
     CTX=131072
     RAM=12288
     
     IP=x.x.x.x
     PORT=8081
     
     ./${TREE}/build/bin/llama-server \
         -m ${MODEL} \
         \
         -t 4 \
         -fa 1 \
         \
         --parallel 2 \
         \
         --ctx-size ${CTX} \
         --cache-ram ${RAM} \
         --cache-type-v q8_0 \
         --cache-type-k q8_0 \
         \
         --jinja \
         --reasoning off\
         \
         --temperature 0.6 \
         --top-k 20 \
         --top-p 0.95 \
         --min-p 0.0 \
         --repeat-penalty 1.0 \
         --presence-penalty 0.0 \
         \
         --batch-size 2048 \
         --ubatch-size 2048 \
         \
         --host ${IP} \
         --port ${PORT}
     

   • Model Embedding PokażUkryj

     TREE=20260723
     MODEL=/home/storage/stripe/llm/models/embeddinggemma-300M/models--unsloth--embeddinggemma-300m-GGUF/snapshots/6661a6504c30d8304af13455cb4a5d4f5bc6011f/embeddinggemma-300M-Q8_0.gguf
     
     CTX=8192     # maksymalny rozmiar jaki był trenowany dla czterech wątków
     RAM=2048
     
     IP=x.x.x.x
     PORT=8082
     
     ./${TREE}/build/bin/llama-server \
         --embedding \
         -m ${MODEL} \
         \
         -t 4 \
         -fa 1 \
         \
         --parallel 4 \
         \
         --ctx-size ${CTX} \
         --cache-ram ${RAM} \
         --cache-type-v q8_0 \
         --cache-type-k q8_0 \
         \
         --jinja \
         --reasoning off\
         \
         --batch-size 2048 \
         --ubatch-size 2048 \
         \
         --host ${IP} \
         --port ${PORT}
     

   • Model Embedding PokażUkryj

     TREE=20260723
     MODEL=/home/storage/stripe/llm/models/jina-reranker-v2-base-multilingual/models--sinjab--jina-reranker-v2-base-multilingual-Q4_K_M-GGUF/snapshots/c1b17df2ce6b810e31e2cae9491b98eecf05947b/jina-reranker-v2-base-multilingual-Q4_K_M.gguf
     
     CTX=4096        # Maksymalny jaki był trenowany
     RAM=2048
     
     IP=x.x.x.x
     PORT=8083
     
     ./${TREE}/build/bin/llama-server \
         --reranking \
         -m ${MODEL} \
         --alias jina-reranker-v2 \
         \
         -fa 1 \
         \
         --parallel 4 \
         \
         --ctx-size ${CTX} \
         --cache-ram ${RAM} \
         --cache-type-v q8_0 \
         --cache-type-k q8_0 \
         \
         --jinja \
         --reasoning off\
         \
         --batch-size 2048 \
         --ubatch-size 2048 \
         \
         --host ${IP} \
         --port ${PORT}
     

Materiały dodatkowe:¶

• https://bibek-poudel.medium.com/how-to-run-qwen3-6-27b-locally-on-intel-arc-pro-b70-what-actually-works-c96dec67c6f7
• https://github.com/intel/compute-runtime/releases
• https://github.com/ggml-org/llama.cpp/releases
• https://www.intel.com/content/www/us/en/developer/tools/oneapi/oneapi-toolkit-download.html?packages=dl-essentials&dl-essentials-os=linux&dl-lin=apt
• https://www.intel.com/content/www/us/en/docs/oneapi-toolkit/installation-guide-linux/latest/install-deep-learning-essentials-with-apt.html#INSTALL-DEEP-LEARNING-ESSENTIALS-WITH-APT