Felipe97/llama-cpp-compiled
01.1k
1:: MIT license2:: Copyright (C) 2024 Intel Corporation3:: SPDX-License-Identifier: MIT4 5@echo off6setlocal EnableExtensions EnableDelayedExpansion7 8set "BIN_FILE=.\build\bin\llama-server.exe"9set "SEED=0"10set "GPUS_SETTING="11 12set "MODEL_FILE=..\models\Qwen3.5-4B-Q4_0.gguf"13set "NGL=99"14set "CONTEXT=4096"15set "GGML_SYCL_DEVICE=-1"16set "SYCL_DEVICES=SYCL0"17set "SPLIT_MODE=layer"18set "LOG_VERBOSE=3"19 20if "%~1"=="" goto after_args21 22:parse_args23if "%~1"=="" goto after_args24 25if /I "%~1"=="-c" (26 if "%~2"=="" goto missing_value27 set "CONTEXT=%~2"28 shift29 shift30 goto parse_args31)32if /I "%~1"=="--context" (33 if "%~2"=="" goto missing_value34 set "CONTEXT=%~2"35 shift36 shift37 goto parse_args38)39 40if /I "%~1"=="-d" (41 if "%~2"=="" goto missing_value42 set "SYCL_DEVICES=%~2"43 shift44 shift45 goto parse_args46)47if /I "%~1"=="--device" (48 if "%~2"=="" goto missing_value49 set "SYCL_DEVICES=%~2"50 shift51 shift52 goto parse_args53)54 55if /I "%~1"=="-m" (56 if "%~2"=="" goto missing_value57 set "MODEL_FILE=%~2"58 shift59 shift60 goto parse_args61)62if /I "%~1"=="--model" (63 if "%~2"=="" goto missing_value64 set "MODEL_FILE=%~2"65 shift66 shift67 goto parse_args68)69 70if /I "%~1"=="-mg" (71 if "%~2"=="" goto missing_value72 set "GGML_SYCL_DEVICE=%~2"73 set "SPLIT_MODE=none"74 shift75 shift76 goto parse_args77)78if /I "%~1"=="--main-gpu" (79 if "%~2"=="" goto missing_value80 set "GGML_SYCL_DEVICE=%~2"81 set "SPLIT_MODE=none"82 shift83 shift84 goto parse_args85)86 87if /I "%~1"=="-sm" (88 if "%~2"=="" goto missing_value89 set "SPLIT_MODE=%~2"90 shift91 shift92 goto parse_args93)94if /I "%~1"=="--split-mode" (95 if "%~2"=="" goto missing_value96 set "SPLIT_MODE=%~2"97 shift98 shift99 goto parse_args100)101 102if /I "%~1"=="-ngl" (103 if "%~2"=="" goto missing_value104 set "NGL=%~2"105 shift106 shift107 goto parse_args108)109if /I "%~1"=="--n-gpu-layers" (110 if "%~2"=="" goto missing_value111 set "NGL=%~2"112 shift113 shift114 goto parse_args115)116 117if /I "%~1"=="-lv" (118 if "%~2"=="" goto missing_value119 set "LOG_VERBOSE=%~2"120 shift121 shift122 goto parse_args123)124if /I "%~1"=="--log-verbosity" (125 if "%~2"=="" goto missing_value126 set "LOG_VERBOSE=%~2"127 shift128 shift129 goto parse_args130)131 132if /I "%~1"=="-h" goto help133if /I "%~1"=="--help" goto help134 135echo Invalid option: %~1136exit /b 1137 138:missing_value139echo Missing value for option: %~1140exit /b 1141 142:help143echo Usage: %~n0 [OPTIONS]144echo.145echo This script processes files with specified options.146echo.147echo Options:148echo -h, --help Display this help message and exit.149echo -d, --device ^<value^> Set SYCL devices (default: SYCL0).150echo -c, --context ^<value^> Set context length. Bigger need more memory.151echo -m, --model ^<value^> Full model file path.152echo -mg,--main-gpu ^<value^> Set main GPU ID (0 - n) for single GPU mode.153echo -sm,--split-mode ^<value^> How to split the model across multiple GPUs, one of:154echo - none: use one GPU only155echo - layer (default): split layers and KV across GPUs156echo - row: split rows across GPUs157echo -ngl,--n-gpu-layers ^<value^> Max. number of layers to store in VRAM (default: -1)158echo -lv,--log-verbosity ^<value^> Set the verbosity threshold. Messages with a higher verbosity will be159echo ignored. Values:160echo - 0: generic output161echo - 1: error162echo - 2: warning163echo - 3: info164echo - 4: debug165exit /b 0166 167:after_args168 169REM In Windows CMD, source is not available; call oneAPI setvars if present.170if exist "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" (171 call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" >nul172) else (173 echo Warning: oneAPI setvars.bat not found. Continuing without environment setup.174)175 176REM Support malloc device memory more than 4GB.177set "UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=1"178echo UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=%UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS%179 180echo ONEAPI_DEVICE_SELECTOR=%ONEAPI_DEVICE_SELECTOR%181 182if not "%GGML_SYCL_DEVICE%"=="-1" (183 echo Use %GGML_SYCL_DEVICE% as main GPU184 REM Use single GPU only.185 set "GPUS_SETTING=-mg %GGML_SYCL_DEVICE% -sm %SPLIT_MODE%"186 ) else (187 echo Use Intel GPUs: %SYCL_DEVICES%188 set "GPUS_SETTING=-sm %SPLIT_MODE%"189)190 191echo run cmd: ZES_ENABLE_SYSMAN=1 %BIN_FILE% -m "%MODEL_FILE%" -ngl %NGL% -s %SEED% -c %CONTEXT% %GPUS_SETTING% -lv %LOG_VERBOSE% --device %SYCL_DEVICES% --load-mode auto --host 0.0.0.0 --port 8000192set "ZES_ENABLE_SYSMAN=1"193%BIN_FILE% -m "%MODEL_FILE%" -ngl %NGL% -s %SEED% -c %CONTEXT% %GPUS_SETTING% -lv %LOG_VERBOSE% --device "%SYCL_DEVICES%" --load-mode auto --host 0.0.0.0 --port 8000194 195endlocal196 197 