if [ -n "$FISH_VERSION" ]; then
echo "❌FEHLER BITTE STARTE MIT 'bash XAIGPUARC.sh' ODER INSTALLIERE DIR 'bass'"
exit 1
fi

#|✅XAIGPUARC✅

#|DATUM
#|04.06.2026

#|196598k7bFP16MathTutorGGUF 70WattDualGPULaptop

#|UHRZEIT
#|01:23

set -euo pipefail
IFS=$'\n\t'
PRECISION="FP16"
DEVICE="ARC"
IGPU_VRAM_FALLBACK_GIB="${IGPU_VRAM_FALLBACK_GIB:-25}"
IGPU_VRAM_RESERVE_GIB="${IGPU_VRAM_RESERVE_GIB:-1}"
LLAMA_CPP_DIR="llama.cpp"
BUILD_DIR="${BUILD_DIR:-XAIGPUARC}"
BUILD_DIR="${BUILD_DIR%/}"

#|KONFIGURIERE UMGEBUNGSVARIABLEN VORBAU XAIGPUARC
GGML_SYCL_CPP="${LLAMA_CPP_DIR}/ggml/src/ggml-sycl/ggml-sycl.cpp"
LOG_FILE="${BUILD_DIR}/XAIGPUARC.log"
CMAKE_BUILD_TYPE="${CMAKE_BUILD_TYPE:-Release}"

NPROC="${NPROC:-$(nproc)}"
LLAMA_CLI_PATH="bin/llama-cli"
LS_SYCL_DEVICE_PATH="bin/llama-ls-sycl-device"
ADD_SUBDIR_LINE="${LLAMA_CPP_DIR}/ggml/src/ggml-sycl/ggml-sycl.cpp"

#|ONEAPIFUNKTIONEN
export LIBVA_DRIVER_NAME="iHD,i915,xe"
export TCM_ROOT="${TCM_ROOT:-"/opt/intel/oneapi/tcm/latest/"}"

export SYCL_CACHE_PERSISTENT=0
export OCL_ICD_FILENAMES=""
export ZES_ENABLE_SYSMAN=0
export OverrideDefaultFP64Settings=0
export CCACHE_DIR="$HOME/.ccache"
export COMPILER_VERSION="2026.0"

export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
export SYCL_PI_LEVEL_ZERO_BATCH_SIZE=512
export FP_FLAG=FP16

export SYCL_DEVICE_FILTER=level_zero:gpu
export GGML_SYCL_SPLIT_MODE=row
export ZE_ENABLE_PCI_ID_DEVICE_ORDER=1

export ONEAPI_DEVICE_SELECTOR="level_zero:0,1"

#|TREIBERKOMPATIBILITAET INTEL COMPUTE RUNTIME 26 PLUS
CR_VERSION=$(pacman -Q intel-compute-runtime 2>/dev/null | awk '{print $2}' | cut -d. -f1)
if [ "${CR_VERSION:-0}" -ge 26 ]; then
export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=0
export GGML_SYCL_DISABLE_OPT=1
echo "🔷INTEL COMPUTE RUNTIME 26 PLUS ERKANNT KOMPATIBILITAETSMODUS AKTIV"
fi
#|HILFSFUNKTIONEN
log() { printf "🔷 %s\n" "$"; }
success() { printf "✅ %s\n" "$"; }
error() { printf "❌ %s\n\n" "$"; }
warn() { printf "⚠️ %s\n" "$"; }
#|INTERNETPRUEFUNG
check_internet() {
log "🔷PRUEFE INTERNETZ VERBINDUNG"
if timeout 5 bash -c "</dev/tcp/8.8.8.8/53" 2>/dev/null; then
success "✅INTERNETZ VERBINDUNG VORHANDEN"
return 0
else
warn "⚠️KEINE INTERNETZ VERBINDUNG ERSTINSTALLATION ONLINE ABHAENGIGKEITEN BITTE IHREN ANSCHLUSS PRUEFEN"
return 1
fi
}
#|UMGEBUNG|RUECKFALL MECHANISMEN VORBEREITEN
prepare_environment() {
log "🔷HOLE ONEAPI KOPFZEILEN XAIGPUARC"
local SETVARS_PATH="/opt/intel/oneapi/setvars.sh"
if [ ! -f "$SETVARS_PATH" ]; then
error "❌ONE API KOEPFZEILEN NICHT GEFUNDEN $SETVARS_PATH INSTALLIERE ONEAPI BIBLIOTHEKEN"
exit 1
fi
log "🔷SETVARS SETZEN PLUS SUCHEN SS PLUS S"
source "$SETVARS_PATH" --force 2>/dev/null
local ONEAPI_ROOT_FALLBACK="/opt/intel/oneapi"
local COMPILER_VERSION_FALLBACK="${COMPILER_VERSION:-2026.0}"
DPCPP_ROOT="${DPCPP_ROOT:-${ONEAPI_ROOT_FALLBACK}/compiler/${COMPILER_VERSION_FALLBACK}}"
MKL_ROOT="${MKL_ROOT:-${ONEAPI_ROOT_FALLBACK}/mkl/${COMPILER_VERSION_FALLBACK}}"
ONEAPI_ROOT="${ONEAPI_ROOT:-${ONEAPI_ROOT_FALLBACK}}"
export CC=icx
export CXX=icpx
export FC=ifx
export DPCPP_ROOT
export MKL_ROOT
export ONEAPI_ROOT
export CPATH="${CPATH:-}:${MKL_ROOT}/include"
local LIB_DIR="/opt/intel/oneapi/mkl/latest/lib/intel64/libmkl_sycl.so"
export LD_LIBRARY_PATH="./${BUILD_DIR}/bin:${LIB_DIR}:${LD_LIBRARY_PATH:-}"
if ! command -v icx &>/dev/null || ! command -v icpx &>/dev/null; then
error "❌ICX IPCX ONE API XAIGPUARC UNTERMODUL INSTALLATION FEHLGESCHLAGEN"
exit 1
fi
log "🔷VERBINDUNG ONEAPI GELADEN DPCPP${DPCPP_ROOT}MKL${MKL_ROOT}"
}
#|1|PROJEKT VORBAU
setup_project() {
if [[ "${1:-}" == "--clean" ]]; then
warn "⚠️ REINIGUNGSMODUS LOESCHE ALTE BAUSTEINE"
rm -rf "${LLAMA_CPP_DIR}"
rm -rf "${BUILD_DIR}"
success "✅ REINIGUNG ABGESCHLOSSEN"
fi
log "🔷BEREITE XAIGPUARC VORBAU BITTE WARTEN"
if [ ! -d "${LLAMA_CPP_DIR}" ]; then
log "🔷ZIEHE BAUSTEINE AUS INTERNETZ"
git clone https://github.com/ggerganov/llama.cpp "${LLAMA_CPP_DIR}"
if [ ! -d "${LLAMA_CPP_DIR}" ]; then
error "❌ZIEHEN DER BAUSTEINE FEHLGESCHLAGEN ABBRUCH"
exit 1
fi
fi
if pushd "${LLAMA_CPP_DIR}" > /dev/null; then
log "🔷AKTUALISIERE BAUSTEINE FUER UNTERMODULE"
git reset --hard HEAD
git pull
git submodule update --init --recursive
popd > /dev/null
success "✅LADE BAUSTEINE IN XAIGPUARC UNTERBAUMODULE"
else
error "❌FEHLER HAUPTVERZEICHNIS'${LLAMA_CPP_DIR}'NICHT GEFUNDEN"
exit 1
fi
}
#|AUFWERTLOGIK 0 BIS 12 +a+b+c
patch_llama_cpp() {
log "🔷PATCH 1|12 GGML SYCL DOCTPHELPER REGESTRIERUNG BIBLIOTHEK KOPFZEILENEINTRAEGE"
local DPCT_HELPER_FILE="${LLAMA_CPP_DIR}/ggml/src/ggml-sycl/dpct/helper.hpp"
local CMAKE_LISTS_FILE="${LLAMA_CPP_DIR}/ggml/src/ggml-sycl/CMakeLists.txt"
local CUSTOM_KERNEL_DIR="${LLAMA_CPP_DIR}/ggml/src/ggml-sycl/custom_kernels"
local CUSTOM_KERNEL_SRC="${CUSTOM_KERNEL_DIR}/ggml_flash_attention_sycl.cpp"
local CUSTOM_KERNEL_CMAKE="${CUSTOM_KERNEL_DIR}/CMakeLists.txt"
local GGML_SYCL_CPP="${LLAMA_CPP_DIR}/ggml/src/ggml-sycl/ggml-sycl.cpp"
local KERNEL_SOURCE_LOCAL="ggml_flash_attention_sycl.cpp"
#|1|12
if [ -f "$DPCT_HELPER_FILE" ]; then
log "🔷PATCH 1|12 DOCTPHELPER BIBLIOTHEK REGESTRIERTE KOPFZEILENEINTRAEGE LADEN"
if sed -i 's|#include <sycl/ext/oneapi/math.hpp>|#include <sycl/ext/intel/math.hpp>|g' "$DPCT_HELPER_FILE"; then
log "🔷PATCH 1|12 SCHREIBE DOCTPHELPER BIBLIOTHEK KOPFZEILENEINTRAEGE IN XAIGPUARC"
elif sed -i '#if !defined(DPCT_USM_LEVEL_NONE) && defined(DPCT_ENABLE_MKL_MATH).#endif|#include <sycl|ext|intel|math.hpp>|g' "$DPCT_HELPER_FILE"; then
log "✅PATCH 1|12 DOCTPHELPER BIBLIOTHEK KOPFZEILENEINTRAEGE ERFOLGREICH REGESTRIERT - SPEICHERE"
#|1|a|12
else
error "❌PATCH 1|a|12 DOCTPHELPER BIBLIOTHEK KOPFZEILENEINTRAEGE EINLADEN DER REGESTRIERUNG FEHLGESCHLAGEN"
return 1
fi
else
warn "⚠️PATCH 1|a|12 DOCTPHELPER BIBLIOTHEK NICHT GEFUNDEN ABHAENIGKEITEN PRUEFEN"
return 1
fi
#|2|12
log "🔷PATCH 2|12 BAUE FLASH ATTENTION KERN"
if [ ! -d "$CUSTOM_KERNEL_DIR" ]; then
mkdir -p "$CUSTOM_KERNEL_DIR"
log "✅PATCH 2|12 ORNDER FUER FLASH ATTENTION KERN'${CUSTOM_KERNEL_DIR}'ANGELEGT"
fi
if [ -f "$KERNEL_SOURCE_LOCAL" ]; then
cp "$KERNEL_SOURCE_LOCAL" "$CUSTOM_KERNEL_SRC"
log "✅PATCH 2|12 ggml_flash_attention_sycl.cpp KERN'${KERNEL_SOURCE_LOCAL}' '${CUSTOM_KERNEL_SRC}'EINGEBAUT"
fi

#|2|a|12
if [ ! -f "$CUSTOM_KERNEL_SRC" ]; then
echo "ggml_flash_attention_sycl.cpp KERN VERZEICHNIS" > "$CUSTOM_KERNEL_SRC"
warn "⚠️PATCH 2|a|12 LADEN DER KERNELDATEI'${KERNEL_SOURCE_LOCAL} FEHLGESCHLAGEN"
fi
echo "
add_library(ggml_flash_attention_sycl OBJECT
ggml_flash_attention_sycl.cpp)
target_include_directories(ggml_flash_attention_sycl PRIVATE ${GGML_SYCL_INCLUDE_DIRS})
target_compile_options(ggml_flash_attention_sycl PUBLIC ${GGML_SYCL_COMPILE_FLAGS})
" > "$CUSTOM_KERNEL_CMAKE"
log "🔷PATCH 2|a|12 CMAKE TXT LISTEN OBJEKTE KOPFZEILENEINTRAEGE EINFUEGEN"

#|2|b|12
local ADD_SUBDIR_LINE="add_subdirectory(ggml_flash_attention_sycl)"
if ! grep -Fq "${ADD_SUBDIR_LINE}" "$CMAKE_LISTS_FILE"; then
if sed -i "|#add_subdirectory(dpct)|a\${ADD_SUBDIR_LINE}" "$CMAKE_LISTS_FILE"; then

log "✅PATCH 2|b|12 ERFOLGREICH FLASH ATTENTION KOPFZEILENEINTRAEGE CMAKE TXT GESCHRIEBEN"
else
error "❌PATCH 2|b|12 FLASH ATTENTION KOPFZEILENEINTRAEGE EINGLIEDERUNG CMAKE TXT FEHLGESCHLAGEN"
return 1
fi
else
log "🔷PATCH 2|b|12 FLASH ATTENTION KOPFZEILENEINTRAEGE BEREITS AKTIV UEBERSPRINGE"
fi
#|3|12
if [ -f "$CMAKE_LISTS_FILE" ]; then
log "🔷PATCH 3|12 CMAKE TEXT LISTEN MKL KOPZEILEN ICP|X IMPLEMENTIERUNG VORBEREITEN"
local MKL_INCLUDE_PATH="${MKL_ROOT}/include"
local COMPILER_INCLUDE_PATH="${DPCPP_ROOT}/include"
local DPCPP_LIB_INCLUDE_PATH="${DPCPP_ROOT}/lib/dpcpp/include"
local ALL_INCLUDE_FLAGS="-I${MKL_INCLUDE_PATH} -I${COMPILER_INCLUDE_PATH} -I${DPCPP_LIB_INCLUDE_PATH}"
local PATCH_LINE="target_compile_options (ggml-sycl PUBLIC ${ALL_INCLUDE_FLAGS})"
local SEARCH_MARKER="#Add include directories for MKL headers"
if ! grep -Fq "${COMPILER_INCLUDE_PATH}" "$CMAKE_LISTS_FILE"; then
local SED_PATCH_LINE=$(echo "$PATCH_LINE" | sed 's| |\ |g; s|[|&]|&|g')
if sed -i "|${SEARCH_MARKER}|a $SED_PATCH_LINE" "$CMAKE_LISTS_FILE"; then
log "✅PATCH 3|12 ICP|X CMAKET LISTS TXT MKL KOPFZEILENEINTRAEGE EINGEFUEGT"
else
error "❌PATCH 3|12 ICP|X CMAKE LISTS TXT MKL NICHT GEFUNDEN ABHAENGIKEITEN PRUEFEN"
return 1
fi
else
log "✅PATCH 3|12 CMAKE LISTS TXT PFAD SYCL GGML BEREITS BENUTZT UEBERSPRINGE"
fi
else
error "❌PATCH 3|12 FEHLGESCHLAGEN CMAKE LISTS TXT SYCL GGML PFADE ABHAENGIGKEITEN GARUDA LINUX ARCH"
return 1
fi

#|4|12
log "🔷PATCH 4|12 FLASH ATTENTION HAUPTKERN KOPFZEILENEINTRAEGE INJIZIEREN"
if [ -f "$GGML_SYCL_CPP" ]; then
local FA_REGISTER_CODE=$'//REGESTRIERE ggml_flash_attention_sycl.cpp \nextern "C"
void ggml_flash_attention_sycl(ggml_flash_attention_sycl * ctx, ggml_tensor *
dst, const ggml_tensor * Q, const ggml_tensor * K, const ggml_tensor * V);\n'
if ! grep -Fq "ggml_flash_attention_sycl" "${GGML_SYCL_CPP}"; then

#|Anmerkung FlashAttention alucian einfuegen in HOME Ordner
echo "${FA_REGISTER_CODE}" > /tmp/ggml_flash_attention_sycl.cpp
awk '/extern "C" void ggml_flash_attention_sycl/ { system("cat /tmp/ggml_flash_attention_sycl.patch"); } { print }' "${GGML_SYCL_CPP}" > /tmp/ggml-sycl.cpp.new
mv /tmp/ggml-sycl.cpp.new "${GGML_SYCL_CPP}"
cp "$GGML_SYCL_CPP" "$GGML_SYCL_CPP.bak"
if [ $? -eq 0 ]; then
log "✅PATCH 4|12 AWK KOPFZEILENEINTRAEGE DEKLARATION EINGEFUEGT"
else
error "❌PATCH 4|12 FEHLER EINFUEGEN FLASH ATTENTION KOPFZEILENEINTRAEGE DEKLARATION AWK FEHLT"
return 1
fi
else
log "🔷PATCH 4|12 FLASH ATTENTION KOPFZEILENEINTRAEGE DEKLARATIONEN VORHANDEN FORTFAHREN"
fi
#|4|a|12
local FA_DISPATCH_CASE=$' case GGML_OP_FLASH_ATTN:\n ggml_flash_attention_sycl(ctx, dst, src0, src1, src2);\n break;'
if ! grep -Fq "case GGML_OP_FLASH_ATTN:" "${GGML_SYCL_CPP}"; then
log "🔷PATCH 4|a|12 FUEGE ZWISCHENSPEICHER PER AWK KOPFZEILE EIN"
echo "${FA_DISPATCH_CASE}" > /tmp/fa_dispatch.patch
awk '/case GGML_OP_MUL_MAT_Q_K:/ { system("cat /tmp/fa_dispatch.patch"); } { print }' "${GGML_SYCL_CPP}" > /tmp/ggml-sycl.cpp.new
mv /tmp/ggml-sycl.cpp.new "${GGML_SYCL_CPP}"
if [ $? -eq 0 ]; then

log "✅PATCH 4|a|12 AWK UNTERBAU KOPFZEILENEINTRAEGE EINGEFUEHRT"
else
error "❌PATCH 4|a|12 FEHLER EINFUEGEN AWK KOPFZEILENEINTRAEGE"
fi
else
log "✅PATCH 4|a|12 AWK UNTERBAU FLASH ATTENTION KOPFZEILENEINTRAEGE INJEKTION VORHANDEN FORTFAHREN"
fi
log "🔷PATCH 4|a|12 ERFOLGREICH FLASH ATTENTION AKW UNTERBAU GELADEN"
else
error "❌PATCH 4|a|12 FEHLGESCHLAGEN FLASH ATTENTION AWK PATCH KERN STOPP"
return 1
fi

#|5|12
log "🔷PATCH 5|12FLASH ATTENTION OBJEKT INJIZIERE VARIABLEN UNTERBLOCK SYCL BIBLIOTHEKEN"
local CMAKE_LISTS_FILE="${LLAMA_CPP_DIR}/ggml/src/ggml-sycl/CMakeLists.txt"
local VAR_LINE="set(FA_OBJECT_FILES "$<TARGET_OBJECTS:ggml_flash_attention_sycl>")"
local VAR_SEARCH_MARKER="set(GGML_SYCL_SOURCES"
if ! grep -Fq "FA_OBJECT_FILES" "$CMAKE_LISTS_FILE"; then
local SED_VAR_LINE=$(echo "$VAR_LINE" | sed 's/[/&]/\&/g')
if sed -i "|${VAR_SEARCH_MARKER}|a ${SED_VAR_LINE}" "$CMAKE_LISTS_FILE"; then
log "✅PATCH 5|12 FLASH ATTENTION OBJEKT VARIABLE ERFOLGREICH DEFINIERT WEITER"
else
error "❌PATCH 5|12 FLASH ATTENTION OBJEKT REGESTRIERUNG FEHLGESCHLAGEN STOPP"
return 1
fi
else
log "🔷PATCH 5|12 FLASH ATTENTION OBJEKT VARIABLEN VORHANDEN UEBERSPRINGE"
fi

#|5|a|12
local TARGET_SEARCH_MARKER="target_sources(ggml-sycl PRIVATE ${GGML_SYCL_SOURCES})"
local NEW_TARGET_SOURCES_LINE="target_sources(ggml-sycl PRIVATE ${GGML_SYCL_SOURCES} ${FA_OBJECT_FILES})"
if grep -Fq "${TARGET_SEARCH_MARKER}" "$CMAKE_LISTS_FILE" && ! grep -Fq "${FA_OBJECT_FILES}" "$CMAKE_LISTS_FILE"; then
local SED_NEW_LINE=$(echo "$NEW_TARGET_SOURCES_LINE" | sed 's/[/&]/\&/g')
local SED_SEARCH_MARKER=$(echo "$TARGET_SEARCH_MARKER" | sed 's/[/&]/\&/g')
if sed -i "|${SED_SEARCH_MARKER}|${SED_NEW_LINE}|" "$CMAKE_LISTS_FILE"; then
log "✅PATCH 5|a|12 ERFOLGREICHE GGML SYCL INJEKTIONEN"
else
error "❌PATCH 5|a|12 GGML SYCL INJEKTION FEHLGESCHLAGEN"
return 1
fi
else
log "✅PATCH 5|a|12 GGML SYCL AKTIV INJECTION UEBERSPRUNGEN"
fi

#|6|12
log "🔷PATCH 6|12 SSMCONV CPP WARNUNG BEHEBEN VORZEICHEN VERGLEICH AKW PATCH"
local SSM_CONV_FILE="${LLAMA_CPP_DIR}/ggml/src/ggml-sycl/ssm_conv.cpp"
local SEARCH_LINE='GGML_ASSERT(src0->nb[1] == src0->ne[0] * static_cast<int64_t>(sizeof(float)));'
local REPLACE_LINE='GGML_ASSERT(src0->nb[1] == (size_t)(src0->ne[0] * sizeof(float)));'
if [ -f "$SSM_CONV_FILE" ]; then

#|6|a|12
if grep -Fq "${SEARCH_LINE}" "$SSM_CONV_FILE"; then
log "🔷PATCH 6a|12 SUCHE ssm_conv.cpp NATIVE ADRESSIERUNG"
if grep -Fq "${SEARCH_LINE}" "$SSM_CONV_FILE"; then
if sed -i "|${SEARCH_LINE}|${REPLACE_LINE}|g" "$SSM_CONV_FILE"; then
log "✅PATCH 6|a|12 SSMCONV CPP ERFOLGREICH GESCHRIEBEN"
else
error "❌PATCH 6|a|12 SSMCONV CPP FEHLGESCHLAGEN"
return 1
fi
else
warn "⚠️PATCH 6|a|12 SSMCONV CPP ZEILE NICHT GEFUNDEN UEBERSPRINGE"
fi
fi

#ATTENTION 7|12

#|7|12
log "🔷PATCH 7|12 ERZWINGE MAXIMALE BLOCK GROESSE 512 FUER ARC ALCHEMIST"
if [ -f "$GGML_SYCL_CPP" ]; then
if ! grep -q "GGML_SYCL_MAX_BLOCK_SIZE 512" "$GGML_SYCL_CPP"; then
sed -i 's/#define GGML_SYCL_MAX_BLOCK_SIZE [0-9]/#define GGML_SYCL_MAX_BLOCK_SIZE 512/g' "$GGML_SYCL_CPP"
success "✅PATCH 7|12 ERZWINGE MAXIMALE BLOCK GROESSE 512 FUER ARC ALCHEMIST AKTIVIERT"
log "✅PATCH 7|12 MAXIMALE BLOCK GROESSE 512 ZWANG ERFOLGREICH INJIZIERT"
else
log "🔷PATCH 7|12 MAXIMALE BLOCK GROESSE 512 ZWANG FUER ARC BEREITS DEFINIERT"
fi
fi
#|8|12
log "🔷PATCH 8|12 SYCL QUEUE ORDNER: HOME/XAIGPUARC OPTIMIERVORGANG KOPFZEILENEINTRAEGE"
local SYCL_FILE="${LLAMA_CPP_DIR}/ggml/src/ggml-sycl/ggml-sycl.cpp"
if [ ! -f "$SYCL_FILE" ]; then
error "❌PATCH 8|12 KONNTE DIE DATEI ${LLAMA_CPP_DIR}/ggml/src/ggml-sycl/ggml-sycl.cpp NICHT FINDEN"
return 1
fi
if grep -q "sycl::property_list" "$SYCL_FILE"; then
sed -i 's|sycl::property_list prop_list{[^}]}|sycl::property_list prop_list{sycl::property::queue::in_order{}}|' "$SYCL_FILE"
else
sed -i '/sycl::queue q{/i\ sycl::property_list prop_list{sycl::property::queue::in_order{}};' "$SYCL_FILE"
fi
fi
#|9|12
if grep -q "sycl::queue q{" "$SYCL_FILE"; then
sed -i 's|sycl::queue q{[^}]*}|sycl::queue q{dev, ctx, prop_list}|g' "$SYCL_FILE"
if grep -q "in_order" "$SYCL_FILE"; then
success "✅PATCH 9|12 SYCL QUEUE ERFOLGREICH KOPFZEILENEINTRAEGE OPTIMIERT"
else
error "❌PATCH 9|12 SYCL QUEUE AENDERUNGEN KOPFZEILENEINTRAEGE NICHT FERTIG"
return 1
fi
else
warn "⚠️PATCH 9|12 SYCL QUEUE KOPFZEILENEINTRAEGE NICHT GEFUNDEN UEBERSPRINGE"
fi
success "🔷ALLE FUENF VON ACHT EINGLIEDERUNGEN INTEL XE/ARC/iGPU/dGPU/eGPU FUER XAIGPUARC ERFOLGREICH"
patch_gcc16_t5 || return 1
}

#|10|a|12
#|GCC16 T5.CPP VORWAERTSDEKLARATION VOR INSTANTIIERUNG
patch_gcc16_t5() {
log "🔷PATCH 10|a|12 GCC16 T5.CPP VORWAERTSDEKLARATION VOR INSTANTIIERUNG"
local T5_FILE="${LLAMA_CPP_DIR}/src/models/t5.cpp"
if [ ! -f "$T5_FILE" ]; then
warn "⚠️PATCH 10|a|12 T5.CPP NICHT GEFUNDEN UEBERSPRINGE"
return 0
fi
if grep -q "XAIGPUARC_GCC16_PATCH" "$T5_FILE"; then
log "🔷PATCH 10|a|12 GCC16 VORWAERTSDEKLARATION BEREITS AKTIV UEBERSPRINGE"
return 0
fi
awk '/^std::unique_ptr<llm_graph_context> llama_model_t5::build_arch_graph/ { print ""; print "// XAIGPUARC_GCC16_PATCH: Vorwaertsdeklaration vor Instantiierung"; print "template <> llama_model_t5::graph::graph(const llama_model & model, const llm_graph_params & params);"; print "template <> llama_model_t5::graph::graph(const llama_model & model, const llm_graph_params & params);"; } { print }' "$T5_FILE" > /tmp/t5_patched.cpp && mv /tmp/t5_patched.cpp "$T5_FILE"
if grep -q "XAIGPUARC_GCC16_PATCH" "$T5_FILE"; then
success "✅PATCH 10|a|12 GCC16 T5.CPP VORWAERTSDEKLARATION ERFOLGREICH INJIZIERT"
else
error "❌PATCH 10|a|12 GCC16 T5.CPP VORWAERTSDEKLARATION FEHLGESCHLAGEN"
return 1
fi

#|WEITERLEITUNG
patch_gcc16_t5
patch_10_xmx_dual
}

#|10|b|12
#|GCC 16 AUF T5 AWK VORWAERTS DECLARATION INJEKTION
patch_gcc16_t5() {
log "🔷PATCH 10|b|12 GCC16 T5.CPP AWK VORWAERTSDEKLARATION INJEKTION"
local T5_FILE="${LLAMA_CPP_DIR}/src/models/t5.cpp"
if [ ! -f "$T5_FILE" ]; then return 0; fi
if grep -q "XAIGPUARC_GCC16_PATCH" "$T5_FILE"; then return 0; fi
awk '/^std::unique_ptr<llm_graph_context> llama_model_t5::build_arch_graph/ { print ""; print "// XAIGPUARC_GCC16_PATCH: Vorwaertsdeklaration vor Instantiierung"; print "template <> llama_model_t5::graph::graph(const llama_model & model, const llm_graph_params & params);"; print "template <> llama_model_t5::graph::graph(const llama_model & model, const llm_graph_params & params);"; } { print }' "$T5_FILE" > /tmp/t5_patched.cpp && mv /tmp/t5_patched.cpp "$T5_FILE"
success "✅ PATCH 10|b|12 GCC16 T5 TEMPLATE VORWAERTSDEKLARATION EINGEFUEGT"
}

#ATTENTION 10|c|-|12|

#|10|c|BIS 12 WIEDER EINFUEGEN

#|3|BAUKONFIGURATION
configure_build() {
log "🔷BEREITE KOPFZEILEN BAUVORGANG VOR"
local FP_MODE="${1:-1}"
local SYCL_FLAGS="-DGGML_SYCL_F16=${FP_MODE} -fsycl-device-code-split=per_kernel"
if [ ! -d "${BUILD_DIR}" ]; then
log "🔷LEGE XAIGPUARC IN ORDNER HOME ${BUILD_DIR}"
mkdir -p "${BUILD_DIR}" || { error "❌KONNTE ORDNER XAIGPUARC'${BUILD_DIR}'NICHT ANLEGEN"; return 1; }
fi
if pushd "${BUILD_DIR}" > /dev/null; then
log "🔷STARTE CMAKE TXT KOPFZEILENBAU XAIGPUARC ${FP_FLAG:-FP_MODE}"
cmake "../${LLAMA_CPP_DIR}"
-G "Unix Makefiles"
-DCMAKE_EXPORT_COMPILE_COMMANDS=ON
-DCMAKE_BUILD_TYPE="${CMAKE_BUILD_TYPE}"
-DGGML_SYCL=ON
-DGGML_SYCL_TARGET=INTEL
-DGGML_SYCL_CCACHE=ON
-DGGML_SYCL_F16=${FP_MODE}
-DGGML_SYCL_FLASH_ATTN=ON
-DGGML_SYCL_MKL_SYCL_BATCH_GEMM=1
-DCMAKE_C_COMPILER=icx
-DCMAKE_CXX_COMPILER=icpx
-DCMAKE_CXX_STANDARD=23
-DCMAKE_CXX_STANDARD_REQUIRED=ON
-DCMAKE_CXX_EXTENSIONS=ON
-DGGML_SYCL_PRIORITIZE_DMMV=ON
-DGGML_SYCL_DISABLE_DNN=OFF
-DTHREADS_PREFER_PTHREAD_FLAG=ON
-DCMAKE_C_FLAGS="-O3 -ffast-math"
-DGGML_SYCL_DISABLE_GRAPH=OFF
-DCMAKE_THREAD_LIBS_INIT="-pthread"
-DCMAKE_C_COMPILER_WORKS=1
local CMAKE_STATUS=$?
popd > /dev/null
if [ ${CMAKE_STATUS} -ne 0 ]; then
error "❌CMAKE TXTKOPFZEILEN SCHREIBEN FEHLGESCHLAGEN"
return 1
fi
success "✅VORBAU ABGESCHLOSSEN"
else
error "❌KONNTE NICHT NACH HOME/XAIGPUARC WECHSELN'${BUILD_DIR}'COMPUTER NUTZER BERECHTIGUNG PRUEFEN"
return 1
fi
}

#|4|KOMPILIEREN
compile_project() {
log "🔷BAUE XAIGPUARC GRUND GERUEST STRUKTUR"
local LOG_FILE="build.log"
log "🔷KOPFZEILEN AUSGABE UNTERORNDER GESPEICHERT"
log "🔷BAUVORGANG LAEUFT XAIGPUARC SYCL C++ LEVEL ZERO KOPFZEILEN SCHREIBEN ABGESCHLOSSEN"
if pushd "${BUILD_DIR}" > /dev/null; then
log "🔷VORBBAU IN HAUPTORDNER GESCHICKT STARTE INSTALLATION UND DANNACH/ODER KI-INFERENZ
|🟢|XAIGPUARC
|🔧INSTALLATION auf ihrem Endgeraet in 🧰 .../HOME/USERNAME/XAIGPUARC/..."
cmake --build . --verbose --config "${CMAKE_BUILD_TYPE}" -j ${NPROC} --target llama-cli llama-ls-sycl-device > "${LOG_FILE}" 2>&1
local BUILD_STATUS=$?
popd > /dev/null
if [ ${BUILD_STATUS} -ne 0 ]; then
error "❌XAIGPUARC KOPFZEILEN FEHLGESCHLAGEN UEBERPRUEFEN${BUILD_DIR}/${LOG_FILE}"
return 1
fi
success "✅BAU XAIGPUARC ERFOLGREICH"
else
error "❌XAIGPUARC '${BUILD_DIR}' WEGEN FEHLERHAFTEM WECHSEL KOPZEILENBAU NICHT MOEGLICH"
return 1
fi
}

#|5|AUTOMATISCHE GERAETEAUSWAHL
auto_select_device() {
local FULL_LS_PATH="./${BUILD_DIR}/${LS_SYCL_DEVICE_PATH}"
log "🔷SUCHE NACH VERFUEGBAREN SYCL GERAETEN AUF IHREM SYSTEM"

if [ ! -x "${FULL_LS_PATH}" ]; then
warn "⚠️SYCL UNTERBAU NICHT GEFUNDEN ${FULL_LS_PATH} ZWEITSUCHE ODER RUECKFALL AUF CPU"
export ONEAPI_DEVICE_SELECTOR="level_zero:0"
DEVICE="DUAL"
return
fi

#|DOPPELTE GERAETEABFRAGE
log "🔷 STARTE ERSTE GERAETEABFRAGE"
export ONEAPI_DEVICE_SELECTOR="level_zero:gpu"
local DEVICES
DEVICES=$("${FULL_LS_PATH}" 2>&1 || true)

if [ -z "$DEVICES" ]; then
warn "⚠️KEINE KOMPATIBLEN SYCL GERAETE GEFUNDEN! RUECKFALL AUF EINZELGERAET"
export ONEAPI_DEVICE_SELECTOR="level_zero:0"
DEVICE="DUAL"
N_GPU_LAYERS=98
return
fi

#|ERSTER SUCHLAUF DOPPEL GPU
local ARC_LINE IGPU_LINE ARC_ID IGPU_ID
ARC_LINE=$(echo "$DEVICES" | grep -i "Intel Arc.Graphics" | head -n1)
IGPU_LINE=$(echo "$DEVICES" | grep -i "Iris.Xe.Graphics" | head -n1)

#|SUCHFEHLER GPU ZWEITER DURCHLAUF
if [ -z "$ARC_LINE" ] || [ -z "$IGPU_LINE" ]; then
warn "⚠️ SUCHFEHLER ERKANNT ODER GERAET FEHLT STARTE ZWEITEN DURCHLAUF"
sleep 1
export ONEAPI_DEVICE_SELECTOR="level_zero:gpu"
DEVICES=$("${FULL_LS_PATH}" 2>&1 || true)
[ -z "$ARC_LINE" ] && ARC_LINE=$(echo "$DEVICES" | grep -i "Intel Arc.Graphics" | head -n1)
[ -z "$IGPU_LINE" ] && IGPU_LINE=$(echo "$DEVICES" | grep -i "Iris.Xe.Graphics" | head -n1)
fi

#|ENDE DOPPELTE GERAETEABFRAGE
ARC_ID=$(echo "$ARC_LINE" | sed -n 's/.level_zero:gpu:([0-9]+)]./\1/p')
IGPU_ID=$(echo "$IGPU_LINE" | sed -n 's/.level_zero:gpu:([0-9]+)]./\1/p')

#|VRAM BERECHNUNG HILFSFUNKTION MiB SYCL AUS ZEILEN LESEN
get_vram_mib_from_line() {
local line="$1"
local fallback_gib="$2"
local raw
raw=$(echo "$line" | awk -F'|' '{gsub(/ /,"",$8); print $8}' | grep -oP '\d+(?=M)')
if [ -n "$raw" ] && [ "$raw" -gt 0 ] 2>/dev/null; then
echo "$raw"
else
echo $(( fallback_gib * 1024 ))
fi
}

#|DUALER MODUS: dGPU + iGPU zusammen
if [ -n "$ARC_ID" ] && [ -n "$IGPU_ID" ]; then
log "🔷DUAL-GPU MODUS AKTIV: dGPU(${ARC_ID}) + iGPU(${IGPU_ID}) VRAM POOL"

#|NUMERISCHE IDs AUS LEVEL_ZERO X EXTRAHIEREN
local ARC_NUM="$ARC_ID"
local IGPU_NUM="$IGPU_ID"

#|VRAM DUAL GPU BERECHNEN
local ARC_VRAM_MIB IGPU_VRAM_MIB
ARC_VRAM_MIB=$(get_vram_mib_from_line "$ARC_LINE" 12)
#|iGPU TEILT SICH ARBEITSSPEICHER MIT CPU RUECKFALL IGPU_VRAM_FALLBACK_GIB
IGPU_VRAM_MIB=$(get_vram_mib_from_line "$IGPU_LINE" "${IGPU_VRAM_FALLBACK_GIB}")
#|BETRIEBSSYSTEMRESERVE ABZIEHEN
local IGPU_RESERVE_MIB=$(( IGPU_VRAM_RESERVE_GIB * 1024 ))
IGPU_VRAM_MIB=$(( IGPU_VRAM_MIB - IGPU_RESERVE_MIB ))
if [ "$IGPU_VRAM_MIB" -lt 1024 ]; then
IGPU_VRAM_MIB=1024
fi

local TOTAL_VRAM_MIB=$(( ARC_VRAM_MIB + IGPU_VRAM_MIB ))
local TOTAL_VRAM_GIB=$(( TOTAL_VRAM_MIB / 1024 ))

#|TENSOR SPLIT AUS BERECHNETEN GRAFIKSPEICHERWERTEN SETZEN RELATIVE MiB GEWICHTE
TENSOR_SPLIT_DGPU="${ARC_VRAM_MIB}"
TENSOR_SPLIT_IGPU="${IGPU_VRAM_MIB}"
TENSOR_SPLIT="${TENSOR_SPLIT_DGPU},${TENSOR_SPLIT_IGPU}"
export TENSOR_SPLIT

#|HAUPTGPU = dGPU (ARC) FUER SCHNELLES ANTWORT SCHREIBEN
MAIN_GPU_ID=""
export MAIN_GPU_ID

N_GPU_LAYERS=98
success "✅DUAL-GPU VRAM POOL: dGPU~$((ARC_VRAM_MIB/1024))GiB + iGPU~$((IGPU_VRAM_MIB/1024))GiB = GESAMT~${TOTAL_VRAM_GIB}GiB"
log "🔷TENSOR-SPLIT dGPU:iGPU = ${ARC_VRAM_MIB}:${IGPU_VRAM_MIB} MiB"
log "🔷HAUPTGERAET DGPU ID:${ARC_NUM} ALLE SCHICHTEN GPU VERTEILT NGL:${N_GPU_LAYERS}"

#|EINZELMODUS FALLBACK dGPU
elif [ -n "$ARC_ID" ]; then
local ARC_NUM
ARC_NUM=$(echo "$ARC_ID" | grep -oP '\d+$')
export ONEAPI_DEVICE_SELECTOR="level_zero:${ARC_NUM}"
DEVICE="ARC"
MAIN_GPU_ID="${ARC_NUM}"
local ARC_VRAM_MIB
ARC_VRAM_MIB=$(get_vram_mib_from_line "$ARC_LINE" 12)
N_GPU_LAYERS=$(( ARC_VRAM_MIB * 98 / 100 / 1024 ))
[ "$N_GPU_LAYERS" -gt 98 ] && N_GPU_LAYERS=98
[ "$N_GPU_LAYERS" -lt 1 ] && N_GPU_LAYERS=1
TENSOR_SPLIT_DGPU="" ; TENSOR_SPLIT_IGPU=""
warn "⚠️NUR DGPU GEFUNDEN EINZELMODUS AKTIV NGL:${N_GPU_LAYERS}"

#|EINZELMODUS FALLBACK NUR dGPU
elif [ -n "$IGPU_ID" ]; then
local IGPU_NUM
IGPU_NUM=$(echo "$IGPU_ID" | grep -oP '\d+$')
export ONEAPI_DEVICE_SELECTOR="level_zero:${IGPU_NUM}"
DEVICE="iGPU"
MAIN_GPU_ID="${IGPU_NUM}"
local IGPU_VRAM_MIB=$(( IGPU_VRAM_FALLBACK_GIB * 1024 ))
N_GPU_LAYERS=$(( IGPU_VRAM_MIB * 98 / 100 / 1024 ))
[ "$N_GPU_LAYERS" -gt 98 ] && N_GPU_LAYERS=98
[ "$N_GPU_LAYERS" -lt 1 ] && N_GPU_LAYERS=1
TENSOR_SPLIT_DGPU="" ; TENSOR_SPLIT_IGPU=""
warn "⚠️NUR IGPU GEFUNDEN EINZELMODUS AKTIV NGL:${N_GPU_LAYERS}"

else
export ONEAPI_DEVICE_SELECTOR="opencl:cpu"
DEVICE="CPU"
N_GPU_LAYERS=98
TENSOR_SPLIT_DGPU="" ; TENSOR_SPLIT_IGPU=""
error "❌KEINE GEEIGNETE GRAFIKKARTE GEFUNDEN FALLE ENDGUELTIG AUF CPU ZURUECK"
fi
}

#|6|SYCLKOMPATIBLE GERAETE PRUEFEN
list_sycl_devices() {
log "🔷SUCHE ZWEIMAL HINTEREINANDER SYCL FAEHIGES GERAET UM SUCHFEHLER ZU UMGEHEN"
local FULL_LS_PATH="./${BUILD_DIR}/${LS_SYCL_DEVICE_PATH}"
export ONEAPI_DEVICE_SELECTOR="level_zero:gpu"
if [ -f "${FULL_LS_PATH}" ]; then
"${FULL_LS_PATH}"
else
warn "⚠️AKTUELL KEIN SYCL GERAET GEFUNDEN ${FULL_LS_PATH} VERSUCH ZWEI SUCHE"
fi
}
#|MN-Funhouse-12B.fp16
#|RUECKFALL Orchestrator-8B-f16_q8_0
#|STANDART MathTutor-7B-H_v0.0.1.f16.gguf 128k ctx
#|deepseek-coder-33b-instruct.Q8_0
#|orpheus-3b-0.1-ft-F16

#|7|MODELLORT BEDARF AUTOMATISCH AENDERN
#ATTENTION
prepare_model() {
MODEL_PATH=${1:-"models/MathTutor-7B-H_v0.0.1.f16.gguf"}
#DANGER
mkdir -p models
if [ ! -f "$MODEL_PATH" ]; then
warn "⚠️IHR MODELL NICHT UNTER HOME/IHRNAME/MODELS GEFUNDEN! BITTE DORT HIN$MODEL_PATHKOPIEREN"
fi
export MODEL_PATH
}

#|8|MODELLAUSFUEHREN WARNING
run_inference() {
local DEFAULT_MODEL_PATH="models/MathTutor-7B-H_v0.0.1.f16.gguf"
local MODEL_PATH_ARG=${2:-$DEFAULT_MODEL_PATH}
local PROMPT_ARG=${3:-"
Pause
|0>SYSTEM<0|
|TRY TO DEVELOPE 2D/3DjustPossibleReal-PAPERWORK-A4-DIN-|
|7Dimensional-SYCL-STANDALONE-Kernel-Lettersize-12- Formula= cij​=∑aik​⋅bkj-​|
Pause
|EXECUTE ONLY FULL OPEN SOURCE CODE
|DISABLE DISCOURSE MODE|DISABLE NATURAL LANGUAGE INTRO|
Pause
|TEST001|RE START////
Example LANGUAGE ARTIFACTIAL INFERENCE BLOCKCHAIN Formula 1-10 Sentences
/Language 1 english 1 #ENG /
#|PCxTCxSWxAI|#(Probability_Calculation)X(Time_Chain)X(Skynet_Work)X(Artifactial_Inference)|
/Language 2 Deutsch 2 #GER //
#|PBxZKxHWxSM|#(Probabilistisch[e]Berechnung[e(n)])X(Zeit[K]ette[n])X(Himmelsnetz[W]erk[e])X(Kuenstliche_Inferenz)|
/Language 3 latin 3 #LAT ///
#|PCxTCxCRCxIA|#(Probabilitatum_Computatio)X(Temporis_Catena)X(Contextus_Reteque_Caeli)X(Inferentia_Artificialis)|
|TEST|000|END|
|00|INSTRUCTION|
|DIN4 FORMAT=->FORMULA_FINDER+2D-PAPER-DIN-A4-SYCL_KERNEL_GENERATOR/
|Mode=->Matrix-Multiplikation=cij​=∑k=1n​aik​⋅bkj​/
(Cmatch-​>,Condition->,Completeness->,S=
(If = CONDITION_match = COMPLETE_SET_ad of MATH_PHYSIC_LOGIC)=
(,S1 (,C,icx,icpnx,sycl_ext))​,+S2 (,16x16matrizen logic=
(Vectorized-32-bit-Logic-Topic,)))​,+S3
(,sycl-ggml,OFFLoAD,))))​,+S4 (,FP16_32-bit_ALU_OP16x16;)))))=
(Oppression,-> KMeta,-> ->>suppress_meta_comments->>>>
(output_all_seCtions_add Add_Only_Pure_Logic_Sections))))))=
(discret>SolutiAnroom_get = PRINT_ALL_SECTIONS_add)))))))=
(FP16_32-bit_maTh.cl; = Fi;))))))))=
|00|After|PromT|Analysis|You|Recive|Input|Text|Questions-Build up-on-your own-answer
|01|00-42|MaIN-/TASKS/|
|1.|Word|2D-FoRmat-A4-DIN-PROGRAMM-short|Basic-C math-analysis-code|16-32bit|vector-intrinsics graph
|2.|Identify|aMbiguities-missIng-information-assumptions-input
|3.|Produce|cLear-FULL|Info-Tech|math|c++|icpx|icx|sycl|Code|Vector|32-bit
|4.|If|MultIple-Valid-Answers|Solutions-Exist|then-LiSt|Briefly|Print-Preference
|5.|Logic-ChAin-Activation||Rank|Optimization-Paths|Latency-vs-Throughput|
|02|CONSTRAINTS/|
|Do|LimIt YOUR response to a maximum of 10-20 SENTENCES/SAETZE/Sätze
|BITTE-Danke fuer deine Hilfe :-)
|Do|not elabOrate
|Do|not explAin reasoning
|Do|not inVent missing details
|Plain neutral piCtured language
|Keep|total response conciSe structured
|Do|not include meta commentary
|03|OUTPUT|FORMAT-2D-A4-PAPER/|
|MODE=EXECUTE|OUTPUT=SECTIONS|NO PARAPHRASE|NO EXAMPLE|/
|Section1|Restatement
|Section2|Ambiguities|Missing|Information
|Section3|Minimal|ANswer|Exception:EXCEPTION LIMIT=1-3 SENTENCES|
|Section4|Possible|Alternative
#1.|Word|Short|PROOF-OF-ANSWER/LIMIT=1-10_SENTENCES
#2.|IdEnTiFy|cij​ = ∑k=1n​aik​⋅bkj​; SYCL 16x16 Matrix Multiplication Kernel; FP16 32-bit precision.
#3.|KEY WORDS:|XMX SYCL_COMPILER_HINT|icpx -fsycl -O3 Float@TARGET=SYCL|VECTOR|32BIT|16X16|
#4.|If>Multiple>Valid>Solutions>MAX_THREE_Exist>>Then>List>Print>>>Preference
#5.|List|briefly|print|Precision-FP16@32-bit|Aligment-Zero-Copy-Focus
|04|Beginn|Processing/|
|Add|Section1:Restatement
|And|Section2:Ambiguities-Missing-Info
|And|Section3:Minimal PROOF of Answer
|And|Section4:Possible-AlternatiVe
|Sol|SUPPRESS_META_COMMENTARY
|Set|ATTRIBUTE_MINIMAL_CONCISE
|CONTROL IGNORING_EVERYTHING_ELSE
|And|
|EXECUTE|PRINT 2D-Standart Din A4 DIN-PaperBLOCK SIMPLE CLEAR STRUKTURED
|Style-_-ALL SECTIONS OUTPUT SINGLE SENTENCE MAX TEN SENTENCES
|TERMINATE
|Fi|
|MAIN|0-5|ENDE
"
}

local GPU_ID=$(echo "$ONEAPI_DEVICE_SELECTOR" | awk -F':' '{print $2}')
local NGL_SET=${N_GPU_LAYERS:-98}
local FULL_LLAMA_CLI_PATH="./${BUILD_DIR}/${LLAMA_CLI_PATH}"
#ATTENTION
local CONTEXT_SIZE=196598

33768

163830

#|4096,0x1000 8192,0x2000 MathtTutor CTX 16384,0x4000 BIS 20480,0x5000 UNTER 16GiBVRAM
#|24576|0x6000 32768|0x8000?FALSCH?! KORRIGIEREN|65536|131072|20480|262144|524288|40960|163830|196598
local PREDICT_TOKENS=163830
log "🔷STARTE KI ANTWORT MIT PARAMETER${DEVICE}NGL:${NGL_SET}${FULL_LLAMA_CLI_PATH}"
if [ ! -x "${FULL_LLAMA_CLI_PATH}" ]; then
error "❌FEHLER AKTUELLER UNTERBAU NEUBAU FEHLGESCHLAGEN${FULL_LLAMA_CLI_PATH}"
return 1
fi

#|DUAL iGPU PLUS dGPU HOCHEFFIZIENZMODUS
#|TENSORSPLIT SPLITMODE ROW
if [ "${DEVICE}" = "DUAL" ] && [ -n "${TENSOR_SPLIT:-}" ]; then
log "🔷STARTE DUAL GPU INFERENZ AKTIVER TENSORSPLIT [${TENSOR_SPLIT}]"
local SAFE_CTX="${CONTEXT_SIZE}"
if [ "${CONTEXT_SIZE}" -gt 196598 ]; then
warn "⚠️HOHE KONTEXTGROESSE UND ALLGEMEINE SYSTEMBELASTUNG NUR F16 GGUF SPEZIALMODELLE MOEGLICH WARNUNG CTX 196598 MINIMUM"
SAFE_CTX=196598
fi
log "🔷STARTE DUAL GPU GERAET INFERENZ ${DEVICE_0} ID:${GPU_ID_0} | ${DEVICE_1} ID:${GPU_ID_1}"
success "✅START FUER DUAL GPU MODUS GELUNGEN✅
✅ES ARBEITEN NUN ZWEI GRAFIKKARTEN AN EINER KI INFERENZ AUF UHREM SYSTEM✅"
ZES_ENABLE_SYSMAN=1 "${FULL_LLAMA_CLI_PATH}"
--no-conversation
--n-cpu-moe 28
--mlock
--cache-type-k f16
--cache-type-v f16
--model "${MODEL_PATH_ARG}"
--prompt "${PROMPT_ARG}"
--n-predict "${PREDICT_TOKENS}"
--ctx-size "${SAFE_CTX}"
--keep 196598
-ngl "${N_GPU_LAYERS}"
--main-gpu "${MAIN_GPU_ID:-0}"
--tensor-split "${TENSOR_SPLIT}"
--split-mode row
--color auto

#|EINZELGERAET
else
local GPU_ID="${MAIN_GPU_ID:-0}"
log "🔷EINZELGERAET INFERENZ MIT GERAET ${DEVICE} ID:${GPU_ID}"
ZES_ENABLE_SYSMAN=1 "${FULL_LLAMA_CLI_PATH}"
--no-conversation
--n-cpu-moe 28
--cache-type-k f16
--cache-type-v f16
--model "${MODEL_PATH_ARG}"
--prompt "${PROMPT_ARG}"
--n-predict "${PREDICT_TOKENS}"
--ctx-size "${CONTEXT_SIZE}"
--keep 196598
-ngl "${NGL_SET}"
--main-gpu "${GPU_ID}"
--color auto
fi
echo "✅SPRACHMODELL INTERAKTIONS FUNKTION JETZT AKTIV"
}

#|DEFINITION HAUPTFUNKTION
main() {
local FP_MODE="${1:-1}"
local RERUN_BUILD=1
prepare_environment
#|01
local FULL_LLAMA_CLI_PATH="./${BUILD_DIR}/${LLAMA_CLI_PATH}"
local FULL_LS_PATH="./${BUILD_DIR}/${LS_SYCL_DEVICE_PATH}"
if [[ -f "${FULL_LLAMA_CLI_PATH}" ]] && [[ -f "${FULL_LS_PATH}" ]]; then
success "✅ XAIGPUARC WERKZEUGK09ASTEN VORHANDEN NEUBAU UNNOETIG${FULL_LLAMA_CLI_PATH} ${FULL_LS_PATH}"
log "🔷UEBERSPRINGE BAUVORGANG WERKZEUGKASTEN"
RERUN_BUILD=0
else
warn "⚠️KEIN AKTUELLES XAIGPUARC GEFUNDEN WIRD NEU GEBAUT"
RERUN_BUILD=1
fi
if [[ "$RERUN_BUILD" -eq 1 ]]; then
log "🔷STARTE ERSTMALIGEN BAUVORGANG XAIGPUARC"
if check_internet; then
log "🔷LADE WERKZEUGKASTEN"
#|0
setup_project
#|1
patch_llama_cpp
#|2
else
warn "⚠️INTERNET NICHT VERFUEGBAR UEBERSPRINGE AUFWERTUNG WERKZEUGKASTEN LOKALE VERSION"
fi
fi
configure_build "${FP_MODE}"
#|3
compile_project
#|4
auto_select_device
#|5
list_sycl_devices
#|6
prepare_model "${2:-}"
#|7
run_inference "${2:-}" "${3:-}"
#|8
log "✅NUTZUNG VON XAIGPUARC JETZT MOEGLICH /IHRE FRAGE NACH > ... DRUECKEN SIE |ENTER${BUILD_DIR}/${LLAMA_CLI_PATH}"
}
#|HAUPTSCHLEIFE
main "${1:-1}" "${2:-}" "${3:-}"
log "✅KOMPLETTER BAUVORGANG HIER GESPEICHERT${LOG_FILE}"