mirrors/ollama-for-amd
1
0
Fork 0
mirror of https://github.com/likelovewant/ollama-for-amd.git synced 2025-12-25 16:08:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'ollama:main' into main

Browse Source
This commit is contained in:
likelovewant
2024-09-19 22:46:10 +08:00
committed by GitHub
parent 486ae433ae 5804cf1723
commit c8d6d9a010
74 changed files with 2609 additions and 1050 deletions
Download Patch File Download Diff File Expand all files Collapse all files

60
llm/ext_server/server.cpp vendored
View File

@@ -913,7 +913,9 @@ struct llama_server_context
slot.sampled = result.tok;
// search stop word and delete it
slot.generated_text += token_str;
if (!llama_token_is_eog(model, result.tok))
slot.generated_text += token_str;
slot.has_next_token = true;
if (slot.ctx_sampling->params.use_penalty_prompt_tokens && result.tok != -1)
@@ -954,30 +956,36 @@ struct llama_server_context
if (!incomplete)
{
size_t pos = std::min(slot.n_sent_text, slot.generated_text.size());
const std::string str_test = slot.generated_text.substr(pos);
bool is_stop_full = false;
size_t stop_pos = find_stopping_strings(str_test, token_str.size(), STOP_FULL, slot);
if (stop_pos != std::string::npos)
{
is_stop_full = true;
slot.generated_text.erase(
slot.generated_text.begin() + pos + stop_pos,
slot.generated_text.end());
pos = std::min(slot.n_sent_text, slot.generated_text.size());
}
else
{
is_stop_full = false;
stop_pos = find_stopping_strings(str_test, token_str.size(), STOP_PARTIAL, slot);
}
// check if there is any token to predict
if (stop_pos == std::string::npos || (!slot.has_next_token && !is_stop_full && stop_pos > 0))
{
// no send the stop word in the response
result.text_to_send = slot.generated_text.substr(pos, std::string::npos);
slot.n_sent_text += result.text_to_send.size();
// add the token to slot queue and cache
if (!llama_token_is_eog(model, result.tok)) {
const std::string str_test = slot.generated_text.substr(pos);
bool is_stop_full = false;
size_t stop_pos = find_stopping_strings(str_test, token_str.size(), STOP_FULL, slot);
if (stop_pos != std::string::npos)
{
is_stop_full = true;
slot.generated_text.erase(
slot.generated_text.begin() + pos + stop_pos,
slot.generated_text.end());
pos = std::min(slot.n_sent_text, slot.generated_text.size());
}
else
{
is_stop_full = false;
stop_pos = find_stopping_strings(str_test, token_str.size(), STOP_PARTIAL, slot);
}
// check if there is any token to predict
if (stop_pos == std::string::npos || (!slot.has_next_token && !is_stop_full && stop_pos > 0))
{
// no send the stop word in the response
result.text_to_send = slot.generated_text.substr(pos, std::string::npos);
slot.n_sent_text += result.text_to_send.size();
// add the token to slot queue and cache
}
} else {
result.text_to_send = slot.generated_text.substr(pos, std::string::npos);
slot.n_sent_text += result.text_to_send.size();
}
if (slot.params.stream)
@@ -1117,9 +1125,7 @@ struct llama_server_context
{"multimodal", multimodal}
};
if (!llama_token_is_eog(model, tkn.tok)) {
res.result_json["content"] = tkn.text_to_send;
}
res.result_json["content"] = tkn.text_to_send;
if (slot.sparams.n_probs > 0)
{

54
llm/generate/gen_common.sh
View File

@@ -31,6 +31,7 @@ init_vars() {
NO_WHOLE_ARCHIVE=""
GCC_ARCH="-arch ${ARCH}"
DIST_BASE=../../dist/darwin-${GOARCH}/
PAYLOAD_BASE=../../build/darwin/${GOARCH}
;;
"Linux")
LIB_EXT="so"
@@ -40,6 +41,7 @@ init_vars() {
# Cross compiling not supported on linux - Use docker
GCC_ARCH=""
DIST_BASE=../../dist/linux-${GOARCH}/
PAYLOAD_BASE=../../build/linux/${GOARCH}
;;
*)
;;
@@ -47,7 +49,8 @@ init_vars() {
if [ -z "${CMAKE_CUDA_ARCHITECTURES}" ] ; then
CMAKE_CUDA_ARCHITECTURES="50;52;61;70;75;80"
fi
GZIP=$(which pigz 2>/dev/null || echo "gzip")
GZIP=$(command -v pigz 2>/dev/null || echo "gzip")
RUNNER_BASE="${DIST_BASE}/lib/ollama/runners"
}
git_module_setup() {
@@ -66,22 +69,10 @@ git_module_setup() {
}
apply_patches() {
# Wire up our CMakefile
if ! grep ollama ${LLAMACPP_DIR}/CMakeLists.txt; then
echo 'add_subdirectory(../ext_server ext_server) # ollama' >>${LLAMACPP_DIR}/CMakeLists.txt
fi
if [ -n "$(ls -A ../patches/*.diff)" ]; then
# apply temporary patches until fix is upstream
for patch in ../patches/*.diff; do
for file in $(grep "^+++ " ${patch} | cut -f2 -d' ' | cut -f2- -d/); do
(cd ${LLAMACPP_DIR}; git checkout ${file})
done
done
for patch in ../patches/*.diff; do
(cd ${LLAMACPP_DIR} && git apply ${patch})
done
fi
# apply temporary patches until fix is upstream
for patch in ../patches/*.patch; do
git -c 'user.name=nobody' -c 'user.email=<>' -C ${LLAMACPP_DIR} am ${patch}
done
}
build() {
@@ -91,17 +82,34 @@ build() {
rm -f ${BUILD_DIR}/bin/ggml-common.h ${BUILD_DIR}/bin/ggml-metal.metal
}
compress() {
echo "Compressing payloads to reduce overall binary size..."
rm -rf ${BUILD_DIR}/bin/*.gz
dist() {
[ -z "${RUNNER}" ] && exit 1
mkdir -p ${RUNNER_BASE}/${RUNNER}/
for f in ${BUILD_DIR}/bin/* ; do
${GZIP} -n --best -f ${f} &
cp ${f} ${RUNNER_BASE}/${RUNNER}/
done
# check for lib directory
if [ -d ${BUILD_DIR}/lib ]; then
for f in ${BUILD_DIR}/lib/* ; do
cp ${f} ${RUNNER_BASE}/${RUNNER}/
done
fi
}
# Compress from the build $BUILD_DIR into the $PAYLOAD_BASE/$RUNNER dir
compress() {
[ -z "${RUNNER}" ] && exit 1
echo "Compressing payloads with ${GZIP} to reduce overall binary size..."
rm -rf "${PAYLOAD_BASE}/${RUNNER}/"
mkdir -p "${PAYLOAD_BASE}/${RUNNER}/"
for f in ${BUILD_DIR}/bin/* ; do
${GZIP} -c --best ${f} > "${PAYLOAD_BASE}/${RUNNER}/$(basename ${f}).gz" &
compress_pids+=" $!"
done
# check for lib directory
if [ -d ${BUILD_DIR}/lib ]; then
for f in ${BUILD_DIR}/lib/* ; do
${GZIP} -n --best -f ${f} &
${GZIP} -c --best ${f} > "${PAYLOAD_BASE}/${RUNNER}/$(basename ${f}).gz" &
compress_pids+=" $!"
done
fi
@@ -117,7 +125,7 @@ wait_for_compress() {
install() {
echo "Installing libraries to bin dir ${BUILD_DIR}/bin/"
for lib in $(find ${BUILD_DIR} -name \*.${LIB_EXT}); do
for lib in $(find ${BUILD_DIR} -name \*.${LIB_EXT} | grep -v "${BUILD_DIR}/bin/" ); do
rm -f "${BUILD_DIR}/bin/$(basename ${lib})"
cp -af "${lib}" "${BUILD_DIR}/bin/"
done

12
llm/generate/gen_darwin.sh
View File

@@ -39,7 +39,8 @@ case "${GOARCH}" in
#
init_vars
CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_ACCELERATE=off -DGGML_BLAS=off -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
BUILD_DIR="../build/darwin/${ARCH}/cpu"
RUNNER=cpu
BUILD_DIR="../build/darwin/${GOARCH}/${RUNNER}"
echo "Building LCD CPU"
build
sign ${BUILD_DIR}/bin/ollama_llama_server
@@ -51,7 +52,8 @@ case "${GOARCH}" in
#
init_vars
CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_ACCELERATE=off -DGGML_BLAS=off -DGGML_AVX=on -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
BUILD_DIR="../build/darwin/${ARCH}/cpu_avx"
RUNNER=cpu_avx
BUILD_DIR="../build/darwin/${GOARCH}/${RUNNER}"
echo "Building AVX CPU"
build
sign ${BUILD_DIR}/bin/ollama_llama_server
@@ -63,7 +65,8 @@ case "${GOARCH}" in
#
init_vars
CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_ACCELERATE=on -DGGML_BLAS=off -DGGML_AVX=on -DGGML_AVX2=on -DGGML_AVX512=off -DGGML_FMA=on -DGGML_F16C=on ${CMAKE_DEFS}"
BUILD_DIR="../build/darwin/${ARCH}/cpu_avx2"
RUNNER=cpu_avx2
BUILD_DIR="../build/darwin/${GOARCH}/${RUNNER}"
echo "Building AVX2 CPU"
EXTRA_LIBS="${EXTRA_LIBS} -framework Accelerate -framework Foundation"
build
@@ -84,7 +87,8 @@ case "${GOARCH}" in
if [ -z "$OLLAMA_SKIP_METAL_GENERATE" ]; then
init_vars
CMAKE_DEFS="${COMMON_DARWIN_DEFS} -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} ${CMAKE_DEFS}"
BUILD_DIR="../build/darwin/${ARCH}/metal"
RUNNER="metal"
BUILD_DIR="../build/darwin/${GOARCH}/${RUNNER}"
EXTRA_LIBS="${EXTRA_LIBS} -framework Accelerate -framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders"
build
sign ${BUILD_DIR}/bin/ollama_llama_server

35
llm/generate/gen_linux.sh
View File

@@ -88,10 +88,12 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
init_vars
echo "OLLAMA_CUSTOM_CPU_DEFS=\"${OLLAMA_CUSTOM_CPU_DEFS}\""
CMAKE_DEFS="${OLLAMA_CUSTOM_CPU_DEFS} -DBUILD_SHARED_LIBS=on -DCMAKE_POSITION_INDEPENDENT_CODE=on ${CMAKE_DEFS}"
BUILD_DIR="../build/linux/${ARCH}/cpu"
RUNNER="cpu"
BUILD_DIR="../build/linux/${GOARCH}/${RUNNER}"
echo "Building custom CPU"
build
install
dist
compress
else
# Darwin Rosetta x86 emulation does NOT support AVX, AVX2, AVX512
@@ -111,10 +113,12 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
#
init_vars
CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
BUILD_DIR="../build/linux/${ARCH}/cpu"
RUNNER=cpu
BUILD_DIR="../build/linux/${GOARCH}/${RUNNER}"
echo "Building LCD CPU"
build
install
dist
compress
fi
@@ -129,10 +133,12 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
#
init_vars
CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_AVX=on -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
BUILD_DIR="../build/linux/${ARCH}/cpu_avx"
RUNNER=cpu_avx
BUILD_DIR="../build/linux/${GOARCH}/${RUNNER}"
echo "Building AVX CPU"
build
install
dist
compress
fi
@@ -143,10 +149,12 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
#
init_vars
CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_AVX=on -DGGML_AVX2=on -DGGML_AVX512=off -DGGML_FMA=on -DGGML_F16C=on ${CMAKE_DEFS}"
BUILD_DIR="../build/linux/${ARCH}/cpu_avx2"
RUNNER=cpu_avx2
BUILD_DIR="../build/linux/${GOARCH}/${RUNNER}"
echo "Building AVX2 CPU"
build
install
dist
compress
fi
fi
@@ -196,11 +204,13 @@ if [ -z "${OLLAMA_SKIP_CUDA_GENERATE}" -a -d "${CUDA_LIB_DIR}" ]; then
fi
export CUDAFLAGS="-t8"
CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} ${ARM64_DEFS} ${CMAKE_CUDA_DEFS} -DGGML_STATIC=off"
BUILD_DIR="../build/linux/${ARCH}/cuda${CUDA_VARIANT}"
RUNNER=cuda${CUDA_VARIANT}
BUILD_DIR="../build/linux/${GOARCH}/${RUNNER}"
export LLAMA_SERVER_LDFLAGS="-L${CUDA_LIB_DIR} -lcudart -lcublas -lcublasLt -lcuda"
CUDA_DIST_DIR="${CUDA_DIST_DIR:-${DIST_BASE}/lib/ollama}"
build
install
dist
echo "Installing CUDA dependencies in ${CUDA_DIST_DIR}"
mkdir -p "${CUDA_DIST_DIR}"
for lib in ${CUDA_LIB_DIR}/libcudart.so* ${CUDA_LIB_DIR}/libcublas.so* ${CUDA_LIB_DIR}/libcublasLt.so* ; do
@@ -221,7 +231,8 @@ if [ -z "${OLLAMA_SKIP_ONEAPI_GENERATE}" -a -d "${ONEAPI_ROOT}" ]; then
source ${ONEAPI_ROOT}/setvars.sh --force # set up environment variables for oneAPI
CC=icx
CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL=ON -DGGML_SYCL_F16=OFF"
BUILD_DIR="../build/linux/${ARCH}/oneapi"
RUNNER=oneapi
BUILD_DIR="../build/linux/${GOARCH}/${RUNNER}"
ONEAPI_DIST_DIR="${DIST_BASE}/lib/ollama"
export LLAMA_SERVER_LDFLAGS="-fsycl -lOpenCL -lmkl_core -lmkl_sycl_blas -lmkl_intel_ilp64 -lmkl_tbb_thread -ltbb"
DEBUG_FLAGS="" # icx compiles with -O0 if we pass -g, so we must remove it
@@ -240,6 +251,7 @@ if [ -z "${OLLAMA_SKIP_ONEAPI_GENERATE}" -a -d "${ONEAPI_ROOT}" ]; then
cp "${ONEAPI_ROOT}/compiler/latest/lib/libsvml.so" "${ONEAPI_DIST_DIR}"
cp "${ONEAPI_ROOT}/compiler/latest/lib/libur_loader.so.0" "${ONEAPI_DIST_DIR}"
install
dist
compress
fi
@@ -268,7 +280,8 @@ if [ -z "${OLLAMA_SKIP_ROCM_GENERATE}" -a -d "${ROCM_PATH}" ]; then
CMAKE_DEFS="${CMAKE_DEFS} ${OLLAMA_CUSTOM_ROCM_DEFS}"
echo "Building custom ROCM GPU"
fi
BUILD_DIR="../build/linux/${ARCH}/rocm${ROCM_VARIANT}"
RUNNER=rocm${ROCM_VARIANT}
BUILD_DIR="../build/linux/${GOARCH}/${RUNNER}"
# ROCm dependencies are too large to fit into a unified bundle
ROCM_DIST_DIR="${DIST_BASE}/../linux-${GOARCH}-rocm/lib/ollama"
# TODO figure out how to disable runpath (rpath)
@@ -278,13 +291,17 @@ if [ -z "${OLLAMA_SKIP_ROCM_GENERATE}" -a -d "${ROCM_PATH}" ]; then
# copy the ROCM dependencies
mkdir -p "${ROCM_DIST_DIR}"
for dep in $(ldd "${BUILD_DIR}/bin/ollama_llama_server" | grep "=>" | cut -f2 -d= | cut -f2 -d' ' | grep -v "${ARCH}/rocm${ROCM_VARIANT}" | grep -e rocm -e amdgpu -e libtinfo ); do
for dep in $(ldd "${BUILD_DIR}/bin/ollama_llama_server" | grep "=>" | cut -f2 -d= | cut -f2 -d' ' | grep -v "${GOARCH}/rocm${ROCM_VARIANT}" | grep -e rocm -e amdgpu -e libtinfo -e libnuma -e libelf ); do
cp -a "${dep}"* "${ROCM_DIST_DIR}"
if [ $(readlink -f "${dep}") != "${dep}" ] ; then
cp $(readlink -f "${dep}") "${ROCM_DIST_DIR}"
fi
done
install
dist
compress
fi
cleanup
wait_for_compress
echo "go generate completed. LLM runners: $(cd ${BUILD_DIR}/..; echo *)"
echo "go generate completed. LLM runners: $(cd ${PAYLOAD_BASE}; echo *)"

24
llm/generate/gen_windows.ps1
View File

@@ -101,29 +101,9 @@ function git_module_setup {
}
function apply_patches {
# Wire up our CMakefile
if (!(Select-String -Path "${script:llamacppDir}/CMakeLists.txt" -Pattern 'ollama')) {
Add-Content -Path "${script:llamacppDir}/CMakeLists.txt" -Value 'add_subdirectory(../ext_server ext_server) # ollama'
}
# Apply temporary patches until fix is upstream
$patches = Get-ChildItem "../patches/*.diff"
foreach ($patch in $patches) {
# Extract file paths from the patch file
$filePaths = Get-Content $patch.FullName | Where-Object { $_ -match '^\+\+\+ ' } | ForEach-Object {
$parts = $_ -split ' '
($parts[1] -split '/', 2)[1]
}
# Checkout each file
foreach ($file in $filePaths) {
git -C "${script:llamacppDir}" checkout $file
}
}
# Apply each patch
foreach ($patch in $patches) {
git -C "${script:llamacppDir}" apply $patch.FullName
foreach ($patch in $(Get-ChildItem "../patches/*.patch")) {
git -c 'user.name=nobody' -c 'user.email=<>' -C "${script:llamacppDir}" am $patch.FullName
}
}

4
llm/llm_darwin_arm64.go → llm/llm_darwin.go
View File

@@ -1,11 +1,7 @@
package llm
import (
"embed"
"syscall"
)
//go:embed build/darwin/arm64/*/bin/*
var libEmbed embed.FS
var LlamaServerSysProcAttr = &syscall.SysProcAttr{}

11
llm/llm_darwin_amd64.go
View File

@@ -1,11 +0,0 @@
package llm
import (
"embed"
"syscall"
)
//go:embed build/darwin/x86_64/*/bin/*
var libEmbed embed.FS
var LlamaServerSysProcAttr = &syscall.SysProcAttr{}

4
llm/llm_linux.go
View File

@@ -1,11 +1,7 @@
package llm
import (
"embed"
"syscall"
)
//go:embed build/linux/*/*/bin/*
var libEmbed embed.FS
var LlamaServerSysProcAttr = &syscall.SysProcAttr{}

4
llm/llm_windows.go
View File

@@ -1,13 +1,9 @@
package llm
import (
"embed"
"syscall"
)
// unused on windows
var libEmbed embed.FS
const CREATE_DEFAULT_ERROR_MODE = 0x04000000
var LlamaServerSysProcAttr = &syscall.SysProcAttr{

22
llm/patches/0000-cmakelist.patch Normal file
View File

@@ -0,0 +1,22 @@
From 8b8d83ffca775840acc5dc700f3b3703e9f5cfe4 Mon Sep 17 00:00:00 2001
From: Michael Yang <mxyng@pm.me>
Date: Fri, 23 Aug 2024 11:27:48 -0700
Subject: [PATCH] patch cmakelist
---
CMakeLists.txt | 2 ++
1 file changed, 2 insertions(+)
diff --git a/CMakeLists.txt b/CMakeLists.txt
index a3132063..6a2a9912 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -199,3 +199,5 @@ if (LLAMA_BUILD_EXAMPLES)
add_subdirectory(examples)
add_subdirectory(pocs)
endif()
+
+add_subdirectory(../ext_server ext_server) # ollama
--
2.45.2

21
llm/patches/01-load-progress.diff → llm/patches/0001-load-progress.patch
View File

@@ -1,8 +1,18 @@
From 2cfaa0a04faa9c87ba8f1ac8527eb953e69c6cde Mon Sep 17 00:00:00 2001
From: Michael Yang <mxyng@pm.me>
Date: Mon, 16 Sep 2024 15:53:10 -0700
Subject: [PATCH] 01-load-progress.diff
---
common/common.cpp | 2 ++
common/common.h | 7 +++++++
2 files changed, 9 insertions(+)
diff --git a/common/common.cpp b/common/common.cpp
index 2c05a4d4..927f0e3d 100644
index 9fa18472..48ff41e9 100644
--- a/common/common.cpp
+++ b/common/common.cpp
@@ -2093,6 +2093,8 @@ struct llama_model_params llama_model_params_from_gpt_params(const gpt_params &
@@ -2573,6 +2573,8 @@ struct llama_model_params llama_model_params_from_gpt_params(const gpt_params &
mparams.use_mmap = params.use_mmap;
mparams.use_mlock = params.use_mlock;
mparams.check_tensors = params.check_tensors;
@@ -12,10 +22,10 @@ index 2c05a4d4..927f0e3d 100644
mparams.kv_overrides = NULL;
} else {
diff --git a/common/common.h b/common/common.h
index 65c0ef81..ebca2c77 100644
index cb5e7f6d..d8f043f7 100644
--- a/common/common.h
+++ b/common/common.h
@@ -184,6 +184,13 @@ struct gpt_params {
@@ -204,6 +204,13 @@ struct gpt_params {
std::string mmproj = ""; // path to multimodal projector
std::vector<std::string> image; // path to image file(s)
@@ -29,3 +39,6 @@ index 65c0ef81..ebca2c77 100644
// embedding
bool embedding = false; // get only sentence embedding
int32_t embd_normalize = 2; // normalisation for embendings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
--
2.46.0

14
llm/patches/02-clip-log.diff → llm/patches/0002-clip-log.patch
View File

@@ -1,5 +1,14 @@
From ba4bba80a744f76ac67b8234451c259a3c5da83b Mon Sep 17 00:00:00 2001
From: Michael Yang <mxyng@pm.me>
Date: Mon, 16 Sep 2024 15:53:11 -0700
Subject: [PATCH] 02-clip-log.diff
---
examples/llava/clip.cpp | 1 +
1 file changed, 1 insertion(+)
diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp
index e431c7f7..f077e688 100644
index 9b890571..cb51793d 100644
--- a/examples/llava/clip.cpp
+++ b/examples/llava/clip.cpp
@@ -3,6 +3,7 @@
@@ -10,3 +19,6 @@ index e431c7f7..f077e688 100644
#include "log.h"
#include "ggml.h"
#include "ggml-alloc.h"
--
2.46.0

18
llm/patches/03-load_exception.diff → llm/patches/0003-load_exception.patch
View File

@@ -1,8 +1,17 @@
From e43bfd3f607a6dfcaba2d490d35f412a52e55e30 Mon Sep 17 00:00:00 2001
From: Michael Yang <mxyng@pm.me>
Date: Mon, 16 Sep 2024 15:53:12 -0700
Subject: [PATCH] 03-load_exception.diff
---
src/llama.cpp | 25 ++++++++++++++++---------
1 file changed, 16 insertions(+), 9 deletions(-)
diff --git a/src/llama.cpp b/src/llama.cpp
index 73f52435..58a00fb1 100644
index 88355971..926bb71a 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -7241,7 +7241,7 @@ static int llama_model_load(const std::string & fname, llama_model & model, llam
@@ -8635,7 +8635,7 @@ static int llama_model_load(const std::string & fname, llama_model & model, llam
}
} catch (const std::exception & err) {
LLAMA_LOG_ERROR("%s: error loading model: %s\n", __func__, err.what());
@@ -11,7 +20,7 @@ index 73f52435..58a00fb1 100644
}
return 0;
@@ -17564,16 +17564,23 @@ struct llama_model * llama_load_model_from_file(
@@ -18022,16 +18022,23 @@ struct llama_model * llama_load_model_from_file(
}
model->rpc_servers.push_back(servers);
}
@@ -43,3 +52,6 @@ index 73f52435..58a00fb1 100644
}
return model;
--
2.46.0

16
llm/patches/04-metal.diff → llm/patches/0004-metal.patch
View File

@@ -1,8 +1,17 @@
From 29411d9a9d2b6a0af6425ffe88498f17f71f7d5d Mon Sep 17 00:00:00 2001
From: Michael Yang <mxyng@pm.me>
Date: Mon, 16 Sep 2024 15:53:12 -0700
Subject: [PATCH] 04-metal.diff
---
ggml/src/ggml-metal.m | 30 +++++++++++++-----------------
1 file changed, 13 insertions(+), 17 deletions(-)
diff --git a/ggml/src/ggml-metal.m b/ggml/src/ggml-metal.m
index 0207b787..b5e9884b 100644
index 91b5e61b..9cfa72ac 100644
--- a/ggml/src/ggml-metal.m
+++ b/ggml/src/ggml-metal.m
@@ -1396,27 +1396,23 @@ static enum ggml_status ggml_metal_graph_compute(
@@ -1734,27 +1734,23 @@ static enum ggml_status ggml_metal_graph_compute(
// to the matrix-vector kernel
int ne11_mm_min = 1;
@@ -43,3 +52,6 @@ index 0207b787..b5e9884b 100644
// for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
// AMD GPU and older A-chips will reuse matrix-vector multiplication kernel
--
2.46.0

14
llm/patches/05-default-pretokenizer.diff → llm/patches/0005-default-pretokenizer.patch
View File

@@ -1,5 +1,14 @@
From b298ac8614d1e38da28f760eb1d2ae8af0fbbe62 Mon Sep 17 00:00:00 2001
From: Michael Yang <mxyng@pm.me>
Date: Mon, 16 Sep 2024 15:53:13 -0700
Subject: [PATCH] 05-default-pretokenizer.diff
---
src/llama.cpp | 14 +++-----------
1 file changed, 3 insertions(+), 11 deletions(-)
diff --git a/src/llama.cpp b/src/llama.cpp
index 88355971..dd7d41ed 100644
index 926bb71a..d1e959fc 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -6083,16 +6083,7 @@ static void llm_load_vocab(
@@ -30,3 +39,6 @@ index 88355971..dd7d41ed 100644
}
} else if (vocab.type == LLAMA_VOCAB_TYPE_SPM) {
vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
--
2.46.0

18
llm/patches/06-embeddings.diff → llm/patches/0006-embeddings.patch
View File

@@ -1,8 +1,17 @@
From c9a6ca9fc039233dee746a4da9705762cd9e515d Mon Sep 17 00:00:00 2001
From: Michael Yang <mxyng@pm.me>
Date: Mon, 16 Sep 2024 15:53:14 -0700
Subject: [PATCH] 06-embeddings.diff
---
src/llama.cpp | 17 ++++++++++-------
1 file changed, 10 insertions(+), 7 deletions(-)
diff --git a/src/llama.cpp b/src/llama.cpp
index 88355971..d7db689b 100644
index d1e959fc..f79bd782 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -15906,7 +15906,7 @@ static size_t llama_output_reserve(llama_context & lctx, size_t n_outputs) {
@@ -15898,7 +15898,7 @@ static size_t llama_output_reserve(llama_context & lctx, size_t n_outputs) {
const auto n_embd = hparams.n_embd;
// TODO: use a per-batch flag for logits presence instead
@@ -11,7 +20,7 @@ index 88355971..d7db689b 100644
const bool has_embd = cparams.embeddings && (cparams.pooling_type == LLAMA_POOLING_TYPE_NONE);
const size_t logits_size = has_logits ? n_vocab*n_outputs_max : 0;
@@ -16175,20 +16175,23 @@ static int llama_decode_internal(
@@ -16167,20 +16167,23 @@ static int llama_decode_internal(
// no output
res = nullptr;
embd = nullptr;
@@ -41,3 +50,6 @@ index 88355971..d7db689b 100644
// LLAMA_LOG_INFO("graph build time: %.3f ms (%d nodes, %d leafs)\n", (ggml_time_us() - t_start_us)/1000.0, gf->n_nodes, gf->n_leafs);
ggml_backend_sched_alloc_graph(lctx.sched, gf);
--
2.46.0

18
llm/patches/07-clip-unicode.diff → llm/patches/0007-clip-unicode.patch
View File

@@ -1,8 +1,17 @@
From ae2b188a679c83ce105aa1e823499441dfab3c57 Mon Sep 17 00:00:00 2001
From: Michael Yang <mxyng@pm.me>
Date: Mon, 16 Sep 2024 15:53:15 -0700
Subject: [PATCH] 07-clip-unicode.diff
---
examples/llava/clip.cpp | 23 +++++++++++++++++++++++
1 file changed, 23 insertions(+)
diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp
index 95fbe3d0..5a02a6ec 100644
index cb51793d..8716472b 100644
--- a/examples/llava/clip.cpp
+++ b/examples/llava/clip.cpp
@@ -32,6 +33,14 @@
@@ -41,6 +41,14 @@
#include <cinttypes>
#include <limits>
@@ -17,7 +26,7 @@ index 95fbe3d0..5a02a6ec 100644
//#define CLIP_DEBUG_FUNCTIONS
// RGB uint8 image
@@ -1055,7 +1064,22 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
@@ -1223,7 +1231,22 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
return nullptr;
}
@@ -40,3 +49,6 @@ index 95fbe3d0..5a02a6ec 100644
if (!fin) {
LOG_TEE("cannot open model file for loading tensors\n");
clip_free(new_clip);
--
2.46.0

402
llm/patches/0008-solar-pro.patch Normal file
View File

@@ -0,0 +1,402 @@
From 8313ce5f43f11f3d84f352f97f3802792e90e18c Mon Sep 17 00:00:00 2001
From: Michael Yang <mxyng@pm.me>
Date: Mon, 16 Sep 2024 15:53:16 -0700
Subject: [PATCH] add solar-pro support
solar-pro introduces block skip connections where blocks are connected
to other, non-sequential blocks with a scale multiple
this change adds 4 new keys to store the skip connections and one new
tensor to store the scalar. the scalar is implemented a 1-dimensional
tensor with 2 elements dervied from the model's bskcn_tv configuration.
in general, the values are (bskcn_tv, 1 - bskcn_tv)
---
src/llama.cpp | 267 +++++++++++++++++++++++++++++++++++++++++++++++---
1 file changed, 254 insertions(+), 13 deletions(-)
diff --git a/src/llama.cpp b/src/llama.cpp
index f79bd782..b7771f53 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -213,6 +213,7 @@ enum llm_arch {
LLM_ARCH_NEMOTRON,
LLM_ARCH_EXAONE,
LLM_ARCH_RWKV6,
+ LLM_ARCH_SOLAR,
LLM_ARCH_UNKNOWN,
};
@@ -261,6 +262,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_NEMOTRON, "nemotron" },
{ LLM_ARCH_EXAONE, "exaone" },
{ LLM_ARCH_RWKV6, "rwkv6" },
+ { LLM_ARCH_SOLAR, "solar" },
{ LLM_ARCH_UNKNOWN, "(unknown)" },
};
@@ -314,6 +316,7 @@ enum llm_kv {
LLM_KV_ATTENTION_KV_LORA_RANK,
LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,
LLM_KV_ATTENTION_SLIDING_WINDOW,
+ LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,
LLM_KV_ROPE_DIMENSION_COUNT,
LLM_KV_ROPE_FREQ_BASE,
@@ -405,19 +408,20 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
{ LLM_KV_TIME_MIX_EXTRA_DIM, "%s.time_mix_extra_dim" },
{ LLM_KV_TIME_DECAY_EXTRA_DIM, "%s.time_decay_extra_dim" },
- { LLM_KV_ATTENTION_HEAD_COUNT, "%s.attention.head_count" },
- { LLM_KV_ATTENTION_HEAD_COUNT_KV, "%s.attention.head_count_kv" },
- { LLM_KV_ATTENTION_MAX_ALIBI_BIAS, "%s.attention.max_alibi_bias" },
- { LLM_KV_ATTENTION_CLAMP_KQV, "%s.attention.clamp_kqv" },
- { LLM_KV_ATTENTION_KEY_LENGTH, "%s.attention.key_length" },
- { LLM_KV_ATTENTION_VALUE_LENGTH, "%s.attention.value_length" },
- { LLM_KV_ATTENTION_LAYERNORM_EPS, "%s.attention.layer_norm_epsilon" },
- { LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, "%s.attention.layer_norm_rms_epsilon" },
- { LLM_KV_ATTENTION_CAUSAL, "%s.attention.causal" },
- { LLM_KV_ATTENTION_Q_LORA_RANK, "%s.attention.q_lora_rank" },
- { LLM_KV_ATTENTION_KV_LORA_RANK, "%s.attention.kv_lora_rank" },
- { LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, "%s.attention.relative_buckets_count" },
- { LLM_KV_ATTENTION_SLIDING_WINDOW, "%s.attention.sliding_window" },
+ { LLM_KV_ATTENTION_HEAD_COUNT, "%s.attention.head_count" },
+ { LLM_KV_ATTENTION_HEAD_COUNT_KV, "%s.attention.head_count_kv" },
+ { LLM_KV_ATTENTION_MAX_ALIBI_BIAS, "%s.attention.max_alibi_bias" },
+ { LLM_KV_ATTENTION_CLAMP_KQV, "%s.attention.clamp_kqv" },
+ { LLM_KV_ATTENTION_KEY_LENGTH, "%s.attention.key_length" },
+ { LLM_KV_ATTENTION_VALUE_LENGTH, "%s.attention.value_length" },
+ { LLM_KV_ATTENTION_LAYERNORM_EPS, "%s.attention.layer_norm_epsilon" },
+ { LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, "%s.attention.layer_norm_rms_epsilon" },
+ { LLM_KV_ATTENTION_CAUSAL, "%s.attention.causal" },
+ { LLM_KV_ATTENTION_Q_LORA_RANK, "%s.attention.q_lora_rank" },
+ { LLM_KV_ATTENTION_KV_LORA_RANK, "%s.attention.kv_lora_rank" },
+ { LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, "%s.attention.relative_buckets_count" },
+ { LLM_KV_ATTENTION_SLIDING_WINDOW, "%s.attention.sliding_window" },
+ { LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION, "%s.attention.block_skip_connection.%d" },
{ LLM_KV_ROPE_DIMENSION_COUNT, "%s.rope.dimension_count" },
{ LLM_KV_ROPE_FREQ_BASE, "%s.rope.freq_base" },
@@ -589,6 +593,7 @@ enum llm_tensor {
LLM_TENSOR_ENC_FFN_DOWN,
LLM_TENSOR_ENC_FFN_UP,
LLM_TENSOR_ENC_OUTPUT_NORM,
+ LLM_TENSOR_BSKCN_TV,
};
static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES = {
@@ -1408,6 +1413,24 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
{ LLM_TENSOR_CHANNEL_MIX_RECEPTANCE, "blk.%d.channel_mix_receptance" },
},
},
+ {
+ LLM_ARCH_SOLAR,
+ {
+ { LLM_TENSOR_TOKEN_EMBD, "token_embd" },
+ { LLM_TENSOR_OUTPUT_NORM, "output_norm" },
+ { LLM_TENSOR_OUTPUT, "output" },
+ { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
+ { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
+ { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
+ { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
+ { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
+ { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
+ { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" },
+ { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
+ { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
+ { LLM_TENSOR_BSKCN_TV, "bskcn_tv" },
+ },
+ },
{
LLM_ARCH_UNKNOWN,
{
@@ -2237,6 +2260,7 @@ enum e_model {
MODEL_15B,
MODEL_16B,
MODEL_20B,
+ MODEL_22B,
MODEL_30B,
MODEL_34B,
MODEL_35B,
@@ -2284,6 +2308,8 @@ struct llama_hparams {
std::array<uint32_t, LLAMA_MAX_LAYERS> n_head_kv_arr;
std::array<uint32_t, LLAMA_MAX_LAYERS> n_ff_arr;
+ std::array<std::array<uint32_t, LLAMA_MAX_LAYERS>, 4> n_bskcn_arr;
+
uint32_t n_layer_dense_lead = 0;
uint32_t n_lora_q = 0;
uint32_t n_lora_kv = 0;
@@ -2349,6 +2375,7 @@ struct llama_hparams {
if (this->n_head_arr != other.n_head_arr) return true;
if (this->n_head_kv_arr != other.n_head_kv_arr) return true;
if (this->n_ff_arr != other.n_ff_arr) return true;
+ if (this->n_bskcn_arr != other.n_bskcn_arr) return true;
if (this->n_rel_attn_bkts != other.n_rel_attn_bkts) return true;
if (this->n_layer_dense_lead != other.n_layer_dense_lead) return true;
@@ -2455,6 +2482,14 @@ struct llama_hparams {
return ssm_d_state * ssm_d_inner;
}
}
+
+ bool n_bskcn(uint32_t n, uint32_t il = 0) const {
+ if (il < n_layer) {
+ return n_bskcn_arr[n][il] > 0;
+ }
+
+ GGML_ABORT("fatal error");
+ }
};
static_assert(std::is_trivially_copyable<llama_hparams>::value, "llama_hparams must be trivially copyable");
@@ -2635,6 +2670,8 @@ struct llama_layer {
struct ggml_tensor * ffn_gate_scale;
struct ggml_tensor * ffn_up_scale;
struct ggml_tensor * ffn_down_scale;
+
+ struct ggml_tensor * bskcn_tv;
};
// very similar to llama_batch,
@@ -5937,6 +5974,21 @@ static void llm_load_hparams(
default: model.type = e_model::MODEL_UNKNOWN;
}
} break;
+ case LLM_ARCH_SOLAR:
+ {
+ ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+ for (int i = 0; i < hparams.n_bskcn_arr.max_size(); ++i) {
+ auto & bskcn = hparams.n_bskcn_arr.at(i);
+ bskcn.fill(0);
+ ml.get_key_or_arr(::format(LLM_KV_NAMES.at(LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION), LLM_ARCH_NAMES.at(ml.llm_kv.arch), i), bskcn, hparams.n_layer, false);
+ }
+
+ switch (hparams.n_layer) {
+ case 64: model.type = e_model::MODEL_22B; break;
+ default: model.type = e_model::MODEL_UNKNOWN;
+ }
+ }
default: (void)0;
}
@@ -8420,6 +8472,38 @@ static bool llm_load_tensors(
}
} break;
+ case LLM_ARCH_SOLAR:
+ {
+ model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+
+ // output
+ {
+ model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+ model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
+ }
+
+ for (int i = 0; i < n_layer; ++i) {
+ ggml_context * ctx_layer = ctx_for_layer(i);
+ ggml_context * ctx_split = ctx_for_layer_split(i);
+
+ auto & layer = model.layers[i];
+
+ layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+
+ layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * n_head});
+ layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa});
+ layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa});
+ layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd});
+
+ layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+
+ layer.bskcn_tv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_BSKCN_TV, "weight"), {2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
+
+ layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff});
+ layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd});
+ layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff});
+ }
+ } break;
default:
throw std::runtime_error("unknown architecture");
}
@@ -15173,6 +15257,158 @@ struct llm_build_context {
return gf;
}
+
+ ggml_cgraph * build_solar() {
+ struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
+
+ // mutable variable, needed during the last layer of the computation to skip unused tokens
+ int32_t n_tokens = this->n_tokens;
+
+ const int64_t n_embd_head = hparams.n_embd_head_v;
+ GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+ GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+ struct ggml_tensor * cur;
+ struct ggml_tensor * inpL;
+
+ inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+ // inp_pos - contains the positions
+ struct ggml_tensor * inp_pos = build_inp_pos();
+
+ // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+ struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+ struct ggml_tensor * bskcn_1;
+ struct ggml_tensor * bskcn_2;
+
+ for (int il = 0; il < n_layer; ++il) {
+ struct ggml_tensor * inpSA = inpL;
+
+ if (hparams.n_bskcn(0, il)) {
+ bskcn_1 = inpSA;
+ }
+
+ if (hparams.n_bskcn(1, il)) {
+ bskcn_2 = inpSA;
+ }
+
+ if (hparams.n_bskcn(2, il)) {
+ inpSA = ggml_add(
+ ctx0,
+ ggml_mul(ctx0, bskcn_1, ggml_view_1d(ctx0, model.layers[il].bskcn_tv, 1, 0)),
+ ggml_mul(ctx0, inpSA, ggml_view_1d(ctx0, model.layers[il].bskcn_tv, 1, ggml_element_size(model.layers[il].bskcn_tv))));
+ }
+
+ if (hparams.n_bskcn(3, il)) {
+ inpSA = ggml_add(
+ ctx0,
+ ggml_mul(ctx0, bskcn_2, ggml_view_1d(ctx0, model.layers[il].bskcn_tv, 1, 0)),
+ ggml_mul(ctx0, inpSA, ggml_view_1d(ctx0, model.layers[il].bskcn_tv, 1, ggml_element_size(model.layers[il].bskcn_tv))));
+ }
+
+ // norm
+ cur = llm_build_norm(ctx0, inpL, hparams,
+ model.layers[il].attn_norm, NULL,
+ LLM_NORM_RMS, cb, il);
+ cb(cur, "attn_norm", il);
+
+ // self-attention
+ {
+ // rope freq factors for llama3; may return nullptr for llama2 and other models
+ struct ggml_tensor * rope_factors = build_rope_factors(il);
+
+ // compute Q and K and RoPE them
+ struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
+ cb(Qcur, "Qcur", il);
+ if (model.layers[il].bq) {
+ Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+ cb(Qcur, "Qcur", il);
+ }
+
+ struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
+ cb(Kcur, "Kcur", il);
+ if (model.layers[il].bk) {
+ Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+ cb(Kcur, "Kcur", il);
+ }
+
+ struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
+ cb(Vcur, "Vcur", il);
+ if (model.layers[il].bv) {
+ Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+ cb(Vcur, "Vcur", il);
+ }
+
+ Qcur = ggml_rope_ext(
+ ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, rope_factors,
+ n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+ ext_factor, attn_factor, beta_fast, beta_slow
+ );
+ cb(Qcur, "Qcur", il);
+
+ Kcur = ggml_rope_ext(
+ ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, rope_factors,
+ n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+ ext_factor, attn_factor, beta_fast, beta_slow
+ );
+ cb(Kcur, "Kcur", il);
+
+ cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+ model.layers[il].wo, model.layers[il].bo,
+ Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+ }
+
+ if (il == n_layer - 1) {
+ // skip computing output for unused tokens
+ struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+ n_tokens = n_outputs;
+ cur = ggml_get_rows(ctx0, cur, inp_out_ids);
+ inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+ }
+
+ struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+ cb(ffn_inp, "ffn_inp", il);
+
+ // feed-forward network
+ cur = llm_build_norm(ctx0, ffn_inp, hparams,
+ model.layers[il].ffn_norm, NULL,
+ LLM_NORM_RMS, cb, il);
+ cb(cur, "ffn_norm", il);
+
+ cur = llm_build_ffn(ctx0, lctx, cur,
+ model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
+ model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+ model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+ NULL,
+ LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+ cb(cur, "ffn_out", il);
+
+ cur = ggml_add(ctx0, cur, ffn_inp);
+ cb(cur, "ffn_out", il);
+
+ cur = lctx.cvec.apply_to(ctx0, cur, il);
+ cb(cur, "l_out", il);
+
+ // input for next layer
+ inpL = cur;
+ }
+
+ cur = inpL;
+
+ cur = llm_build_norm(ctx0, cur, hparams,
+ model.output_norm, NULL,
+ LLM_NORM_RMS, cb, -1);
+ cb(cur, "result_norm", -1);
+
+ // lm_head
+ cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+ cb(cur, "result_output", -1);
+
+ ggml_build_forward_expand(gf, cur);
+
+ return gf;
+ }
};
static struct ggml_cgraph * llama_build_graph_defrag(llama_context & lctx, const std::vector<uint32_t> & ids) {
@@ -15423,6 +15659,10 @@ static struct ggml_cgraph * llama_build_graph(
{
result = llm.build_rwkv6();
} break;
+ case LLM_ARCH_SOLAR:
+ {
+ result = llm.build_solar();
+ } break;
default:
GGML_ABORT("fatal error");
}
@@ -18503,6 +18743,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
case LLM_ARCH_ARCTIC:
case LLM_ARCH_DEEPSEEK2:
case LLM_ARCH_CHATGLM:
+ case LLM_ARCH_SOLAR:
return LLAMA_ROPE_TYPE_NORM;
// the pairs of head values are offset by n_rot/2
--
2.46.0

233
llm/payload.go
View File

@@ -1,233 +0,0 @@
package llm
import (
"compress/gzip"
"errors"
"fmt"
"io"
"io/fs"
"log/slog"
"os"
"path/filepath"
"runtime"
"slices"
"strings"
"golang.org/x/sync/errgroup"
"github.com/ollama/ollama/gpu"
)
var errPayloadMissing = errors.New("expected payloads not included in this build of ollama")
func Init() error {
payloadsDir, err := gpu.PayloadsDir()
if err != nil {
return err
}
if runtime.GOOS != "windows" {
slog.Info("extracting embedded files", "dir", payloadsDir)
binGlob := "build/*/*/*/bin/*"
// extract server libraries
err = extractFiles(payloadsDir, binGlob)
if err != nil {
return fmt.Errorf("extract binaries: %v", err)
}
}
var variants []string
for v := range getAvailableServers() {
variants = append(variants, v)
}
slog.Info(fmt.Sprintf("Dynamic LLM libraries %v", variants))
slog.Debug("Override detection logic by setting OLLAMA_LLM_LIBRARY")
return nil
}
// binary names may contain an optional variant separated by '_'
// For example, "ollama_rocm_v6" and "ollama_rocm_v5" or "ollama_cpu" and "ollama_cpu_avx2"
// Any library without a variant is the lowest common denominator
func getAvailableServers() map[string]string {
payloadsDir, err := gpu.PayloadsDir()
if err != nil {
slog.Error("payload lookup error", "error", err)
return nil
}
// glob payloadsDir for files that start with ollama_
pattern := filepath.Join(payloadsDir, "*", "ollama_*")
files, err := filepath.Glob(pattern)
if err != nil {
slog.Debug("could not glob", "pattern", pattern, "error", err)
return nil
}
servers := make(map[string]string)
for _, file := range files {
slog.Debug("availableServers : found", "file", file)
servers[filepath.Base(filepath.Dir(file))] = filepath.Dir(file)
}
return servers
}
// serversForGpu returns a list of compatible servers give the provided GPU
// info, ordered by performance. assumes Init() has been called
// TODO - switch to metadata based mapping
func serversForGpu(info gpu.GpuInfo) []string {
// glob workDir for files that start with ollama_
availableServers := getAvailableServers()
requested := info.Library
if info.Variant != gpu.CPUCapabilityNone.String() {
requested += "_" + info.Variant
}
servers := []string{}
// exact match first
for a := range availableServers {
if a == requested {
servers = []string{a}
if a == "metal" {
return servers
}
break
}
}
alt := []string{}
// Then for GPUs load alternates and sort the list for consistent load ordering
if info.Library != "cpu" {
for a := range availableServers {
if info.Library == strings.Split(a, "_")[0] && a != requested {
alt = append(alt, a)
}
}
slices.Sort(alt)
servers = append(servers, alt...)
}
if !(runtime.GOOS == "darwin" && runtime.GOARCH == "arm64") {
// Load up the best CPU variant if not primary requested
if info.Library != "cpu" {
variant := gpu.GetCPUCapability()
// If no variant, then we fall back to default
// If we have a variant, try that if we find an exact match
// Attempting to run the wrong CPU instructions will panic the
// process
if variant != gpu.CPUCapabilityNone {
for cmp := range availableServers {
if cmp == "cpu_"+variant.String() {
servers = append(servers, cmp)
break
}
}
} else {
servers = append(servers, "cpu")
}
}
if len(servers) == 0 {
servers = []string{"cpu"}
}
}
return servers
}
// Return the optimal server for this CPU architecture
func serverForCpu() string {
if runtime.GOOS == "darwin" && runtime.GOARCH == "arm64" {
return "metal"
}
variant := gpu.GetCPUCapability()
availableServers := getAvailableServers()
if variant != gpu.CPUCapabilityNone {
for cmp := range availableServers {
if cmp == "cpu_"+variant.String() {
return cmp
}
}
}
return "cpu"
}
// extract extracts the embedded files to the target directory
func extractFiles(targetDir string, glob string) error {
files, err := fs.Glob(libEmbed, glob)
if err != nil || len(files) == 0 {
return errPayloadMissing
}
if err := os.MkdirAll(targetDir, 0o755); err != nil {
return fmt.Errorf("extractFiles could not mkdir %s: %v", targetDir, err)
}
g := new(errgroup.Group)
// build/$OS/$GOARCH/$VARIANT/{bin,lib}/$FILE
for _, file := range files {
filename := file
variant := filepath.Base(filepath.Dir(filepath.Dir(filename)))
slog.Debug("extracting", "variant", variant, "file", filename)
g.Go(func() error {
srcf, err := libEmbed.Open(filename)
if err != nil {
return err
}
defer srcf.Close()
src := io.Reader(srcf)
if strings.HasSuffix(filename, ".gz") {
src, err = gzip.NewReader(src)
if err != nil {
return fmt.Errorf("decompress payload %s: %v", filename, err)
}
filename = strings.TrimSuffix(filename, ".gz")
}
variantDir := filepath.Join(targetDir, variant)
if err := os.MkdirAll(variantDir, 0o755); err != nil {
return fmt.Errorf("extractFiles could not mkdir %s: %v", variantDir, err)
}
base := filepath.Base(filename)
destFilename := filepath.Join(variantDir, base)
_, err = os.Stat(destFilename)
switch {
case errors.Is(err, os.ErrNotExist):
destFile, err := os.OpenFile(destFilename, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, 0o755)
if err != nil {
return fmt.Errorf("write payload %s: %v", filename, err)
}
defer destFile.Close()
if _, err := io.Copy(destFile, src); err != nil {
return fmt.Errorf("copy payload %s: %v", filename, err)
}
case err != nil:
return fmt.Errorf("stat payload %s: %v", filename, err)
}
return nil
})
}
err = g.Wait()
if err != nil {
// If we fail to extract, the payload dir is unusable, so cleanup whatever we extracted
gpu.Cleanup()
return err
}
return nil
}

31
llm/server.go
View File

@@ -24,9 +24,11 @@ import (
"golang.org/x/sync/semaphore"
"github.com/ollama/ollama/api"
"github.com/ollama/ollama/build"
"github.com/ollama/ollama/envconfig"
"github.com/ollama/ollama/format"
"github.com/ollama/ollama/gpu"
"github.com/ollama/ollama/runners"
)
type LlamaServer interface {
@@ -106,7 +108,7 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
gpus = gpu.GetCPUInfo()
}
if len(gpus) == 1 && gpus[0].Library == "cpu" {
cpuRunner = serverForCpu()
cpuRunner = runners.ServerForCpu()
estimate = EstimateGPULayers(gpus, ggml, projectors, opts)
} else {
estimate = EstimateGPULayers(gpus, ggml, projectors, opts)
@@ -118,7 +120,7 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
opts.NumGPU = 0
case gpus[0].Library != "metal" && estimate.Layers == 0:
// Don't bother loading into the GPU if no layers can fit
cpuRunner = serverForCpu()
cpuRunner = runners.ServerForCpu()
gpus = gpu.GetCPUInfo()
case opts.NumGPU < 0 && estimate.Layers > 0 && gpus[0].Library != "cpu":
opts.NumGPU = estimate.Layers
@@ -145,25 +147,20 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
return nil, errors.New("ollama supports only one lora adapter, but multiple were provided")
}
availableServers := getAvailableServers()
rDir, err := runners.Refresh(build.EmbedFS)
if err != nil {
return nil, err
}
availableServers := runners.GetAvailableServers(rDir)
if len(availableServers) == 0 {
if runtime.GOOS != "windows" {
slog.Warn("llama server binary disappeared, reinitializing payloads")
err = Init()
if err != nil {
slog.Warn("failed to reinitialize payloads", "error", err)
return nil, err
}
availableServers = getAvailableServers()
} else {
return nil, finalErr
}
return nil, finalErr
}
var servers []string
if cpuRunner != "" {
servers = []string{cpuRunner}
} else {
servers = serversForGpu(gpus[0]) // All GPUs in the list are matching Library and Variant
servers = runners.ServersForGpu(gpus[0]) // All GPUs in the list are matching Library and Variant
}
demandLib := envconfig.LLMLibrary()
if demandLib != "" {
@@ -274,7 +271,7 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
params = append(params, "--tensor-split", estimate.TensorSplit)
}
for i := range len(servers) {
for i := range servers {
dir := availableServers[servers[i]]
if dir == "" {
// Shouldn't happen
@@ -330,7 +327,7 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
_, err := os.Stat(server)
if errors.Is(err, os.ErrNotExist) {
slog.Warn("llama server disappeared, reinitializing payloads", "path", server, "error", err)
err = Init()
_, err = runners.Refresh(build.EmbedFS)
if err != nil {
slog.Warn("failed to reinitialize payloads", "error", err)
return nil, err