feat(llama.cpp): upgrade and use libmtmd (#5379)

* WIP

* wip

* wip

* Make it compile

* Update json.hpp

* this shouldn't be private for now

* Add logs

* Reset auto detected template

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* Re-enable grammars

* This seems to be broken - 360a9c98e1 (diff-a18a8e64e12a01167d8e98fc)[…]cccf0d4eed09d76d879L2998-L3207

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* Placeholder

* Simplify image loading

* use completion type

* disable streaming

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* correctly return timings

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* Remove some debug logging

* Adapt tests

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* Keep header

* embedding: do not use oai type

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* Sync from server.cpp

* Use utils and json directly from llama.cpp

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* Sync with upstream

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* fix: copy json.hpp from the correct location

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* fix: add httplib

* sync llama.cpp

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* Embeddiongs: set OAICOMPAT_TYPE_EMBEDDING

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* feat: sync with server.cpp by including it

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* make it darwin-compatible

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

---------

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

backend/cpp/llama/prepare.sh

@@ -7,22 +7,46 @@ for patch in $(ls patches); do
     patch -d llama.cpp/ -p1 < patches/$patch
 done 
 
+set -e
+
 cp -r CMakeLists.txt llama.cpp/tools/grpc-server/
 cp -r grpc-server.cpp llama.cpp/tools/grpc-server/
-cp -rfv json.hpp llama.cpp/tools/grpc-server/
-cp -rfv utils.hpp llama.cpp/tools/grpc-server/
-    
+cp -rfv llama.cpp/common/json.hpp llama.cpp/tools/grpc-server/
+cp -rfv llama.cpp/tools/server/utils.hpp llama.cpp/tools/grpc-server/
+cp -rfv llama.cpp/tools/server/httplib.h llama.cpp/tools/grpc-server/
+
+set +e
 if grep -q "grpc-server" llama.cpp/tools/CMakeLists.txt; then
     echo "grpc-server already added"
 else
     echo "add_subdirectory(grpc-server)" >> llama.cpp/tools/CMakeLists.txt
 fi
+set -e
 
-## XXX: In some versions of CMake clip wasn't being built before llama.
-## This is an hack for now, but it should be fixed in the future.
-cp -rfv llama.cpp/tools/mtmd/clip.h llama.cpp/tools/grpc-server/clip.h
-cp -rfv llama.cpp/tools/mtmd/clip-impl.h llama.cpp/tools/grpc-server/clip-impl.h
-cp -rfv llama.cpp/tools/mtmd/llava.cpp llama.cpp/tools/grpc-server/llava.cpp
-echo '#include "llama.h"' > llama.cpp/tools/grpc-server/llava.h
-cat llama.cpp/tools/mtmd/llava.h >> llama.cpp/tools/grpc-server/llava.h
-cp -rfv llama.cpp/tools/mtmd/clip.cpp llama.cpp/tools/grpc-server/clip.cpp
+# Now to keep maximum compatibility with the original server.cpp, we need to remove the index.html.gz.hpp and loading.html.hpp includes
+# and remove the main function
+# TODO: upstream this to the original server.cpp by extracting the upstream main function to a separate file
+awk '
+/int[ \t]+main[ \t]*\(/ {          # If the line starts the main function
+    in_main=1;                     # Set a flag
+    open_braces=0;                 # Track number of open braces
+}
+in_main {
+    open_braces += gsub(/\{/, "{"); # Count opening braces
+    open_braces -= gsub(/\}/, "}"); # Count closing braces
+    if (open_braces == 0) {         # If all braces are closed
+        in_main=0;                  # End skipping
+    }
+    next;                           # Skip lines inside main
+}
+!in_main                           # Print lines not inside main
+' "llama.cpp/tools/server/server.cpp" > llama.cpp/tools/grpc-server/server.cpp
+
+# remove index.html.gz.hpp and loading.html.hpp includes
+if [[ "$OSTYPE" == "darwin"* ]]; then
+    # macOS
+    sed -i '' '/#include "index\.html\.gz\.hpp"/d; /#include "loading\.html\.hpp"/d' llama.cpp/tools/grpc-server/server.cpp
+else
+    # Linux and others
+    sed -i '/#include "index\.html\.gz\.hpp"/d; /#include "loading\.html\.hpp"/d' llama.cpp/tools/grpc-server/server.cpp
+fi