feat(llama.cpp): add turboquant support

This PR adds patchset from the great work of @TheTom in
https://github.com/TheTom/llama-cpp-turboquant and creates a pipeline
that updates the patches against upstream automatically.

It also creates necessary scaffolding for doing this with other patches
sources.

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

Author: mudler

.github/workflows/bump_deps.yaml

@@ -63,6 +63,3 @@ jobs:
           branch: "update/${{ matrix.variable }}"
           body: ${{ steps.bump.outputs.message }}
           signoff: true
-
-
-

backend/cpp/llama-cpp/patches/sources.yaml

@@ -0,0 +1,12 @@
+# Patch sources for the llama-cpp backend.
+# Each source declares a fork whose commits are extracted as patches
+# and applied on top of upstream llama.cpp during the build.
+# See scripts/patch_utils/upgrade_patches.sh for the rebase/regenerate tool.
+sources:
+  - name: turboquant
+    repo: https://github.com/TheTom/llama-cpp-turboquant.git
+    branch: feature/turboquant-kv-cache
+    fork_base: ded446b34c0cd803a0122446b848619adbb458cf
+    upstream_repo: https://github.com/ggml-org/llama.cpp.git
+    version_file: Makefile
+    version_var: LLAMA_VERSION

backend/cpp/llama-cpp/prepare.sh

@@ -1,16 +1,66 @@
 #!/bin/bash
+set -e
 
-## Patches
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
 
-## Apply patches from the `patches` directory
-if [ -d "patches" ]; then
-    for patch in $(ls patches); do
-        echo "Applying patch $patch"
-        patch -d llama.cpp/ -p1 < patches/$patch
-    done 
+## Generate patches from fork sources if not already present.
+## If patches/<name>/ already contains .patch files (vendored), skip fetching.
+## Otherwise, clone the fork and generate them on the fly.
+if [ -f "patches/sources.yaml" ] && command -v yq &>/dev/null; then
+    SOURCE_COUNT=$(yq '.sources | length' patches/sources.yaml)
+
+    for i in $(seq 0 $((SOURCE_COUNT - 1))); do
+        NAME=$(yq ".sources[$i].name" patches/sources.yaml)
+        REPO=$(yq ".sources[$i].repo" patches/sources.yaml)
+        BRANCH=$(yq ".sources[$i].branch" patches/sources.yaml)
+        FORK_BASE=$(yq ".sources[$i].fork_base" patches/sources.yaml)
+
+        SOURCE_DIR="patches/$NAME"
+        EXISTING=$(ls "$SOURCE_DIR"/*.patch 2>/dev/null | wc -l)
+
+        if [ "$EXISTING" -gt 0 ]; then
+            echo "Source '$NAME': $EXISTING patches already present — skipping fetch."
+        else
+            echo "Source '$NAME': no patches found — fetching from $REPO ($BRANCH)"
+
+            TMPDIR=$(mktemp -d)
+
+            if git clone --single-branch -b "$BRANCH" --depth=500 "$REPO" "$TMPDIR/fork" 2>&1; then
+                cd "$TMPDIR/fork"
+                PATCH_COUNT=$(git rev-list --count "$FORK_BASE"..HEAD 2>/dev/null || echo "0")
+                echo "  $PATCH_COUNT commits to extract"
+
+                if [ "$PATCH_COUNT" -gt 0 ]; then
+                    mkdir -p "$SCRIPT_DIR/$SOURCE_DIR"
+                    git format-patch "$FORK_BASE"..HEAD -o "$SCRIPT_DIR/$SOURCE_DIR/" >/dev/null 2>&1
+                    echo "  Generated $PATCH_COUNT patches in $SOURCE_DIR/"
+                fi
+                cd "$SCRIPT_DIR"
+            else
+                echo "WARNING: Failed to clone $REPO — skipping source '$NAME'"
+            fi
+
+            rm -rf "$TMPDIR"
+        fi
+    done
+elif [ -f "patches/sources.yaml" ]; then
+    echo "WARNING: yq not found — skipping source-based patch generation. Install yq to enable."
 fi
 
-set -e
+## Apply patches from source subdirectories (alphabetical order)
+if [ -d "patches" ]; then
+    for source_dir in $(find patches -mindepth 1 -maxdepth 1 -type d | sort); do
+        for p in $(ls "$source_dir"/*.patch 2>/dev/null | sort); do
+            echo "Applying: $p"
+            patch -d llama.cpp/ -p1 < "$p" || { echo "FAILED: $p"; exit 1; }
+        done
+    done
+    # Apply any top-level .patch files (manual patches)
+    for p in $(ls patches/*.patch 2>/dev/null | sort); do
+        echo "Applying: $p"
+        patch -d llama.cpp/ -p1 < "$p" || { echo "FAILED: $p"; exit 1; }
+    done
+fi
 
 for file in $(ls llama.cpp/tools/server/); do
     cp -rfv llama.cpp/tools/server/$file llama.cpp/tools/grpc-server/
@@ -28,4 +78,3 @@ else
     echo "add_subdirectory(grpc-server)" >> llama.cpp/tools/CMakeLists.txt
 fi
 set -e
-