koboldcpp launch script

update llama-cpp-python and koboldcpp
llama.sh supports min_p sampling
llama.sh supports capybara

bin/llama-koboldcpp.sh

@@ -0,0 +1,39 @@
+#!/bin/bash
+
+# if conf does not exist, create it
+if [ ! -f "$HOME/.config/llama/llama-koboldcpp.conf" ]; then
+    mkdir -p "$HOME/.config/llama"
+    cat <<EOF > "$HOME/.config/llama/llama-koboldcpp.conf"
+LLAMA_MODEL_NAME=$HOME/.ai/models/llama/teknium/OpenHermes-2.5-Mistral-7B/openhermes-2.5-mistral-7b-f16.gguf
+LLAMA_CONTEXT_SIZE=8192
+
+KOBOLD_PORT=5000
+KOBOLD_LOG=$HOME/.local/var/llama-koboldcpp.log
+PYTHON_EXEC=$HOME/.virtualenvs/koboldcpp/bin/python
+EOF
+fi
+
+source "$HOME/.config/llama/llama-koboldcpp.conf"
+
+# if arg1 is "stop" then pkill the koboldcpp.py server
+if [[ $# -eq 1 ]] && [[ $1 == "stop" ]]; then
+    echo "Stopping koboldcpp server..."
+    pkill -f "koboldcpp.py"
+    echo "ok"
+    exit
+fi
+
+# start server and pipe stdout+stderr to log file
+cd ~/Work/koboldcpp || exit
+"$PYTHON_EXEC" koboldcpp.py \
+    --gpulayers 1 \
+    --model "$LLAMA_MODEL_NAME" \
+    --contextsize "$LLAMA_CONTEXT_SIZE" \
+    --threads 6 \
+    --skiplauncher \
+    --smartcontext \
+    --noblas \
+    --host "0.0.0.0" \
+    --port "$KOBOLD_PORT" \
+    > "$KOBOLD_LOG" 2>&1 &
+echo "Started koboldcpp server on port $KOBOLD_PORT"

bin/llama-update.sh

@@ -5,11 +5,12 @@ if [ ! -f "$HOME/.config/llama/llama-update.conf" ]; then
     mkdir -p "$HOME/.config/llama"
     cat <<EOF > "$HOME/.config/llama/llama-update.conf"
 LLAMA_CPP_GIT=$HOME/Work/llama.cpp
+KOBOLDCPP_GIT=$HOME/Work/koboldcpp
 BIN_PATH=$HOME/.local/bin
 EOF
 fi
 
-source $HOME/.config/llama/llama-update.conf
+source "$HOME"/.config/llama/llama-update.conf
 
 # if llama.cpp directory does not exist, clone it
 if [ ! -d "$LLAMA_CPP_GIT" ]; then
@@ -33,3 +34,23 @@ install -c -v bin/ggml-metal.metal "$BIN_PATH"
 install -c -v -m 755 bin/llava-cli "$BIN_PATH"/llava
 install -c -v -m 755 bin/finetune "$BIN_PATH/llama-finetune"
 install -c -v -m 755 bin/speculative "$BIN_PATH/llama-speculative"
+
+# configured for Metal
+CMAKE_ARGS="-DLLAMA_METAL=on" "$HOME/.virtualenvs/llama.cpp/bin/pip" install \
+    'llama-cpp-python[server]' \
+    --force-reinstall \
+    --upgrade \
+    --no-cache-dir
+
+# if koboldcpp directory does not exist, clone it
+if [ ! -d "$KOBOLDCPP_GIT" ]; then
+    git clone https://github.com/LostRuins/koboldcpp "$KOBOLDCPP_GIT"
+else
+    cd "$KOBOLDCPP_GIT" || exit
+    git pull
+fi
+
+# update koboldcpp
+cd "$KOBOLDCPP_GIT" || exit
+make clean
+make LLAMA_METAL=1 # LLAMA_CLBLAST=1

bin/llama.sh

@@ -8,8 +8,9 @@ function llama_init_environment {
 LLAMA_TEMPERATURE=0.1
 LLAMA_CONTEXT_SIZE=4096
 LLAMA_REPETITION_PENALTY=1.15
-LLAMA_TOP_P=0.9
-LLAMA_TOP_K=20
+LLAMA_TOP_P=1.0
+LLAMA_MIN_P=0.3
+LLAMA_TOP_K=0
 LLAMA_TEMPLATE=chatml
 LLAMA_MODEL_NAME=teknium/OpenHermes-2.5-Mistral-7B/openhermes-2.5-mistral-7b-f16.gguf
 LLAMA_MODELS_PATH=$HOME/.ai/models/llama
@@ -52,10 +53,12 @@ function llama_interactive {
         --in-suffix "$(get_model_suffix $LLAMA_TEMPLATE)" \
         --reverse-prompt "$(get_model_prefix $LLAMA_TEMPLATE)" \
         --reverse-prompt "<|im_end|>" \
+        --reverse-prompt "</s>" \
         --threads "6" \
         --temp "$LLAMA_TEMPERATURE" \
         --top-p "$LLAMA_TOP_P" \
         --top-k "$LLAMA_TOP_K" \
+        --min-p "$LLAMA_MIN_P" \
         --repeat-penalty "$LLAMA_REPETITION_PENALTY" \
         --ctx-size "$LLAMA_CONTEXT_SIZE" \
         --batch-size 1024 \
@@ -79,7 +82,7 @@ function get_model_prefix {
         *alpaca*|*upstage*|*airoboros*|*hermes*)
             printf "### Instruction: "
             ;;
-        *vicuna*|*wizardlm*|*samantha*|*scarlett*)
+        *vicuna*|*wizardlm*|*samantha*|*scarlett*|*capybara*)
             printf "USER: "
             ;;
         *based*|*yi*)
@@ -105,7 +108,7 @@ function get_model_suffix {
         *alpaca*|*upstage*|*airoboros*|*hermes*)
             printf "### Response:"
             ;;
-        *vicuna*|*wizardlm*|*samantha*|*scarlett*)
+        *vicuna*|*wizardlm*|*samantha*|*scarlett*|*capybara*)
             printf "ASSISTANT:"
             ;;
         *based*|*yi*)
@@ -149,6 +152,9 @@ function get_model_prompt {
         *orca*)
             echo "$LLAMA_PROMPT_PATH/orca.txt"
             ;;
+        *capybara*)
+            echo "$LLAMA_PROMPT_PATH/idm-gpt-lite.txt"
+            ;;
         *)
             echo "$LLAMA_PROMPT_PATH/idm-gpt-lite.txt"
             ;;

install.sh

@@ -13,6 +13,7 @@ install -C -v ./bin/llama-finetune.sh ~/.local/bin
 install -C -v ./bin/llama.sh ~/.local/bin
 install -C -v ./bin/llama-menu.sh ~/.local/bin
 install -C -v ./bin/llama-server.sh ~/.local/bin
+install -C -v ./bin/llama-koboldcpp.sh ~/.local/bin
 install -C -v ./bin/llava-describe.sh ~/.local/bin
 install -C -v ./bin/llama-prompt.sh ~/.local/bin