Enhance LLM configuration and installation scripts for improved flexibility

- Added new environment variables in `env.example` and `gpu-rent.vars.example` for fine-tuning LLM settings, including CUDA build options and model version pinning.
- Updated `llm.md` documentation to include detailed descriptions of new configuration options and usage cases for LLM setups.
- Enhanced the `provision.py` script to forward new environment variables during remote installations, improving the installation process for LLM components.
- Modified the `install_llamacpp.sh` script to support conditional CUDA builds and asset URL overrides, ensuring better compatibility with various environments.
- Improved logging in the installation scripts to provide clearer feedback during the setup process.

src/gpu_rent/remote/install_llamacpp.sh

@@ -1,6 +1,9 @@
 #!/usr/bin/env bash
-# Install llama-server (CUDA) for OpenAI-compatible API on loopback :8080.
-# Official GitHub releases ship Windows CUDA only — on Linux we build from source.
+# Install llama-server for OpenAI-compatible API on loopback :8080.
+#
+# Default: official Linux release asset (Ubuntu Vulkan — GPU without compile).
+# CUDA source build only as last resort (or LLAMACPP_BUILD_CUDA=1).
+# Pin: LLAMACPP_TAG=b10545  LLAMACPP_ASSET_URL=...  LLAMACPP_SHA256=...
 set -euo pipefail
 
 SWARM_USER="${SWARM_USER:-ubuntu}"
@@ -28,8 +31,21 @@ SERVER_BIN="${BIN_DIR}/llama-server"
 LLAMACPP_TAG="${LLAMACPP_TAG:-}"
 LLAMACPP_ASSET_URL="${LLAMACPP_ASSET_URL:-}"
 LLAMACPP_SHA256="${LLAMACPP_SHA256:-}"
+# 1 = force CUDA compile; 0/empty = prebuilt first, build only if prebuilt fails
+LLAMACPP_BUILD_CUDA="${LLAMACPP_BUILD_CUDA:-}"
+LLAMACPP_FORCE_REINSTALL="${LLAMACPP_FORCE_REINSTALL:-}"
+LLAMACPP_NGL="${LLAMACPP_NGL:-}"
+LLAMACPP_CTX="${LLAMACPP_CTX:-}"
+LLAMACPP_HOST="${LLAMACPP_HOST:-127.0.0.1}"
+LLAMACPP_PORT="${LLAMACPP_PORT:-8080}"
+LLAMACPP_EXTRA_ARGS="${LLAMACPP_EXTRA_ARGS:-}"
 
-# Pick a Linux release asset (never Windows/macOS). Prefer ubuntu CUDA → vulkan → cpu.
+if [[ "${LLAMACPP_FORCE_REINSTALL}" == "1" ]]; then
+  log "LLAMACPP_FORCE_REINSTALL=1 — удаляю старый бинарь"
+  rm -f "$SERVER_BIN" "$STAMP"
+fi
+
+# Prefer ubuntu CUDA (rare) → vulkan → cpu. Never Windows/macOS/cudart-only.
 pick_linux_asset_url() {
   python3 -c '
 import json,sys
@@ -107,7 +123,7 @@ ensure_build_deps() {
     export PATH="/usr/local/cuda/bin:${PATH}"
     return 0
   fi
-  log "ставлю nvidia-cuda-toolkit (нужен nvcc для сборки)…"
+  log "ставлю nvidia-cuda-toolkit (нужен nvcc)…"
   apt-get install -y -qq nvidia-cuda-toolkit >/dev/null
   if command -v nvcc >/dev/null 2>&1; then
     return 0
@@ -119,14 +135,24 @@ ensure_build_deps() {
   return 1
 }
 
+ensure_vulkan_runtime() {
+  if ldconfig -p 2>/dev/null | grep -q 'libvulkan\.so'; then
+    return 0
+  fi
+  export DEBIAN_FRONTEND=noninteractive
+  log "ставлю libvulkan1 (для Ubuntu Vulkan prebuilt)…"
+  apt-get install -y -qq libvulkan1 mesa-vulkan-drivers >/dev/null 2>&1 || \
+    apt-get install -y -qq libvulkan1 >/dev/null 2>&1 || true
+}
+
 install_from_archive_url() {
   local url="$1"
   local tmp
   tmp="$(mktemp -d)"
   (
     cd "$tmp"
-    log "asset $url"
-    curl -fL "$url" -o pkg.bin
+    log "скачиваю prebuilt: $url"
+    curl -fL --progress-bar "$url" -o pkg.bin
     if [[ -n "$LLAMACPP_SHA256" ]]; then
       echo "${LLAMACPP_SHA256}  pkg.bin" | sha256sum -c -
     else
@@ -149,58 +175,19 @@ install_from_archive_url() {
       exit 1
     fi
     install -m 755 "$found" "$SERVER_BIN"
-    chown "${SWARM_USER}:${SWARM_USER}" "$SERVER_BIN"
+    # Shared libs next to binary (release tarballs ship .so alongside).
+    find out -type f \( -name '*.so' -o -name '*.so.*' \) -print0 2>/dev/null \
+      | while IFS= read -r -d '' so; do
+          install -m 755 "$so" "${BIN_DIR}/$(basename "$so")"
+        done
+    chown -R "${SWARM_USER}:${SWARM_USER}" "$BIN_DIR"
   )
   local rc=$?
   rm -rf "$tmp"
   return "$rc"
 }
 
-build_cuda_from_source() {
-  local tag="$1"
-  local arch
-  arch="$(cuda_architectures)"
-  log "собираю llama-server CUDA из исходников (tag=${tag}, arch=${arch})…"
-  if ! ensure_build_deps; then
-    log "нет nvcc — CUDA-сборку пропускаем"
-    return 1
-  fi
-  log "nvcc $(nvcc --version 2>/dev/null | tail -n1 || echo '?')"
-
-  mkdir -p "$SRC_DIR"
-  if [[ -d "${SRC_DIR}/.git" ]]; then
-    git -C "$SRC_DIR" fetch --depth 1 origin tag "$tag" 2>/dev/null || true
-    if ! git -C "$SRC_DIR" checkout -f "$tag" 2>/dev/null; then
-      rm -rf "$SRC_DIR"
-      git clone --depth 1 --branch "$tag" "$REPO" "$SRC_DIR"
-    fi
-  else
-    rm -rf "$SRC_DIR"
-    git clone --depth 1 --branch "$tag" "$REPO" "$SRC_DIR"
-  fi
-
-  cmake -S "$SRC_DIR" -B "${SRC_DIR}/build" \
-    -DCMAKE_BUILD_TYPE=Release \
-    -DGGML_CUDA=ON \
-    -DCMAKE_CUDA_ARCHITECTURES="${arch}" \
-    -DLLAMA_CURL=ON \
-    -DLLAMA_BUILD_SERVER=ON
-  cmake --build "${SRC_DIR}/build" -j"$(nproc)" --target llama-server
-
-  local built="${SRC_DIR}/build/bin/llama-server"
-  if [[ ! -x "$built" ]]; then
-    log "сборка не дала ${built}"
-    return 1
-  fi
-  install -m 755 "$built" "$SERVER_BIN"
-  chown "${SWARM_USER}:${SWARM_USER}" "$SERVER_BIN"
-  echo "cuda:${tag}:${arch}" >"$STAMP"
-  chown "${SWARM_USER}:${SWARM_USER}" "$STAMP"
-  log "CUDA binary → ${SERVER_BIN}"
-  return 0
-}
-
-install_linux_release_fallback() {
+install_linux_release() {
   local tag="$1"
   local api url
   api="${API_BASE}/releases/tags/${tag}"
@@ -210,36 +197,134 @@ install_linux_release_fallback() {
     return 1
   fi
   if [[ "$url" == *vulkan* ]]; then
-    log "WARN: официального Linux CUDA нет — беру Ubuntu Vulkan prebuilt"
-  elif [[ "$url" != *cuda* ]]; then
-    log "WARN: беру Linux prebuilt без CUDA (CPU) — лучше собрать с nvcc"
+    log "беру Ubuntu Vulkan prebuilt (GPU без compile; CUDA-сборка — LLAMACPP_BUILD_CUDA=1)"
+    ensure_vulkan_runtime
+  elif [[ "$url" == *cuda* ]]; then
+    log "беру Linux CUDA prebuilt"
+  else
+    log "WARN: Linux prebuilt без GPU backend (CPU) — ${url##*/}"
   fi
   install_from_archive_url "$url"
   echo "asset:${tag}" >"$STAMP"
   chown "${SWARM_USER}:${SWARM_USER}" "$STAMP"
 }
 
+# Quiet CUDA build: no cmake spam; heartbeat every 30s with last %.
+build_cuda_from_source() {
+  local tag="$1"
+  local arch build_log pid pct line
+  arch="$(cuda_architectures)"
+  build_log="${LLAMA_ROOT}/build-cuda.log"
+  log "крайний случай: сборка CUDA из исходников (tag=${tag}, arch=${arch}, 5–15 мин)…"
+  log "полный лог: ${build_log}"
+  if ! ensure_build_deps; then
+    log "нет nvcc — CUDA-сборку пропускаем"
+    return 1
+  fi
+  log "nvcc $(nvcc --version 2>/dev/null | tail -n1 || echo '?')"
+
+  mkdir -p "$SRC_DIR"
+  export GIT_TERMINAL_PROMPT=0
+  if [[ -d "${SRC_DIR}/.git" ]]; then
+    git -C "$SRC_DIR" -c advice.detachedHead=false fetch --depth 1 origin tag "$tag" 2>>"$build_log" || true
+    if ! git -C "$SRC_DIR" -c advice.detachedHead=false checkout -f "$tag" >>"$build_log" 2>&1; then
+      rm -rf "$SRC_DIR"
+      git -c advice.detachedHead=false clone --depth 1 --branch "$tag" "$REPO" "$SRC_DIR" >>"$build_log" 2>&1
+    fi
+  else
+    rm -rf "$SRC_DIR"
+    git -c advice.detachedHead=false clone --depth 1 --branch "$tag" "$REPO" "$SRC_DIR" >>"$build_log" 2>&1
+  fi
+
+  cmake -S "$SRC_DIR" -B "${SRC_DIR}/build" \
+    -DCMAKE_BUILD_TYPE=Release \
+    -DGGML_CUDA=ON \
+    -DCMAKE_CUDA_ARCHITECTURES="${arch}" \
+    -DLLAMA_BUILD_SERVER=ON \
+    -DLLAMA_BUILD_UI=OFF \
+    -DLLAMA_USE_PREBUILT_UI=OFF \
+    -DGGML_CCACHE=OFF \
+    >>"$build_log" 2>&1
+
+  # Background build + heartbeat (keeps SSH stream alive without 200 cmake lines).
+  cmake --build "${SRC_DIR}/build" -j"$(nproc)" --target llama-server \
+    >>"$build_log" 2>&1 &
+  pid=$!
+  while kill -0 "$pid" 2>/dev/null; do
+    pct="$(grep -oE '\[[[:space:]]*[0-9]+%\]' "$build_log" 2>/dev/null | tail -n1 || true)"
+    line="$(grep -E 'Building CUDA|Built target|Linking' "$build_log" 2>/dev/null | tail -n1 || true)"
+    if [[ -n "$pct" ]]; then
+      log "сборка CUDA ещё идёт… ${pct}${line:+ · ${line}}"
+    else
+      log "сборка CUDA ещё идёт… (cmake/nvcc, см. build.log)"
+    fi
+    sleep 30
+  done
+  if ! wait "$pid"; then
+    log "сборка упала — хвост ${build_log}:"
+    tail -n 40 "$build_log" >&2 || true
+    return 1
+  fi
+
+  local built="${SRC_DIR}/build/bin/llama-server"
+  if [[ ! -x "$built" ]]; then
+    log "сборка не дала ${built}"
+    return 1
+  fi
+  install -m 755 "$built" "$SERVER_BIN"
+  # CUDA build may need libs from build/bin
+  find "${SRC_DIR}/build/bin" -maxdepth 1 -type f \( -name '*.so' -o -name '*.so.*' \) -print0 2>/dev/null \
+    | while IFS= read -r -d '' so; do
+        install -m 755 "$so" "${BIN_DIR}/$(basename "$so")"
+      done
+  chown -R "${SWARM_USER}:${SWARM_USER}" "$BIN_DIR"
+  echo "cuda:${tag}:${arch}" >"$STAMP"
+  chown "${SWARM_USER}:${SWARM_USER}" "$STAMP"
+  log "CUDA binary → ${SERVER_BIN}"
+  return 0
+}
+
+normalize_tag() {
+  printf '%s' "$1" | tr -d '\r' | head -n1 | awk 'NF{print; exit}'
+}
+
 if [[ -x "$SERVER_BIN" ]]; then
   log "llama-server уже есть: ${SERVER_BIN}"
 else
   if [[ -n "$LLAMACPP_ASSET_URL" ]]; then
-    log "скачиваю llama-server по LLAMACPP_ASSET_URL…"
+    log "скачиваю по LLAMACPP_ASSET_URL…"
     install_from_archive_url "$LLAMACPP_ASSET_URL"
     echo "asset-url" >"$STAMP"
     chown "${SWARM_USER}:${SWARM_USER}" "$STAMP"
   else
     if [[ -z "$LLAMACPP_TAG" ]]; then
-      log "WARN: LLAMACPP_TAG/ASSET_URL не заданы — берём latest (нет pin). См. docs/llm.md"
+      log "WARN: LLAMACPP_TAG не задан — latest (см. docs/llm.md)"
     fi
-    tag="$(resolve_release_tag)"
-    tag="$(printf '%s' "$tag" | tr -d '\r' | head -n1 | awk 'NF{print; exit}')"
+    tag="$(normalize_tag "$(resolve_release_tag)")"
     if [[ -z "$tag" || "$tag" == *" "* || "$tag" == *"["* ]]; then
       log "не удалось определить release tag (got: ${tag:-empty})"
       exit 1
     fi
-    if ! build_cuda_from_source "$tag"; then
-      log "CUDA-сборка не удалась — fallback на Linux release asset"
-      install_linux_release_fallback "$tag"
+
+    installed=0
+    if [[ "${LLAMACPP_BUILD_CUDA}" == "1" ]]; then
+      log "LLAMACPP_BUILD_CUDA=1 — сразу CUDA-сборка"
+      if build_cuda_from_source "$tag"; then
+        installed=1
+      fi
+    else
+      if install_linux_release "$tag"; then
+        installed=1
+      else
+        log "prebuilt не вышел — крайний случай: CUDA из исходников"
+        if build_cuda_from_source "$tag"; then
+          installed=1
+        fi
+      fi
+    fi
+    if [[ "$installed" != "1" ]]; then
+      log "не удалось поставить llama-server"
+      exit 1
     fi
   fi
 fi
@@ -292,7 +377,14 @@ else:
 PY
 )" || true
 fi
-log "llama.cpp -ngl ${NGL} -c ${CTX}"
+# Explicit overrides from gpu-rent.vars / .env (forwarded by provision).
+if [[ -n "$LLAMACPP_NGL" ]]; then
+  NGL="$LLAMACPP_NGL"
+fi
+if [[ -n "$LLAMACPP_CTX" ]]; then
+  CTX="$LLAMACPP_CTX"
+fi
+log "llama.cpp -ngl ${NGL} -c ${CTX} host=${LLAMACPP_HOST} port=${LLAMACPP_PORT}${LLAMACPP_EXTRA_ARGS:+ extra=${LLAMACPP_EXTRA_ARGS}}"
 
 cat >/etc/systemd/system/${UNIT}.service <<EOF
 [Unit]
@@ -305,7 +397,8 @@ Type=simple
 User=${SWARM_USER}
 Group=${SWARM_USER}
 WorkingDirectory=${LLAMA_ROOT}
-ExecStart=${SERVER_BIN} ${MODEL_ARG} ${MMPROJ_ARG} --host 127.0.0.1 --port 8080 -ngl ${NGL} -c ${CTX}
+Environment=LD_LIBRARY_PATH=${BIN_DIR}
+ExecStart=${SERVER_BIN} ${MODEL_ARG} ${MMPROJ_ARG} --host ${LLAMACPP_HOST} --port ${LLAMACPP_PORT} -ngl ${NGL} -c ${CTX} ${LLAMACPP_EXTRA_ARGS}
 Restart=on-failure
 RestartSec=8