fix: make CUDA runtime setup self-healing

CHANGELOG.md

@@ -1,3 +1,21 @@
+## 0.12.9
+
+- Preserve the NextPlaid log offset in resident-service health views so stale
+  CUDA diagnostics cannot trigger CPU fallback after service startup.
+
+## 0.12.8
+
+- Fix CUDA provider diagnostics so stale errors from a previous NextPlaid
+  process cannot force a healthy CUDA runtime onto CPU fallback.
+
+## 0.12.7
+
+- Fix explicit CUDA installer selection and require the CUDA math libraries
+  used by the ONNX Runtime provider.
+- Add version-aware CUDA library path detection and prevent stale provider log
+  diagnostics from forcing a false CPU fallback.
+
+
 ## 0.12.6
 
 - feat: harden dependencies and release workflows

README.md

@@ -47,10 +47,11 @@ updates your `PATH`.
 
 On Linux x86-64, the installer performs a GPU/driver/CUDA preflight. When a
 compatible visible NVIDIA GPU is present, it offers to install whichever CUDA
-12 runtime (`cuda-cudart-12-9`) and cuDNN 9 (`cudnn9-cuda-12`) libraries are
-missing; otherwise it installs the CPU package and prints the corrective
-action. Use `LLM_CONTEXT_ACCELERATOR_PACKAGE=auto`, `cpu`, or `cuda` to choose
-explicitly. Inspect or repair the host later with:
+12 runtime (`cuda-cudart-12-9`), CUDA math libraries (`cuda-libraries-12-9`,
+including cuBLAS/cuRAND), and cuDNN 9 (`cudnn9-cuda-12`) libraries are missing;
+otherwise it installs the CPU package and prints the corrective action. Use
+`LLM_CONTEXT_ACCELERATOR_PACKAGE=auto`, `cpu`, or `cuda` to choose explicitly.
+Inspect or repair the host later with:
 
 ```bash
 llm-context setup
@@ -74,7 +75,7 @@ curl -fsSL https://raw.githubusercontent.com/devame/llm-context-tools/main/insta
   | LLM_CONTEXT_SKIP_SEMANTIC=1 sh
 ```
 
-Set `LLM_CONTEXT_VERSION=0.12.6` to pin the current release. See the
+Set `LLM_CONTEXT_VERSION=0.12.9` to pin the current release. See the
 [installation and troubleshooting guide](docs/user-guide.md#installation-and-troubleshooting)
 for custom locations, CUDA, and verified model packages.
 
@@ -261,7 +262,7 @@ around the same JAR:
 
 ```bash
 npm pack
-npm install --global ./llm-context-0.12.6.tgz
+npm install --global ./llm-context-0.12.9.tgz
 llm-context doctor
 ```
 

build.clj

@@ -4,7 +4,7 @@
            [java.security MessageDigest]))
 
 (def lib 'devame/llm-context)
-(def version "0.12.6")
+(def version "0.12.9")
 (def class-dir "target/classes")
 (def basis (delay (b/create-basis {:project "deps.edn"})))
 (def uber-file (format "target/llm-context-%s-standalone.jar" version))

docs/user-guide.md

@@ -323,13 +323,15 @@ llm-context doctor
 ```
 
 `setup` reports the visible GPU, NVIDIA driver version, minimum supported
-driver (`525.60.13` for the CUDA 12 package), `libcuda.so.1`, CUDA 12, and
-cuDNN 9. These are static host checks; the first semantic service startup is
-still the authoritative runtime probe. If it finds missing CUDA 12 or cuDNN
-packages on Debian/Ubuntu, setup offers:
+driver (`525.60.13` for the CUDA 12 package), `libcuda.so.1`, CUDA 12, the CUDA
+math libraries used by ONNX Runtime (cuBLAS/cuRAND), and cuDNN 9. These are
+static host checks; the first semantic service startup is still the
+authoritative runtime probe. If it finds missing CUDA packages on Debian/Ubuntu,
+setup offers:
 
 ```bash
-sudo apt-get update && sudo apt-get -y install cuda-cudart-12-9 cudnn9-cuda-12
+sudo apt-get update && sudo apt-get -y install \
+  cuda-cudart-12-9 cuda-libraries-12-9 cudnn9-cuda-12
 ```
 
 Use `llm-context setup --install-cudnn --yes` after authenticating with sudo and
@@ -346,9 +348,10 @@ do not install a Linux NVIDIA driver inside WSL.
 On Linux x86-64, `install.sh` defaults to
 `LLM_CONTEXT_ACCELERATOR_PACKAGE=auto`. When a visible GPU and compatible
 driver are present, the installer asks whether it should install whichever of
-`cuda-cudart-12-9` and `cudnn9-cuda-12` is missing, using `apt-get` and `sudo`
-when needed. A declined prompt, a non-interactive terminal, or an unavailable
-package manager falls back to the CPU package and prints the corrective action.
+`cuda-cudart-12-9`, `cuda-libraries-12-9`, and `cudnn9-cuda-12` is missing, using
+`apt-get` and `sudo` when needed. A declined prompt, a non-interactive terminal,
+or an unavailable package manager falls back to the CPU package and prints the
+corrective action.
 Set `LLM_CONTEXT_CUDA_INSTALL=yes` or `no` for automation;
 `LLM_CONTEXT_CUDNN_INSTALL` remains accepted as a compatibility alias. Use
 `LLM_CONTEXT_ACCELERATOR_PACKAGE=cpu` to force CPU or

install.sh

@@ -90,6 +90,7 @@ verify_hash() {
 
 CUDA_MINIMUM_DRIVER="525.60.13"
 CUDA_RUNTIME_PACKAGE="cuda-cudart-12-9"
+CUDA_LIBRARIES_PACKAGE="cuda-libraries-12-9"
 CUDA_CUDNN_PACKAGE="cudnn9-cuda-12"
 
 version_at_least() {
@@ -129,6 +130,9 @@ cuda_library_present() {
   for candidate in \
     "/usr/lib/wsl/lib/$library" \
     "/usr/local/cuda/lib64/$library" \
+    "/usr/local/cuda/targets/x86_64-linux/lib/$library" \
+    "/usr/local/cuda-12.9/lib64/$library" \
+    "/usr/local/cuda-12.9/targets/x86_64-linux/lib/$library" \
     "/usr/lib/x86_64-linux-gnu/$library" \
     "/usr/lib64/$library"; do
     if [ -f "$candidate" ]; then
@@ -151,6 +155,7 @@ cuda_host_preflight() {
   CUDA_DEVICE_VISIBLE=0
   CUDA_LIBCUDA_PRESENT=0
   CUDA_RUNTIME_PRESENT=0
+  CUDA_LIBRARIES_PRESENT=0
   CUDA_CUDNN_PRESENT=0
 
   if [ -n "$CUDA_NVIDIA_SMI" ]; then
@@ -184,6 +189,10 @@ cuda_host_preflight() {
   if cuda_library_present libcudart.so.12; then
     CUDA_RUNTIME_PRESENT=1
   fi
+  if cuda_library_present libcublasLt.so.12 &&
+     cuda_library_present libcurand.so.10; then
+    CUDA_LIBRARIES_PRESENT=1
+  fi
   if cuda_library_present libcudnn.so.9; then
     CUDA_CUDNN_PRESENT=1
   fi
@@ -193,6 +202,7 @@ cuda_host_preflight() {
      [ "$CUDA_DRIVER_COMPATIBLE" -eq 1 ] &&
      [ "$CUDA_LIBCUDA_PRESENT" -eq 1 ] &&
      [ "$CUDA_RUNTIME_PRESENT" -eq 1 ] &&
+     [ "$CUDA_LIBRARIES_PRESENT" -eq 1 ] &&
      [ "$CUDA_CUDNN_PRESENT" -eq 1 ]; then
     CUDA_HOST_READY=1
   else
@@ -201,13 +211,14 @@ cuda_host_preflight() {
 }
 
 print_cuda_host_preflight() {
-  printf 'GPU preflight: GPU=%s; device=%s; driver=%s (minimum %s); libcuda=%s; CUDA 12 runtime=%s; cuDNN 9=%s\n' \
+  printf 'GPU preflight: GPU=%s; device=%s; driver=%s (minimum %s); libcuda=%s; CUDA 12 runtime=%s; CUDA libraries=%s; cuDNN 9=%s\n' \
     "${CUDA_GPU_NAME:-not detected}" \
     "$(if [ "$CUDA_DEVICE_VISIBLE" -eq 1 ]; then printf visible; else printf missing; fi)" \
     "${CUDA_DRIVER_VERSION:-not detected}" \
     "$CUDA_MINIMUM_DRIVER" \
     "$(if [ "$CUDA_LIBCUDA_PRESENT" -eq 1 ]; then printf present; else printf missing; fi)" \
     "$(if [ "$CUDA_RUNTIME_PRESENT" -eq 1 ]; then printf present; else printf missing; fi)" \
+    "$(if [ "$CUDA_LIBRARIES_PRESENT" -eq 1 ]; then printf present; else printf missing; fi)" \
     "$(if [ "$CUDA_CUDNN_PRESENT" -eq 1 ]; then printf present; else printf missing; fi)"
 }
 
@@ -225,6 +236,10 @@ print_cuda_host_actions() {
     printf 'Action: install or expose the CUDA 12 runtime (libcudart.so.12; package %s).\n' \
       "$CUDA_RUNTIME_PACKAGE"
   fi
+  if [ "$CUDA_LIBRARIES_PRESENT" -eq 0 ]; then
+    printf 'Action: install or expose CUDA 12 math libraries (libcublasLt.so.12 and libcurand.so.10; package %s).\n' \
+      "$CUDA_LIBRARIES_PACKAGE"
+  fi
   if [ "$CUDA_DEVICE_VISIBLE" -eq 1 ] &&
      [ "$CUDA_DRIVER_PRESENT" -eq 1 ] &&
      [ "$CUDA_DRIVER_COMPATIBLE" -eq 1 ] &&
@@ -402,6 +417,11 @@ cuda_configure_apt_repositories() {
       "$CUDA_RUNTIME_PACKAGE" >&2
     return 1
   }
+  cuda_apt_package_available "$CUDA_LIBRARIES_PACKAGE" || {
+    printf 'ERROR: NVIDIA CUDA repository still has no candidate for %s.\n' \
+      "$CUDA_LIBRARIES_PACKAGE" >&2
+    return 1
+  }
   cuda_apt_package_available "$CUDA_CUDNN_PACKAGE" || {
     printf 'ERROR: NVIDIA CUDA repository still has no candidate for %s.\n' \
       "$CUDA_CUDNN_PACKAGE" >&2
@@ -437,12 +457,21 @@ maybe_install_cuda_dependencies() {
   [ "$CUDA_DEVICE_VISIBLE" -eq 1 ] &&
     [ "$CUDA_DRIVER_PRESENT" -eq 1 ] &&
     [ "$CUDA_DRIVER_COMPATIBLE" -eq 1 ] &&
-    { [ "$CUDA_RUNTIME_PRESENT" -eq 0 ] || [ "$CUDA_CUDNN_PRESENT" -eq 0 ]; } || return 0
+    { [ "$CUDA_RUNTIME_PRESENT" -eq 0 ] ||
+      [ "$CUDA_LIBRARIES_PRESENT" -eq 0 ] ||
+      [ "$CUDA_CUDNN_PRESENT" -eq 0 ]; } || return 0
 
   CUDA_INSTALL_PACKAGES=""
   if [ "$CUDA_RUNTIME_PRESENT" -eq 0 ]; then
     CUDA_INSTALL_PACKAGES="$CUDA_RUNTIME_PACKAGE"
   fi
+  if [ "$CUDA_LIBRARIES_PRESENT" -eq 0 ]; then
+    if [ -n "$CUDA_INSTALL_PACKAGES" ]; then
+      CUDA_INSTALL_PACKAGES="$CUDA_INSTALL_PACKAGES $CUDA_LIBRARIES_PACKAGE"
+    else
+      CUDA_INSTALL_PACKAGES="$CUDA_LIBRARIES_PACKAGE"
+    fi
+  fi
   if [ "$CUDA_CUDNN_PRESENT" -eq 0 ]; then
     if [ -n "$CUDA_INSTALL_PACKAGES" ]; then
       CUDA_INSTALL_PACKAGES="$CUDA_INSTALL_PACKAGES $CUDA_CUDNN_PACKAGE"
@@ -536,6 +565,7 @@ if [ "$INSTALL_SEMANTIC" -eq 1 ]; then
           print_cuda_host_actions
           fail "CUDA was requested but the host preflight is incomplete; fix the reported prerequisites or set LLM_CONTEXT_ACCELERATOR_PACKAGE=cpu"
         }
+        RUNTIME_FLAVOR=cuda
       elif [ "$RUNTIME_FLAVOR_REQUESTED" = "auto" ]; then
         cuda_host_preflight
         maybe_install_cuda_dependencies

package.json

@@ -1,6 +1,6 @@
 {
   "name": "llm-context",
-  "version": "0.12.6",
+  "version": "0.12.9",
   "description": "Datalevin-backed semantic code graph and context generator for AI assistants",
   "type": "module",
   "bin": {

resources/llm_context/dependencies.edn

@@ -6,7 +6,7 @@
  ;; agree with this file; script/verify-dependencies.clj enforces that rule.
  :project
  {:name "llm-context"
-  :version "0.12.6"
+  :version "0.12.9"
   :repository "devame/llm-context-tools"}
 
  :toolchain
@@ -85,6 +85,7 @@
    :cudnn-version "9"
    :minimum-driver "525.60.13"
    :runtime-debian-package "cuda-cudart-12-9"
+   :libraries-debian-package "cuda-libraries-12-9"
    :debian-package "cudnn9-cuda-12"}}
 
  :github

scripts/install-dependencies.sh

@@ -671,6 +671,9 @@ cuda_library_present() {
   for candidate in \
     "/usr/lib/wsl/lib/$library" \
     "/usr/local/cuda/lib64/$library" \
+    "/usr/local/cuda/targets/x86_64-linux/lib/$library" \
+    "/usr/local/cuda-12.9/lib64/$library" \
+    "/usr/local/cuda-12.9/targets/x86_64-linux/lib/$library" \
     "/usr/lib/x86_64-linux-gnu/$library" \
     "/usr/lib64/$library"; do
     [[ -f "$candidate" ]] && return 0
@@ -688,9 +691,11 @@ cuda_host_preflight() {
   CUDA_DRIVER_PRESENT=0
   CUDA_DRIVER_COMPATIBLE=0
   CUDA_RUNTIME_PRESENT=0
+  CUDA_LIBRARIES_PRESENT=0
   CUDA_CUDNN_PRESENT=0
   CUDA_MINIMUM_DRIVER=$(cuda_manifest_value ':minimum-driver')
   CUDA_RUNTIME_PACKAGE=$(cuda_manifest_value ':runtime-debian-package')
+  CUDA_LIBRARIES_PACKAGE=$(cuda_manifest_value ':libraries-debian-package')
   CUDA_CUDNN_PACKAGE=$(cuda_manifest_value ':debian-package')
 
   if [[ -n "$CUDA_NVIDIA_SMI" ]]; then
@@ -711,18 +716,23 @@ cuda_host_preflight() {
     CUDA_DEVICE_VISIBLE=1
   fi
   cuda_library_present libcudart.so.12 && CUDA_RUNTIME_PRESENT=1 || true
+  if cuda_library_present libcublasLt.so.12 &&
+     cuda_library_present libcurand.so.10; then
+    CUDA_LIBRARIES_PRESENT=1
+  fi
   cuda_library_present libcudnn.so.9 && CUDA_CUDNN_PRESENT=1 || true
 }
 
 cuda_host_report() {
-  printf 'GPU: %s; device: %s; NVIDIA driver: %s (%s, minimum %s); CUDA 12 runtime: %s; cuDNN 9: %s\n' \
+  printf 'GPU: %s; device: %s; NVIDIA driver: %s (%s, minimum %s); CUDA 12 runtime: %s; CUDA libraries: %s; cuDNN 9: %s\n' \
     "${CUDA_GPU_NAME:-not detected}" \
     "$([[ "$CUDA_DEVICE_VISIBLE" -eq 1 ]] && printf visible || printf missing)" \
     "${CUDA_DRIVER_VERSION:-not detected}" \
     "$([[ "$CUDA_DRIVER_PRESENT" -eq 0 ]] && printf missing || \
        ([[ "$CUDA_DRIVER_COMPATIBLE" -eq 1 ]] && printf compatible || printf too-old))" \
     "$CUDA_MINIMUM_DRIVER" \
     "$([[ "$CUDA_RUNTIME_PRESENT" -eq 1 ]] && printf present || printf missing)" \
+    "$([[ "$CUDA_LIBRARIES_PRESENT" -eq 1 ]] && printf present || printf missing)" \
     "$([[ "$CUDA_CUDNN_PRESENT" -eq 1 ]] && printf present || printf missing)"
 }
 
@@ -906,12 +916,17 @@ cuda_configure_apt_repositories() {
       "$CUDA_RUNTIME_PACKAGE" >&2
     return 1
   }
+  cuda_apt_package_available "$CUDA_LIBRARIES_PACKAGE" || {
+    printf 'ERROR: NVIDIA CUDA repository still has no candidate for %s.\n' \
+      "$CUDA_LIBRARIES_PACKAGE" >&2
+    return 1
+  }
   cuda_apt_package_available "$CUDA_CUDNN_PACKAGE" || {
     printf 'ERROR: NVIDIA CUDA repository still has no candidate for %s.\n' \
       "$CUDA_CUDNN_PACKAGE" >&2
     return 1
   }
-  notify "NVIDIA CUDA repositories configured; both package candidates are available."
+  notify "NVIDIA CUDA repositories configured; all required package candidates are available."
 }
 
 cuda_host_step() {
@@ -928,8 +943,10 @@ cuda_host_step() {
     notify "CUDA host installation not offered: a visible GPU and compatible NVIDIA driver are required."
     return 0
   fi
-  if [[ "$CUDA_RUNTIME_PRESENT" -eq 1 && "$CUDA_CUDNN_PRESENT" -eq 1 ]]; then
-    notify "CUDA 12 runtime and cuDNN 9 host libraries verified."
+  if [[ "$CUDA_RUNTIME_PRESENT" -eq 1 &&
+        "$CUDA_LIBRARIES_PRESENT" -eq 1 &&
+        "$CUDA_CUDNN_PRESENT" -eq 1 ]]; then
+    notify "CUDA 12 runtime, CUDA math libraries, and cuDNN 9 host libraries verified."
     return 0
   fi
   if [[ "$ACTION" != install ]]; then
@@ -939,6 +956,7 @@ cuda_host_step() {
 
   cuda_install_packages=()
   [[ "$CUDA_RUNTIME_PRESENT" -eq 1 ]] || cuda_install_packages+=("$CUDA_RUNTIME_PACKAGE")
+  [[ "$CUDA_LIBRARIES_PRESENT" -eq 1 ]] || cuda_install_packages+=("$CUDA_LIBRARIES_PACKAGE")
   [[ "$CUDA_CUDNN_PRESENT" -eq 1 ]] || cuda_install_packages+=("$CUDA_CUDNN_PACKAGE")
   printf 'CUDA action: install missing host packages: %s\n' \
     "${cuda_install_packages[*]}"
@@ -984,11 +1002,13 @@ cuda_host_step() {
     retry "Install CUDA host packages" 2 5 sudo apt-get -y install "${cuda_install_packages[@]}"
   fi
   cuda_host_preflight
-  [[ "$CUDA_RUNTIME_PRESENT" -eq 1 && "$CUDA_CUDNN_PRESENT" -eq 1 ]] || {
-    printf 'ERROR: CUDA host package command completed but libcudart.so.12/cuDNN 9 verification failed.\n' >&2
+  [[ "$CUDA_RUNTIME_PRESENT" -eq 1 &&
+     "$CUDA_LIBRARIES_PRESENT" -eq 1 &&
+     "$CUDA_CUDNN_PRESENT" -eq 1 ]] || {
+    printf 'ERROR: CUDA host package command completed but libcudart.so.12, CUDA math libraries, or cuDNN 9 verification failed.\n' >&2
     return 1
   }
-  notify "Verified libcudart.so.12 and libcudnn.so.9 after installation."
+  notify "Verified libcudart.so.12, CUDA math libraries, and libcudnn.so.9 after installation."
 }
 
 local_native_platform() {