Don't use compat libcuda on CPU-only systems (triton-inference-server…

…#594)
Marshall2M · Aug 28, 2019 · 265fa41 · 265fa41
1 parent 110a760
commit 265fa41
Show file tree

Hide file tree

Showing 6 changed files with 9 additions and 19 deletions.
diff --git a/nvidia_entrypoint.sh b/nvidia_entrypoint.sh
@@ -46,9 +46,7 @@ if [[ "$(find /usr -name libcuda.so.1 | grep -v "compat") " == " " || "$(ls /dev
   echo "WARNING: The NVIDIA Driver was not detected.  GPU functionality will not be available."
   echo "   Use 'nvidia-docker run' to start this container; see"
   echo "   https://github.com/NVIDIA/nvidia-docker/wiki/nvidia-docker ."
-  ln -s `find / -name libcuda.so.1 -print -quit` /opt/tensorrtserver/lib/libcuda.so.1
   ln -s `find / -name libnvidia-ml.so -print -quit` /opt/tensorrtserver/lib/libnvidia-ml.so.1
-  ln -s `find / -name libnvidia-fatbinaryloader.so.${CUDA_DRIVER_VERSION} -print -quit` /opt/tensorrtserver/lib/libnvidia-fatbinaryloader.so.${CUDA_DRIVER_VERSION}
   export TENSORRT_SERVER_CPU_ONLY=1
 else
   ( /usr/local/bin/checkSMVER.sh )

diff --git a/src/core/backend_context.cc b/src/core/backend_context.cc
@@ -61,7 +61,7 @@ BackendContext::CreateCudaStream(const int cuda_stream_priority)
   auto cuerr = cudaGetDeviceCount(&device_cnt);
   // Do nothing if there is no CUDA device since all data transfer will be done
   // within CPU memory
-  if (cuerr != cudaErrorNoDevice) {
+  if ((cuerr != cudaErrorNoDevice) && (cuerr != cudaErrorInsufficientDriver)) {
     if (cuerr == cudaSuccess) {
       cuerr = cudaStreamCreateWithPriority(
           &stream_, cudaStreamDefault, cuda_stream_priority);

diff --git a/src/core/model_config_utils.cc b/src/core/model_config_utils.cc
@@ -946,12 +946,12 @@ GetSupportedGPUs(std::set<int>& supported_gpus)
 
   int device_cnt;
   cudaError_t cuerr = cudaGetDeviceCount(&device_cnt);
-  if (cuerr == cudaErrorNoDevice) {
+  if ((cuerr == cudaErrorNoDevice) || (cuerr == cudaErrorInsufficientDriver)) {
     device_cnt = 0;
   } else if (cuerr != cudaSuccess) {
     return Status(
         RequestStatusCode::INTERNAL,
-        "unable to get number of CUDA devices" +
+        "unable to get number of CUDA devices: " +
             std::string(cudaGetErrorString(cuerr)));
   }
 

diff --git a/src/core/model_repository_manager.cc b/src/core/model_repository_manager.cc
@@ -93,12 +93,13 @@ BuildBackendConfigMap(
 #ifdef TRTIS_ENABLE_GPU
     int device_cnt = 0;
     cudaError_t cuerr = cudaGetDeviceCount(&device_cnt);
-    if (cuerr == cudaErrorNoDevice) {
+    if ((cuerr == cudaErrorNoDevice) ||
+        (cuerr == cudaErrorInsufficientDriver)) {
       device_cnt = 0;
     } else if (cuerr != cudaSuccess) {
       LOG_ERROR << "unable to get number of CUDA devices while building "
-                   "BackendConfigMap: "
-                << cudaGetErrorString(cuerr);
+                   "BackendConfigMap: ("
+                << cuerr << ") " << cudaGetErrorString(cuerr);
       device_cnt = 0;
     }
 

diff --git a/src/core/profile.cc b/src/core/profile.cc
@@ -39,7 +39,7 @@ ProfileStartAll()
 #ifdef TRTIS_ENABLE_GPU
   int dcnt;
   cudaError_t cuerr = cudaGetDeviceCount(&dcnt);
-  if (cuerr == cudaErrorNoDevice) {
+  if ((cuerr == cudaErrorNoDevice) || (cuerr == cudaErrorInsufficientDriver)) {
     dcnt = 0;
   } else if (cuerr != cudaSuccess) {
     return Status(
@@ -76,7 +76,7 @@ ProfileStopAll()
 #ifdef TRTIS_ENABLE_GPU
   int dcnt;
   cudaError_t cuerr = cudaGetDeviceCount(&dcnt);
-  if (cuerr == cudaErrorNoDevice) {
+  if ((cuerr == cudaErrorNoDevice) || (cuerr == cudaErrorInsufficientDriver)) {
     dcnt = 0;
   } else if (cuerr != cudaSuccess) {
     return Status(

diff --git a/src/core/server.cc b/src/core/server.cc
@@ -44,7 +44,6 @@
 #include "src/core/model_config.pb.h"
 #include "src/core/model_config_utils.h"
 #include "src/core/model_repository_manager.h"
-#include "src/core/profile.h"
 #include "src/core/provider.h"
 #include "src/core/server.h"
 #include "src/core/server_status.pb.h"
@@ -116,14 +115,6 @@ InferenceServer::Init()
         RequestStatusCode::INVALID_ARG, "--model-repository must be specified");
   }
 
-  // Disable profiling at server start. Server API can be used to
-  // start/stop profiling.
-  status = ProfileStopAll();
-  if (!status.IsOk()) {
-    ready_state_ = ServerReadyState::SERVER_FAILED_TO_INITIALIZE;
-    return status;
-  }
-
   // Create the shared memory manager that registers / unregisters and returns
   // the shared memory regions that are current registered.
   status =