Adding deterministic mode (unfortunately that did not make it deterministic on GPU -> general problem with ONNX)

Author: ChSonnabend

Common/ML/include/ML/OrtInterface.h

@@ -116,7 +116,7 @@ class OrtModel
   int32_t mInputsTotal = 0, mOutputsTotal = 0;                                                        // Total number of inputs and outputs
 
   // Environment settings
-  bool mInitialized = false;
+  bool mInitialized = false, mDeterministicMode = false;
   std::string mModelPath, mEnvName = "", mDeviceType = "CPU", mThreadAffinity = ""; // device options should be cpu, rocm, migraphx, cuda
   int32_t mIntraOpNumThreads = 1, mInterOpNumThreads = 1, mDeviceId = -1, mEnableProfiling = 0, mLoggingLevel = 0, mAllocateDeviceMemory = 0, mEnableOptimizations = 0;
 

Common/ML/src/OrtInterface.cxx

@@ -68,8 +68,10 @@ void OrtModel::initOptions(std::unordered_map<std::string, std::string> optionsM
     mEnableProfiling = (optionsMap.contains("enable-profiling") ? std::stoi(optionsMap["enable-profiling"]) : 0);
     mEnableOptimizations = (optionsMap.contains("enable-optimizations") ? std::stoi(optionsMap["enable-optimizations"]) : 0);
     mEnvName = (optionsMap.contains("onnx-environment-name") ? optionsMap["onnx-environment-name"] : "onnx_model_inference");
+    mDeterministicMode = (optionsMap.contains("deterministic-compute") ? std::stoi(optionsMap["deterministic-compute"]) : 0);
 
-    if (mDeviceType == "CPU") {
+      if (mDeviceType == "CPU")
+    {
       (mPImplOrt->sessionOptions).SetIntraOpNumThreads(mIntraOpNumThreads);
       (mPImplOrt->sessionOptions).SetInterOpNumThreads(mInterOpNumThreads);
       if (mIntraOpNumThreads > 1 || mInterOpNumThreads > 1) {
@@ -99,6 +101,10 @@ void OrtModel::initOptions(std::unordered_map<std::string, std::string> optionsM
       (mPImplOrt->sessionOptions).DisableProfiling();
     }
 
+    if (mDeterministicMode > 0) {
+      (mPImplOrt->sessionOptions).AddConfigEntry("session_options.use_deterministic_compute", "1");
+    }
+
     (mPImplOrt->sessionOptions).SetGraphOptimizationLevel(GraphOptimizationLevel(mEnableOptimizations));
     (mPImplOrt->sessionOptions).SetLogSeverityLevel(OrtLoggingLevel(mLoggingLevel));
 

GPU/GPUTracking/Definitions/GPUSettingsList.h

@@ -256,6 +256,7 @@ AddOption(nnInferenceOutputDType, std::string, "FP32", "", 0, "(std::string) Spe
 AddOption(nnInferenceIntraOpNumThreads, int, 1, "", 0, "Number of threads used to evaluate one neural network (ONNX: SetIntraOpNumThreads). 0 = auto-detect, can lead to problems on SLURM systems.")
 AddOption(nnInferenceInterOpNumThreads, int, 1, "", 0, "Number of threads used to evaluate one neural network (ONNX: SetInterOpNumThreads). 0 = auto-detect, can lead to problems on SLURM systems.")
 AddOption(nnInferenceEnableOrtOptimization, unsigned int, 99, "", 0, "Enables graph optimizations in ONNX Runtime. Can be [0, 1, 2, 99] -> see https://github.com/microsoft/onnxruntime/blob/3f71d637a83dc3540753a8bb06740f67e926dc13/include/onnxruntime/core/session/onnxruntime_c_api.h#L347")
+AddOption(nnInferenceUseDeterministicCompute, int, 0, "", 0, "Enables deterministic compute in ONNX Runtime were possible. Can be [0, 1] -> see https://github.com/microsoft/onnxruntime/blob/3b97d79b3c12dbf93aa0d563f345714596dc8ab6/onnxruntime/core/framework/session_options.h#L208")
 AddOption(nnInferenceOrtProfiling, int, 0, "", 0, "Enables profiling of model execution in ONNX Runtime")
 AddOption(nnInferenceOrtProfilingPath, std::string, ".", "", 0, "If nnInferenceOrtProfiling is set, the path to store the profiling data")
 AddOption(nnInferenceVerbosity, int, 1, "", 0, "0: No messages; 1: Warnings; 2: Warnings + major debugs; >3: All debugs")

GPU/GPUTracking/TPCClusterFinder/GPUTPCNNClusterizerHost.cxx

@@ -54,6 +54,7 @@ void GPUTPCNNClusterizerHost::init(const GPUSettingsProcessingNNclusterizer& set
     {"intra-op-num-threads", std::to_string(settings.nnInferenceIntraOpNumThreads)},
     {"inter-op-num-threads", std::to_string(settings.nnInferenceInterOpNumThreads)},
     {"enable-optimizations", std::to_string(settings.nnInferenceEnableOrtOptimization)},
+    {"deterministic-compute", std::to_string(settings.nnInferenceUseDeterministicCompute)}, // TODO: This unfortunately doesn't guarantee determinism (25.07.2025)
     {"enable-profiling", std::to_string(settings.nnInferenceOrtProfiling)},
     {"profiling-output-path", settings.nnInferenceOrtProfilingPath},
     {"logging-level", std::to_string(settings.nnInferenceVerbosity)},