@FIR-1032 - LLama.cpp: wrong model cause crash the llama-cli process

Author: Anoop Kapoor

ggml/include/ggml-tsavorite.h

@@ -214,6 +214,8 @@ extern void ggml_tsi_log_tensor_data(tensor_log log_data);
 // GGML supports tensors with a maximum rank of 4
 #define MEM_REF_DESCRIPTOR_RANK 4
 #define TSI_TVU_MEM_ALIGN 128
+void
+ggml_tsi_finalize();
 
 //
 // backend API

ggml/src/ggml-tsavorite/ggml-tsavorite.cpp

@@ -32,10 +32,10 @@ namespace tsirt = ::tsi::runtime;
 typedef struct _txe_device_t *txe_device_s;
 typedef struct _txe_compute_pipeline_state_t *txe_compute_pipeline_state_s;
 FILE *tsi_op_log_file;
+bool runtime_initialized = false;
 uint64_t num_of_op;
 // Centralized TSI runtime initialization - called once globally
 static void ensure_tsi_runtime_initialized() {
-  static bool runtime_initialized = false;
   if (!runtime_initialized) {
     std::string mainProfilerName = "OPU ";
     tsirt::utils::TSIProfiler::initialize();
@@ -693,6 +693,8 @@ static struct ggml_backend_tsavorite_context *ggml_tsavorite_init(ggml_backend_d
 static void ggml_tsavorite_free(struct ggml_backend_tsavorite_context *ctx) {
   GGML_TSAVORITE_LOG_INFO("Start %s\n", __func__);
 
+  if (!ctx)
+      return;
 
   for (int i = 0; i < GGML_TSAVORITE_KERNEL_TYPE_COUNT; ++i) {
     if (ctx->kernels[i].pipeline) {
@@ -724,6 +726,25 @@ static void ggml_tsavorite_free(struct ggml_backend_tsavorite_context *ctx) {
   std::cout << tsirt::utils::TSIProfiler::getFormattedResults(
                    /*truncateFuncNames*/ true)
             << std::endl;
+  sleep(2);
+}
+
+void
+ggml_tsi_finalize() {
+  //if (tsi_op_log_file == NULL)
+  if (runtime_initialized != true)
+      return;
+  sleep(2);
+  tsi_finalize();
+  GGML_TSAVORITE_LOG_INFO("Start %s\n", __func__);
+  tsirt::utils::TSIProfiler::finalize();
+  std::cout << "\nOPU Profiling Results:" << std::endl;
+  std::cout << tsirt::utils::TSIProfiler::getFormattedResults(
+                   /*truncateFuncNames*/ true)
+            << std::endl;
+  sleep(2);
+  GGML_TSAVORITE_LOG_INFO("End %s\n", __func__);
+  return;
 }
 
 #if 0

tools/main/main.cpp

@@ -14,6 +14,7 @@
 #include <sstream>
 #include <string>
 #include <vector>
+#include <ggml-tsavorite.h>
 
 #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
 #include <signal.h>
@@ -153,6 +154,7 @@ int main(int argc, char ** argv) {
 
     if (model == NULL) {
         LOG_ERR("%s: error: unable to load model\n", __func__);
+	ggml_tsi_finalize();
         return 1;
     }