CUDA mode is pretty fast.

Author: calvertdw

ihmc-high-level-behaviors/build.gradle.kts

@@ -18,8 +18,8 @@ mainDependencies {
       exclude(group = "org.lwjgl.lwjgl") // exclude lwjgl 2
    }
    api("us.ihmc:promp-java:1.0.1")
-   api("de.kherud:llama:3.4.1")
-//   api("de.kherud:llama:3.4.1:cuda12-linux-x86-64")
+//   api("de.kherud:llama:3.4.1") // CPU mode
+   api("de.kherud:llama:3.4.1:cuda12-linux-x86-64")
 }
 
 libgdxDependencies {

ihmc-high-level-behaviors/src/main/java/us/ihmc/behaviors/reasoning/BehaviorTreeNextActionReasoning.java

@@ -4,6 +4,7 @@
 import de.kherud.llama.LlamaModel;
 import de.kherud.llama.ModelParameters;
 import de.kherud.llama.args.MiroStat;
+import us.ihmc.commons.time.Stopwatch;
 import us.ihmc.log.LogTools;
 import us.ihmc.tools.IHMCCommonPaths;
 
@@ -122,9 +123,11 @@ public int queryNextLeafToExecuteIndex()
 
       String reponse = model.complete(inferParams);
 
-      LogTools.info(prompt + reponse);
+//      LogTools.info(prompt + reponse);
+//
+//      LogTools.info("Response: {}", reponse);
 
-      return 0;
+      return Integer.parseInt(reponse.trim());
    }
 
    public void destroy()
@@ -135,7 +138,14 @@ public void destroy()
    public static void main(String[] args)
    {
       BehaviorTreeNextActionReasoning reasoning = new BehaviorTreeNextActionReasoning();
-      reasoning.queryNextLeafToExecuteIndex();
+
+      for  (int i = 0; i < 10; i++)
+      {
+         Stopwatch stopwatch = new Stopwatch().start();
+         int leafIndex = reasoning.queryNextLeafToExecuteIndex();
+         LogTools.info("Returned {} in {} seconds", leafIndex, stopwatch.totalElapsed());
+      }
+
       reasoning.destroy();
 
       System.exit(0); // FIXME: Not sure why it's not exiting automatically.