Fixed LLAMA generation bug (#14320)

* fixed LLAMA generation bug

* update params

Author: prabod

src/main/scala/com/johnsnowlabs/ml/ai/LLAMA2.scala

@@ -164,7 +164,7 @@ private[johnsnowlabs] class LLAMA2(
       randomSeed,
       ignoreTokenIdsInt,
       session,
-      applySoftmax = false,
+      applySoftmax = true,
       ovInferRequest = ovInferRequest)
 
     modelOutputs

src/main/scala/com/johnsnowlabs/ml/ai/util/Generation/Generate.scala

@@ -392,10 +392,7 @@ trait Generate {
         seededRandom = new scala.util.Random(seed.get)
       }
       for (i <- 0 until k) {
-        var rand = scala.util.Random.nextDouble()
-        if (seed.isDefined) {
-          rand = new scala.util.Random(seed.get).nextDouble()
-        }
+        val rand = seededRandom.nextDouble()
         var cumProb = 0.0
         var j = 0
         while (j < probabilities.length - i) {

src/main/scala/com/johnsnowlabs/ml/ai/util/Generation/Logit/LogitWarper/TopKLogitWarper.scala

@@ -20,27 +20,25 @@ import scala.collection.mutable.ArrayBuffer
 class TopKLogitWarper(
     val k: Int,
     val filterValue: Float = Float.NegativeInfinity,
-    val minTokensToKeep: Int = 1)
+    val minTokensToKeep: Int = 100)
     extends LogitWarper {
   override def call(
       inputIds: Seq[Array[Int]],
       scores: Array[Array[Float]],
       currentLength: Int): Array[Array[Float]] = {
-    var logitsUpd = scores
-    val logitsShape = Array(scores.length, scores(0).length)
-    if (k > 0) {
-      val topKup = k.max(minTokensToKeep).min(logitsShape.last) // Safety check
+    val logitsUpd = scores.map(_.clone()) // Deep copy of the scores
 
-      /** Remove all tokens with a probability less than the last token of the top-k */
+    if (k > 0) {
+      val logitsShape = Array(scores.length, scores.head.length)
+      val effectiveTopK = k.max(minTokensToKeep).min(logitsShape.last) // Safety check
 
-      val topKLogits = new ArrayBuffer[Array[Float]]()
-      for (logits <- scores) {
-        val topKIndices = getTopKIndices(logits, topKup)
+      for ((logits, i) <- scores.zipWithIndex) {
+        val topKIndices = getTopKIndices(logits, effectiveTopK)
         val maskedValues = maskNotTopKValues(logits, topKIndices)
-        topKLogits += maskedValues
+        logitsUpd(i) = maskedValues
       }
-      topKLogits.toArray
     }
+
     logitsUpd
   }
 

src/main/scala/com/johnsnowlabs/ml/ai/util/Generation/Logit/LogitWarper/TopPLogitWarper.scala

@@ -21,46 +21,34 @@ class TopPLogitWarper(val p: Double, val minTokensToKeep: Int = 1) extends Logit
       inputIds: Seq[Array[Int]],
       scores: Array[Array[Float]],
       currentLength: Int): Array[Array[Float]] = {
-    var scoresUpd = scores
-    val scoresShape = Array(scores.length, scores(0).length)
-    if (this.p < 1.0) {
-      val (sortedscores, sortedIndices) = scores(0).zipWithIndex.sorted.reverse.unzip
+    val logitsUpd = scores.map(_.clone()) // Deep copy of the scores
 
-      val cumulativeProbs = this.scanLeft(this.softmax(sortedscores))(0.0)(_ + _).drop(1)
+    if (p < 1.0) {
+      val scoresFiltered = scores.map(_.filterNot(_.isInfinite)) // Filter out infinite values
+      val scoresShape = Array(scoresFiltered.length, scoresFiltered.head.length)
+      val topPThreshold = math.ceil(p * scoresShape.last).toInt // Determine top-p threshold
 
-      /** Remove tokens with cumulative probability above the threshold (token with 0 are kept) */
-      var sortedIndicesToRemove =
-        for (prob <- cumulativeProbs)
-          yield if (prob > this.p) true else false
-
-      if (minTokensToKeep > 1) {
-
-        /** Keep at least minTokensToKeep (set to minTokensToKeep-1 because we add the first one
-          * below)
-          */
-        sortedIndicesToRemove = List.fill(sortedIndicesToRemove.take(minTokensToKeep).length)(
-          false) ++ sortedIndicesToRemove.drop(minTokensToKeep)
+      for ((logits, i) <- scores.zipWithIndex) {
+        val topPIndices = getTopPIndices(logits, topPThreshold)
+        val maskedValues = maskNotTopPValues(logits, topPIndices)
+        logitsUpd(i) = maskedValues
       }
+    }
+
+    logitsUpd
+  }
 
-      /** Shift the indices to the right to keep also the first token above the threshold */
-      sortedIndicesToRemove = sortedIndicesToRemove.takeRight(1) ++ sortedIndicesToRemove
-        .dropRight(1)
-      sortedIndicesToRemove =
-        List.fill(sortedIndicesToRemove.take(1).length)(false) ++ sortedIndicesToRemove
-          .drop(1)
+  private def getTopPIndices(logits: Array[Float], k: Int): Array[Int] = {
+    logits.zipWithIndex.sortBy(-_._1).take(k).map(_._2)
+  }
 
-      /** scatter sorted tensors to original indexing */
-      val indicesToRemove =
-        this.scatterValuesOnBatchIndices(sortedIndicesToRemove.toList, sortedIndices)
-      scoresUpd =
-        for ((nextTokenLogit, indexToRemove) <- scores.zip(
-            IndexedSeq.fill(scores.length)(indicesToRemove)))
-          yield setTensorByIndicesToValue(
-            nextTokenLogit,
-            indexToRemove.toIndexedSeq,
-            Float.NegativeInfinity)
+  private def maskNotTopPValues(logits: Array[Float], topPIndices: Array[Int]): Array[Float] = {
+    val maskedValues = logits.clone()
+    for (i <- logits.indices) {
+      if (!topPIndices.contains(i)) {
+        maskedValues(i) = Float.NegativeInfinity
+      }
     }
-    scoresUpd
+    maskedValues
   }
-
 }

src/main/scala/com/johnsnowlabs/nlp/annotators/seq2seq/LLAMA2Transformer.scala

@@ -227,11 +227,11 @@ class LLAMA2Transformer(override val uid: String)
     minOutputLength -> 0,
     maxOutputLength -> 20,
     doSample -> false,
-    temperature -> 0.6,
-    topK -> 50,
+    temperature -> 0.9,
+    topK -> 100,
     topP -> 0.9,
     repetitionPenalty -> 1.0,
-    noRepeatNgramSize -> 3,
+    noRepeatNgramSize -> 0,
     ignoreTokenIds -> Array(),
     batchSize -> 1,
     beamSize -> 1,