Sparknlp 876: Introducing LLAMA2 (#14148)

* introducing LLAMA2

* Added option to read model from model path to onnx wrapper

* Added option to read model from model path to onnx wrapper

* updated text description

* LLAMA2 python API

* added method to save onnx_data

* added position ids

* - updated Generate.scala to accept onnx tensors
- added beam search support for LLAMA2

* updated max input length

* updated python default params
changed test to slow test

* fixed serialization bug

python/sparknlp/annotator/seq2seq/__init__.py

@@ -17,3 +17,4 @@
 from sparknlp.annotator.seq2seq.marian_transformer import *
 from sparknlp.annotator.seq2seq.t5_transformer import *
 from sparknlp.annotator.seq2seq.bart_transformer import *
+from sparknlp.annotator.seq2seq.llama2_transformer import *

python/sparknlp/annotator/seq2seq/llama2_transformer.py

@@ -0,0 +1,343 @@
+#  Copyright 2017-2022 John Snow Labs
+#
+#  Licensed under the Apache License, Version 2.0 (the "License");
+#  you may not use this file except in compliance with the License.
+#  You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+#  Unless required by applicable law or agreed to in writing, software
+#  distributed under the License is distributed on an "AS IS" BASIS,
+#  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#  See the License for the specific language governing permissions and
+#  limitations under the License.
+"""Contains classes for the LLAMA2Transformer."""
+
+from sparknlp.common import *
+
+
+class LLAMA2Transformer(AnnotatorModel, HasBatchedAnnotate, HasEngine):
+    """Llama 2: Open Foundation and Fine-Tuned Chat Models
+
+    The Llama 2 release introduces a family of pretrained and fine-tuned LLMs, ranging in scale
+    from 7B to 70B parameters (7B, 13B, 70B). The pretrained models come with significant
+    improvements over the Llama 1 models, including being trained on 40% more tokens, having a
+    much longer context length (4k tokens 🤯), and using grouped-query attention for fast
+    inference of the 70B model🔥!
+
+    However, the most exciting part of this release is the fine-tuned models (Llama 2-Chat), which
+    have been optimized for dialogue applications using Reinforcement Learning from Human Feedback
+    (RLHF). Across a wide range of helpfulness and safety benchmarks, the Llama 2-Chat models
+    perform better than most open models and achieve comparable performance to ChatGPT according
+    to human evaluations.
+
+    Pretrained models can be loaded with :meth:`.pretrained` of the companion
+    object:
+
+    >>> llama2 = LLAMA2Transformer.pretrained() \\
+    ...     .setInputCols(["document"]) \\
+    ...     .setOutputCol("generation")
+
+
+    The default model is ``"llam2-7b"``, if no name is provided. For available
+    pretrained models please see the `Models Hub
+    <https://sparknlp.org/models?q=llama2>`__.
+
+    ====================== ======================
+    Input Annotation types Output Annotation type
+    ====================== ======================
+    ``DOCUMENT``           ``DOCUMENT``
+    ====================== ======================
+
+    Parameters
+    ----------
+    configProtoBytes
+        ConfigProto from tensorflow, serialized into byte array.
+    minOutputLength
+        Minimum length of the sequence to be generated, by default 0
+    maxOutputLength
+        Maximum length of output text, by default 20
+    doSample
+        Whether or not to use sampling; use greedy decoding otherwise, by default False
+    temperature
+        The value used to module the next token probabilities, by default 1.0
+    topK
+        The number of highest probability vocabulary tokens to keep for
+        top-k-filtering, by default 50
+    topP
+        Top cumulative probability for vocabulary tokens, by default 1.0
+
+        If set to float < 1, only the most probable tokens with probabilities
+        that add up to ``topP`` or higher are kept for generation.
+    repetitionPenalty
+        The parameter for repetition penalty, 1.0 means no penalty. , by default
+        1.0
+    noRepeatNgramSize
+        If set to int > 0, all ngrams of that size can only occur once, by
+        default 0
+    ignoreTokenIds
+        A list of token ids which are ignored in the decoder's output, by
+        default []
+
+    Notes
+    -----
+    This is a very computationally expensive module especially on larger
+    sequence. The use of an accelerator such as GPU is recommended.
+
+    References
+    ----------
+    - `Llama 2: Open Foundation and Fine-Tuned Chat Models
+      <https://ai.meta.com/research/publications/llama-2-open-foundation-and-fine-tuned-chat-models/>`__
+    - https://github.com/facebookresearch/llama
+
+    **Paper Abstract:**
+
+    *In this work, we develop and release Llama 2, a collection of pretrained and fine-tuned
+    large language models (LLMs) ranging in scale from 7 billion to 70 billion parameters. Our
+    fine-tuned LLMs, called Llama 2-Chat, are optimized for dialogue use cases. Our models
+    outperform open-source chat models on most benchmarks we tested, and based on our human
+    evaluations for helpfulness and safety, may be a suitable substitute for closed-source models.
+    We provide a detailed description of our approach to fine-tuning and safety improvements of
+    Llama 2-Chat in order to enable the community to build on our work and contribute to the
+    responsible development of LLMs.*
+
+    Examples
+    --------
+    >>> import sparknlp
+    >>> from sparknlp.base import *
+    >>> from sparknlp.annotator import *
+    >>> from pyspark.ml import Pipeline
+    >>> documentAssembler = DocumentAssembler() \\
+    ...     .setInputCol("text") \\
+    ...     .setOutputCol("documents")
+    >>> llama2 = LLAMA2Transformer.pretrained("llama2-7b") \\
+    ...     .setInputCols(["documents"]) \\
+    ...     .setMaxOutputLength(50) \\
+    ...     .setOutputCol("generation")
+    >>> pipeline = Pipeline().setStages([documentAssembler, llama2])
+    >>> data = spark.createDataFrame([["My name is Leonardo."]]).toDF("text")
+    >>> result = pipeline.fit(data).transform(data)
+    >>> result.select("summaries.generation").show(truncate=False)
+    +----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
+    |result                                                                                                                                                                                              |
+    +----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
+    |[My name is Leonardo. I am a man of letters. I have been a man for many years. I was born in the year 1776. I came to the United States in 1776, and I have lived in the United Kingdom since 1776.]|
+    -----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
+    """
+
+    name = "LLAMA2Transformer"
+
+    inputAnnotatorTypes = [AnnotatorType.DOCUMENT]
+
+    outputAnnotatorType = AnnotatorType.DOCUMENT
+
+
+    configProtoBytes = Param(Params._dummy(),
+                             "configProtoBytes",
+                             "ConfigProto from tensorflow, serialized into byte array. Get with config_proto.SerializeToString()",
+                             TypeConverters.toListInt)
+
+    minOutputLength = Param(Params._dummy(), "minOutputLength", "Minimum length of the sequence to be generated",
+                            typeConverter=TypeConverters.toInt)
+
+    maxOutputLength = Param(Params._dummy(), "maxOutputLength", "Maximum length of output text",
+                            typeConverter=TypeConverters.toInt)
+
+    doSample = Param(Params._dummy(), "doSample", "Whether or not to use sampling; use greedy decoding otherwise",
+                     typeConverter=TypeConverters.toBoolean)
+
+    temperature = Param(Params._dummy(), "temperature", "The value used to module the next token probabilities",
+                        typeConverter=TypeConverters.toFloat)
+
+    topK = Param(Params._dummy(), "topK",
+                 "The number of highest probability vocabulary tokens to keep for top-k-filtering",
+                 typeConverter=TypeConverters.toInt)
+
+    topP = Param(Params._dummy(), "topP",
+                 "If set to float < 1, only the most probable tokens with probabilities that add up to ``top_p`` or higher are kept for generation",
+                 typeConverter=TypeConverters.toFloat)
+
+    repetitionPenalty = Param(Params._dummy(), "repetitionPenalty",
+                              "The parameter for repetition penalty. 1.0 means no penalty. See `this paper <https://arxiv.org/pdf/1909.05858.pdf>`__ for more details",
+                              typeConverter=TypeConverters.toFloat)
+
+    noRepeatNgramSize = Param(Params._dummy(), "noRepeatNgramSize",
+                              "If set to int > 0, all ngrams of that size can only occur once",
+                              typeConverter=TypeConverters.toInt)
+
+    ignoreTokenIds = Param(Params._dummy(), "ignoreTokenIds",
+                           "A list of token ids which are ignored in the decoder's output",
+                           typeConverter=TypeConverters.toListInt)
+
+
+    def setIgnoreTokenIds(self, value):
+        """A list of token ids which are ignored in the decoder's output.
+
+        Parameters
+        ----------
+        value : List[int]
+            The words to be filtered out
+        """
+        return self._set(ignoreTokenIds=value)
+
+    def setConfigProtoBytes(self, b):
+        """Sets configProto from tensorflow, serialized into byte array.
+
+        Parameters
+        ----------
+        b : List[int]
+            ConfigProto from tensorflow, serialized into byte array
+        """
+        return self._set(configProtoBytes=b)
+
+    def setMinOutputLength(self, value):
+        """Sets minimum length of the sequence to be generated.
+
+        Parameters
+        ----------
+        value : int
+            Minimum length of the sequence to be generated
+        """
+        return self._set(minOutputLength=value)
+
+    def setMaxOutputLength(self, value):
+        """Sets maximum length of output text.
+
+        Parameters
+        ----------
+        value : int
+            Maximum length of output text
+        """
+        return self._set(maxOutputLength=value)
+
+    def setDoSample(self, value):
+        """Sets whether or not to use sampling, use greedy decoding otherwise.
+
+        Parameters
+        ----------
+        value : bool
+            Whether or not to use sampling; use greedy decoding otherwise
+        """
+        return self._set(doSample=value)
+
+    def setTemperature(self, value):
+        """Sets the value used to module the next token probabilities.
+
+        Parameters
+        ----------
+        value : float
+            The value used to module the next token probabilities
+        """
+        return self._set(temperature=value)
+
+    def setTopK(self, value):
+        """Sets the number of highest probability vocabulary tokens to keep for
+        top-k-filtering.
+
+        Parameters
+        ----------
+        value : int
+            Number of highest probability vocabulary tokens to keep
+        """
+        return self._set(topK=value)
+
+    def setTopP(self, value):
+        """Sets the top cumulative probability for vocabulary tokens.
+
+        If set to float < 1, only the most probable tokens with probabilities
+        that add up to ``topP`` or higher are kept for generation.
+
+        Parameters
+        ----------
+        value : float
+            Cumulative probability for vocabulary tokens
+        """
+        return self._set(topP=value)
+
+    def setRepetitionPenalty(self, value):
+        """Sets the parameter for repetition penalty. 1.0 means no penalty.
+
+        Parameters
+        ----------
+        value : float
+            The repetition penalty
+
+        References
+        ----------
+        See `Ctrl: A Conditional Transformer Language Model For Controllable
+        Generation <https://arxiv.org/pdf/1909.05858.pdf>`__ for more details.
+        """
+        return self._set(repetitionPenalty=value)
+
+    def setNoRepeatNgramSize(self, value):
+        """Sets size of n-grams that can only occur once.
+
+        If set to int > 0, all ngrams of that size can only occur once.
+
+        Parameters
+        ----------
+        value : int
+            N-gram size can only occur once
+        """
+        return self._set(noRepeatNgramSize=value)
+
+    @keyword_only
+    def __init__(self, classname="com.johnsnowlabs.nlp.annotators.seq2seq.LLAMA2Transformer", java_model=None):
+        super(LLAMA2Transformer, self).__init__(
+            classname=classname,
+            java_model=java_model
+        )
+        self._setDefault(
+            minOutputLength=0,
+            maxOutputLength=20,
+            doSample=False,
+            temperature=0.6,
+            topK=50,
+            topP=0.9,
+            repetitionPenalty=1.0,
+            noRepeatNgramSize=0,
+            ignoreTokenIds=[],
+            batchSize=1
+        )
+
+    @staticmethod
+    def loadSavedModel(folder, spark_session):
+        """Loads a locally saved model.
+
+        Parameters
+        ----------
+        folder : str
+            Folder of the saved model
+        spark_session : pyspark.sql.SparkSession
+            The current SparkSession
+
+        Returns
+        -------
+        LLAMA2Transformer
+            The restored model
+        """
+        from sparknlp.internal import _LLAMA2Loader
+        jModel = _LLAMA2Loader(folder, spark_session._jsparkSession)._java_obj
+        return LLAMA2Transformer(java_model=jModel)
+
+    @staticmethod
+    def pretrained(name="llama2-7b", lang="en", remote_loc=None):
+        """Downloads and loads a pretrained model.
+
+        Parameters
+        ----------
+        name : str, optional
+            Name of the pretrained model, by default "llama2-7b"
+        lang : str, optional
+            Language of the pretrained model, by default "en"
+        remote_loc : str, optional
+            Optional remote address of the resource, by default None. Will use
+            Spark NLPs repositories otherwise.
+
+        Returns
+        -------
+        LLAMA2Transformer
+            The restored model
+        """
+        from sparknlp.pretrained import ResourceDownloader
+        return ResourceDownloader.downloadModel(LLAMA2Transformer, name, lang, remote_loc)

python/sparknlp/internal/__init__.py

@@ -158,6 +158,10 @@ def __init__(self, path, jspark):
         super(_GPT2Loader, self).__init__(
             "com.johnsnowlabs.nlp.annotators.seq2seq.GPT2Transformer.loadSavedModel", path, jspark)
 
+class _LLAMA2Loader(ExtendedJavaWrapper):
+    def __init__(self, path, jspark):
+        super(_LLAMA2Loader, self).__init__(
+            "com.johnsnowlabs.nlp.annotators.seq2seq.LLAMA2Transformer.loadSavedModel", path, jspark)
 
 class _LongformerLoader(ExtendedJavaWrapper):
     def __init__(self, path, jspark):