[breaking] Change internal model serialization to UBJSON. (#7556)

* Use typed array for models.
* Change the memory snapshot format.
* Add new C API for saving to raw format.

include/xgboost/c_api.h

@@ -1081,14 +1081,32 @@ XGB_DLL int XGBoosterSaveModel(BoosterHandle handle,
 XGB_DLL int XGBoosterLoadModelFromBuffer(BoosterHandle handle,
                                          const void *buf,
                                          bst_ulong len);
+
 /*!
- * \brief save model into binary raw bytes, return header of the array
- * user must copy the result out, before next xgboost call
+ * \brief Save model into raw bytes, return header of the array.  User must copy the
+ *        result out, before next xgboost call
+ *
  * \param handle handle
- * \param out_len the argument to hold the output length
- * \param out_dptr the argument to hold the output data pointer
+ * \param json_config JSON encoded string storing parameters for the function.  Following
+ *                    keys are expected in the JSON document:
+ *
+ *     "format": str
+ *       - json: Output booster will be encoded as JSON.
+ *       - ubj:  Output booster will be encoded as Univeral binary JSON.
+ *       - deprecated: Output booster will be encoded as old custom binary format.  Do not use
+ *         this format except for compatibility reasons.
+ *
+ * \param out_len  The argument to hold the output length
+ * \param out_dptr The argument to hold the output data pointer
+ *
  * \return 0 when success, -1 when failure happens
  */
+XGB_DLL int XGBoosterSaveModelToBuffer(BoosterHandle handle, char const *json_config,
+                                       bst_ulong *out_len, char const **out_dptr);
+
+/*!
+ * \brief Deprecated, use `XGBoosterSaveModelToBuffer` instead.
+ */
 XGB_DLL int XGBoosterGetModelRaw(BoosterHandle handle, bst_ulong *out_len,
                                  const char **out_dptr);
 

include/xgboost/data.h

@@ -1,5 +1,5 @@
 /*!
- * Copyright (c) 2015-2021 by Contributors
+ * Copyright (c) 2015-2022 by Contributors
  * \file data.h
  * \brief The input data structure of xgboost.
  * \author Tianqi Chen
@@ -36,10 +36,7 @@ enum class DataType : uint8_t {
   kStr = 5
 };
 
-enum class FeatureType : uint8_t {
-  kNumerical,
-  kCategorical
-};
+enum class FeatureType : uint8_t { kNumerical = 0, kCategorical = 1 };
 
 /*!
  * \brief Meta information about dataset, always sit in memory.

include/xgboost/linalg.h

@@ -1,5 +1,5 @@
 /*!
- * Copyright 2021 by XGBoost Contributors
+ * Copyright 2021-2022 by XGBoost Contributors
  * \file linalg.h
  * \brief Linear algebra related utilities.
  */
@@ -567,7 +567,7 @@ template <typename T, int32_t D>
 Json ArrayInterface(TensorView<T const, D> const &t) {
   Json array_interface{Object{}};
   array_interface["data"] = std::vector<Json>(2);
-  array_interface["data"][0] = Integer(reinterpret_cast<int64_t>(t.Values().data()));
+  array_interface["data"][0] = Integer{reinterpret_cast<int64_t>(t.Values().data())};
   array_interface["data"][1] = Boolean{true};
   if (t.DeviceIdx() >= 0) {
     // Change this once we have different CUDA stream.

include/xgboost/tree_model.h

@@ -1,5 +1,5 @@
 /*!
- * Copyright 2014-2019 by Contributors
+ * Copyright 2014-2022 by Contributors
  * \file tree_model.h
  * \brief model structure for tree
  * \author Tianqi Chen
@@ -42,7 +42,7 @@ struct TreeParam : public dmlc::Parameter<TreeParam> {
   /*! \brief maximum depth, this is a statistics of the tree */
   int deprecated_max_depth;
   /*! \brief number of features used for tree construction */
-  int num_feature;
+  bst_feature_t num_feature;
   /*!
    * \brief leaf vector size, used for vector tree
    * used to store more than one dimensional information in tree
@@ -629,6 +629,7 @@ class RegTree : public Model {
   }
 
  private:
+  template <bool typed>
   void LoadCategoricalSplit(Json const& in);
   void SaveCategoricalSplit(Json* p_out) const;
   // vector of nodes