De-duplicate GPU parameters. (#4454)

* Only define `gpu_id` and `n_gpus` in `LearnerTrainParam`
* Pass LearnerTrainParam through XGBoost vid factory method.
* Disable all GPU usage when GPU related parameters are not specified (fixes XGBoost choosing GPU over aggressively).
* Test learner train param io.
* Fix gpu pickling.

src/predictor/cpu_predictor.cc

@@ -82,7 +82,7 @@ class CPUPredictor : public Predictor {
       for (bst_omp_uint i = nsize - rest; i < nsize; ++i) {
         RegTree::FVec& feats = thread_temp[0];
         const auto ridx = static_cast<int64_t>(batch.base_rowid + i);
-         auto inst = batch[i];
+        auto inst = batch[i];
         for (int gid = 0; gid < num_group; ++gid) {
           const size_t offset = ridx * num_group + gid;
           preds[offset] +=

src/predictor/gpu_predictor.cu

@@ -20,20 +20,6 @@ namespace predictor {
 
 DMLC_REGISTRY_FILE_TAG(gpu_predictor);
 
-/*! \brief prediction parameters */
-struct GPUPredictionParam : public dmlc::Parameter<GPUPredictionParam> {
-  int gpu_id;
-  int n_gpus;
-  // declare parameters
-  DMLC_DECLARE_PARAMETER(GPUPredictionParam) {
-    DMLC_DECLARE_FIELD(gpu_id).set_lower_bound(0).set_default(0).describe(
-        "Device ordinal for GPU prediction.");
-    DMLC_DECLARE_FIELD(n_gpus).set_lower_bound(-1).set_default(1).describe(
-        "Number of devices to use for prediction.");
-  }
-};
-DMLC_REGISTER_PARAMETER(GPUPredictionParam);
-
 template <typename IterT>
 void IncrementOffset(IterT begin_itr, IterT end_itr, size_t amount) {
   thrust::transform(begin_itr, end_itr, begin_itr,
@@ -387,14 +373,15 @@ class GPUPredictor : public xgboost::Predictor {
   }
 
  public:
-  GPUPredictor()                                               // NOLINT
-      : cpu_predictor_(Predictor::Create("cpu_predictor")) {}  // NOLINT
+  GPUPredictor()
+      : cpu_predictor_(Predictor::Create("cpu_predictor", learner_param_)) {}
 
   void PredictBatch(DMatrix* dmat, HostDeviceVector<bst_float>* out_preds,
                     const gbm::GBTreeModel& model, int tree_begin,
                     unsigned ntree_limit = 0) override {
-    GPUSet devices = GPUSet::All(
-        param_.gpu_id, param_.n_gpus, dmat->Info().num_row_);
+    GPUSet devices = GPUSet::All(learner_param_->gpu_id, learner_param_->n_gpus,
+                                 dmat->Info().num_row_);
+    CHECK_NE(devices.Size(), 0);
     ConfigureShards(devices);
 
     if (this->PredictFromCache(dmat, out_preds, model, ntree_limit)) {
@@ -508,9 +495,8 @@ class GPUPredictor : public xgboost::Predictor {
             const std::vector<std::shared_ptr<DMatrix>>& cache) override {
     Predictor::Init(cfg, cache);
     cpu_predictor_->Init(cfg, cache);
-    param_.InitAllowUnknown(cfg);
 
-    GPUSet devices = GPUSet::All(param_.gpu_id, param_.n_gpus);
+    GPUSet devices = GPUSet::All(learner_param_->gpu_id, learner_param_->n_gpus);
     ConfigureShards(devices);
   }
 
@@ -527,7 +513,6 @@ class GPUPredictor : public xgboost::Predictor {
       });
   }
 
-  GPUPredictionParam param_;
   std::unique_ptr<Predictor> cpu_predictor_;
   std::vector<DeviceShard> shards_;
   GPUSet devices_;

src/predictor/predictor.cc

@@ -15,12 +15,14 @@ void Predictor::Init(
     cache_[d.get()].data = d;
   }
 }
-Predictor* Predictor::Create(std::string name) {
+Predictor* Predictor::Create(std::string const& name, LearnerTrainParam const* learner_param) {
   auto* e = ::dmlc::Registry<PredictorReg>::Get()->Find(name);
   if (e == nullptr) {
     LOG(FATAL) << "Unknown predictor type " << name;
   }
-  return (e->body)();
+  auto p_predictor =  (e->body)();
+  p_predictor->learner_param_ = learner_param;
+  return p_predictor;
 }
 }  // namespace xgboost