Better message when no GPU is found. (#5594)

src/gbm/gbtree.cc

@@ -407,6 +407,7 @@ GBTree::GetPredictor(HostDeviceVector<float> const *out_pred,
   if (tparam_.predictor != PredictorType::kAuto) {
     if (tparam_.predictor == PredictorType::kGPUPredictor) {
 #if defined(XGBOOST_USE_CUDA)
+      CHECK_GE(common::AllVisibleGPUs(), 1) << "No visible GPU is found for XGBoost.";
       CHECK(gpu_predictor_);
       return gpu_predictor_;
 #else
@@ -429,6 +430,7 @@ GBTree::GetPredictor(HostDeviceVector<float> const *out_pred,
   // Use GPU Predictor if data is already on device and gpu_id is set.
   if (on_device && generic_param_->gpu_id >= 0) {
 #if defined(XGBOOST_USE_CUDA)
+    CHECK_GE(common::AllVisibleGPUs(), 1) << "No visible GPU is found for XGBoost.";
     CHECK(gpu_predictor_);
     return gpu_predictor_;
 #else
@@ -454,6 +456,7 @@ GBTree::GetPredictor(HostDeviceVector<float> const *out_pred,
 
   if (tparam_.tree_method == TreeMethod::kGPUHist) {
 #if defined(XGBOOST_USE_CUDA)
+    CHECK_GE(common::AllVisibleGPUs(), 1) << "No visible GPU is found for XGBoost.";
     CHECK(gpu_predictor_);
     return gpu_predictor_;
 #else

src/learner.cc

@@ -187,8 +187,13 @@ void GenericParameter::ConfigureGpuId(bool require_gpu) {
   // raw model), number of available GPUs could be different.  Wrap around it.
   int32_t n_gpus = common::AllVisibleGPUs();
   if (n_gpus == 0) {
+    if (gpu_id != kCpuId) {
+      LOG(WARNING) << "No visible GPU is found, setting `gpu_id` to -1";
+    }
     this->UpdateAllowUnknown(Args{{"gpu_id", std::to_string(kCpuId)}});
   } else if (gpu_id != kCpuId && gpu_id >= n_gpus) {
+    LOG(WARNING) << "Only " << n_gpus
+                 << " GPUs are visible, setting `gpu_id` to " << gpu_id % n_gpus;
     this->UpdateAllowUnknown(Args{{"gpu_id", std::to_string(gpu_id % n_gpus)}});
   }
 #else

tests/python-gpu/load_pickle.py

@@ -2,11 +2,17 @@
 `test_gpu_with_dask.py`'''
 import unittest
 import os
+import numpy as np
 import xgboost as xgb
 import json
+import pytest
+import sys
 
 from test_gpu_pickling import build_dataset, model_path, load_pickle
 
+sys.path.append("tests/python")
+import test_basic as tb
+
 
 class TestLoadPickle(unittest.TestCase):
     def test_load_pkl(self):
@@ -49,3 +55,15 @@ class TestLoadPickle(unittest.TestCase):
         test_x = xgb.DMatrix(x)
         res = bst.predict(test_x)
         assert len(res) == 10
+
+    def test_training_on_cpu_only_env(self):
+        assert os.environ['CUDA_VISIBLE_DEVICES'] == '-1'
+        rng = np.random.RandomState(1994)
+        X = rng.randn(10, 10)
+        y = rng.randn(10)
+        with tb.captured_output() as (out, err):
+            # Test no thrust exception is thrown
+            with pytest.raises(xgb.core.XGBoostError):
+                xgb.train({'tree_method': 'gpu_hist'}, xgb.DMatrix(X, y))
+
+            assert out.getvalue().find('No visible GPU is found') != -1

tests/python-gpu/test_gpu_pickling.py

@@ -157,3 +157,14 @@ class TestPickling(unittest.TestCase):
         bst.set_param({'predictor': 'cpu_predictor'})
         cpu_pred = model.predict(x, output_margin=True)
         np.testing.assert_allclose(cpu_pred, gpu_pred, rtol=1e-5)
+
+    def test_training_on_cpu_only_env(self):
+        cuda_environment = {'CUDA_VISIBLE_DEVICES': '-1'}
+        env = os.environ.copy()
+        env.update(cuda_environment)
+        args = self.args_template.copy()
+        args.append(
+            "./tests/python-gpu/"
+            "load_pickle.py::TestLoadPickle::test_training_on_cpu_only_env")
+        status = subprocess.call(args, env=env)
+        assert status == 0