Add test for invalid categorical data values. (#7380)

* Add test for invalid categorical data values.

* Add check during sketching.

src/common/categorical.h

@@ -42,9 +42,9 @@ inline XGBOOST_DEVICE bool Decision(common::Span<uint32_t const> cats, bst_cat_t
   return !s_cats.Check(cat);
 }
 
-inline void CheckCat(bst_cat_t cat) {
-  CHECK_GE(cat, 0) << "Invalid categorical value detected.  Categorical value "
-                      "should be non-negative.";
+inline void InvalidCategory() {
+  LOG(FATAL) << "Invalid categorical value detected.  Categorical value "
+                "should be non-negative.";
 }
 
 struct IsCatOp {

src/common/quantile.cu

@@ -580,6 +580,19 @@ void SketchContainer::AllReduce() {
   timer_.Stop(__func__);
 }
 
+namespace {
+struct InvalidCat {
+  Span<float const> values;
+  Span<uint32_t const> ptrs;
+  Span<FeatureType const> ft;
+
+  XGBOOST_DEVICE bool operator()(size_t i) {
+    auto fidx = dh::SegmentId(ptrs, i);
+    return IsCat(ft, fidx) && values[i] < 0;
+  }
+};
+}  // anonymous namespace
+
 void SketchContainer::MakeCuts(HistogramCuts* p_cuts) {
   timer_.Start(__func__);
   dh::safe_cuda(cudaSetDevice(device_));
@@ -669,6 +682,19 @@ void SketchContainer::MakeCuts(HistogramCuts* p_cuts) {
     assert(idx+1 < in_column.size());
     out_column[idx] = in_column[idx+1].value;
   });
+
+  if (has_categorical_) {
+    dh::XGBCachingDeviceAllocator<char> alloc;
+    auto ptrs = p_cuts->cut_ptrs_.ConstDeviceSpan();
+    auto it = thrust::make_counting_iterator(0ul);
+    CHECK_EQ(p_cuts->Ptrs().back(), out_cut_values.size());
+    auto invalid =
+        thrust::any_of(thrust::cuda::par(alloc), it, it + out_cut_values.size(),
+                       InvalidCat{out_cut_values, ptrs, d_ft});
+    if (invalid) {
+      InvalidCategory();
+    }
+  }
   timer_.Stop(__func__);
 }
 }  // namespace common

src/tree/updater_gpu_hist.cu

@@ -580,7 +580,9 @@ struct GPUHistMakerDevice {
       CHECK_LT(candidate.split.fvalue, std::numeric_limits<bst_cat_t>::max())
           << "Categorical feature value too large.";
       auto cat = common::AsCat(candidate.split.fvalue);
-      common::CheckCat(cat);
+      if (cat < 0) {
+        common::InvalidCategory();
+      }
       std::vector<uint32_t> split_cats(LBitField32::ComputeStorageSize(std::max(cat+1, 1)), 0);
       LBitField32 cats_bits(split_cats);
       cats_bits.Set(cat);

tests/python-gpu/test_gpu_updaters.py

@@ -95,6 +95,21 @@ class TestGPUUpdaters:
         rounds = 4
         self.run_categorical_basic(rows, cols, rounds, cats)
 
+    def test_invalid_categorical(self):
+        import cupy as cp
+        rng = np.random.default_rng()
+        X = rng.normal(loc=0, scale=1, size=1000).reshape(100, 10)
+        y = rng.normal(loc=0, scale=1, size=100)
+
+        # Check is performe during sketching.
+        Xy = xgb.DMatrix(X, y, feature_types=["c"] * 10)
+        with pytest.raises(ValueError):
+            xgb.train({"tree_method": "gpu_hist"}, Xy)
+
+        X, y = cp.array(X), cp.array(y)
+        with pytest.raises(ValueError):
+            Xy = xgb.DeviceQuantileDMatrix(X, y, feature_types=["c"] * 10)
+
     @pytest.mark.skipif(**tm.no_cupy())
     @given(parameter_strategy, strategies.integers(1, 20),
            tm.dataset_strategy)