debug.

Author: trivialfis

demo/guide-python/external_memory.py

@@ -75,7 +75,7 @@ def reset(self) -> None:
 
 def main(tmpdir: str) -> xgboost.Booster:
     # generate some random data for demo
-    files = make_batches(1024, 17, 31, tmpdir)
+    files = make_batches(2 ** 16, 17, 31, tmpdir)
     it = Iterator(files)
     # For non-data arguments, specify it here once instead of passing them by the `next`
     # method.
@@ -85,12 +85,11 @@ def main(tmpdir: str) -> xgboost.Booster:
     # Other tree methods including ``hist`` and ``gpu_hist`` also work, see tutorial in
     # doc for details.
     booster = xgboost.train(
-        {"tree_method": "approx", "max_depth": 2},
+        {"tree_method": "gpu_hist", "max_depth": 6, "sampling_method": "gradient_based", "subsample": 0.5},
         Xy,
         evals=[(Xy, "Train")],
-        num_boost_round=10,
+        num_boost_round=2,
     )
-    return booster
 
 
 if __name__ == "__main__":

src/data/ellpack_page_raw_format.cu

@@ -28,6 +28,7 @@ class EllpackPageRawFormat : public SparsePageFormat<EllpackPage> {
     if (!fi->Read(&impl->base_rowid)) {
       return false;
     }
+    std::cout << "impl brd:" << impl->base_rowid << std::endl;
     return true;
   }
 
@@ -41,6 +42,7 @@ class EllpackPageRawFormat : public SparsePageFormat<EllpackPage> {
     bytes += sizeof(impl->is_dense);
     fo->Write(impl->row_stride);
     bytes += sizeof(impl->row_stride);
+    std::cout << "write brd:" << impl->base_rowid << std::endl;
     CHECK(!impl->gidx_buffer.ConstHostVector().empty());
     fo->Write(impl->gidx_buffer.HostVector());
     bytes += impl->gidx_buffer.ConstHostSpan().size_bytes() + sizeof(uint64_t);

src/data/ellpack_page_source.cu

@@ -24,6 +24,7 @@ void EllpackPageSource::Fetch() {
     auto *impl = this->page_->Impl();
     *impl = EllpackPageImpl(device_, *cuts_, *csr, is_dense_, row_stride_, feature_types_);
     page_->SetBaseRowId(csr->base_rowid);
+    std::cout << "csr br:" << csr->base_rowid << std::endl;
     this->WriteCache();
   }
 }

src/data/sparse_page_source.h

@@ -117,7 +117,7 @@ class SparsePageSourceImpl : public BatchIteratorImpl<S> {
     }
     // An heuristic for number of pre-fetched batches.  We can make it part of BatchParam
     // to let user adjust number of pre-fetched batches when needed.
-    uint32_t constexpr kPreFetch = 3;
+    uint32_t constexpr kPreFetch = 1;
 
     size_t n_prefetch_batches = std::min(kPreFetch, n_batches_);
     CHECK_GT(n_prefetch_batches, 0) << "total batches:" << n_batches_;
@@ -140,6 +140,7 @@ class SparsePageSourceImpl : public BatchIteratorImpl<S> {
 
         std::uint64_t offset = self->cache_info_->offset.at(fetch_it);
         std::uint64_t length = self->cache_info_->bytes.at(fetch_it);
+        // std::cout << typeid(S).name() << " offset:" << offset << " length:" << length << std::endl;
 
         auto fi = std::make_unique<common::PrivateMmapStream>(n, true, offset, length);
         CHECK(fmt->Read(page.get(), fi.get()));