MAINT: Bump legate dependencies to 25.08 (#236)

seberg · RAMitchell · web-flow · commit 859433072160 · 2025-08-12T17:41:48.000+02:00
* MAINT: Bump legate dependencies to 25.08

Signed-off-by: Sebastian Berg &lt;sebastianb@nvidia.com&gt;

* Transition from tcb/span to cuda/std/span

Signed-off-by: Sebastian Berg &lt;sebastianb@nvidia.com&gt;

* Fix clang-tidy

Signed-off-by: Sebastian Berg &lt;sebastianb@nvidia.com&gt;

* Fix exp function

* More fixes

* One more 404

---------

Signed-off-by: Sebastian Berg &lt;sebastianb@nvidia.com&gt;
Co-authored-by: Rory Mitchell &lt;r.a.mitchell.nz@gmail.com&gt;
diff --git a/conda/environments/all_cuda-122.yaml b/conda/environments/all_cuda-122.yaml
@@ -15,9 +15,9 @@ dependencies:
 - cuda-nvcc
 - cuda-toolkit
 - cuda-version>=12.2
-- cupynumeric==25.07.*,>=0.0.0.dev0
+- cupynumeric==25.08.*,>=0.0.0.dev0
 - hypothesis>=6
-- legate==25.07.*,>=0.0.0.dev0
+- legate==25.08.*,>=0.0.0.dev0
 - libcublas-dev
 - llvm-openmp
 - make
diff --git a/conda/recipes/legate-boost/conda_build_config.yaml b/conda/recipes/legate-boost/conda_build_config.yaml
@@ -26,6 +26,6 @@ gpu_enabled:
   - false
 
 cupynumeric_version:
-  - "=25.07.*,>=0.0.0.dev0"
+  - "=25.08.*,>=0.0.0.dev0"
 legate_version:
-  - "=25.07.*,>=0.0.0.dev0"
+  - "=25.08.*,>=0.0.0.dev0"
diff --git a/dependencies.yaml b/dependencies.yaml
@@ -81,7 +81,7 @@ dependencies:
       - output_types: [conda, pyproject, requirements]
         packages:
           - cmake>=3.26.4,!=3.30.0
-          - &legate legate==25.07.*,>=0.0.0.dev0
+          - &legate legate==25.08.*,>=0.0.0.dev0
           - ninja>=1.11.1.1
   clang_tidy:
     common:
@@ -160,7 +160,7 @@ dependencies:
           - typing-extensions>=4.0
       - output_types: [conda, pyproject, requirements]
         packages:
-          - cupynumeric==25.07.*,>=0.0.0.dev0
+          - cupynumeric==25.08.*,>=0.0.0.dev0
           - *legate
   test:
     common:
diff --git a/legateboost/metrics.py b/legateboost/metrics.py
@@ -343,8 +343,8 @@ def metric(self, y: cn.ndarray, pred: cn.ndarray, w: cn.ndarray) -> cn.ndarray:
         f = cn.log(pred) * (K - self.one)  # undo softmax
         y_k = cn.full((y.size, K), -self.one / (K - self.one))
 
-        set_col_by_idx(y_k, y.astype(cn.int32), self.one)
-        # y_k[cn.arange(y.size), y.astype(cn.int32)] = 1.0
+        # y_k[cn.arange(y.size), y.astype(cn.int32)] = self.one
+        y_k = set_col_by_idx(y_k, y.astype(cn.int32), self.one)
 
         exp = cn.exp(-1 / K * cn.sum(y_k * f, axis=1))
         return (exp * w).sum() / w.sum()
diff --git a/legateboost/objectives.py b/legateboost/objectives.py
@@ -598,7 +598,7 @@ def gradient(self, y: cn.ndarray, pred: cn.ndarray) -> GradPair:
         f = cn.log(pred) * (K - 1)  # undo softmax
         y_k = cn.full((y.size, K), -self.one / (K - self.one))
         labels = y.astype(cn.int32).squeeze()
-        set_col_by_idx(y_k, labels, self.one)
+        y_k = set_col_by_idx(y_k, labels, self.one)
         # y_k[cn.arange(y.size), labels] = 1.0
         exp = cn.exp(-1 / K * cn.sum(y_k * f, axis=1))
 
diff --git a/legateboost/test/models/test_tree.py b/legateboost/test/models/test_tree.py
@@ -82,7 +82,9 @@ def get_feature_distribution(model):
     histogram = cn.zeros(model.n_features_in_)
     for m in model:
         histogram += cn.histogram(
-            m.feature, bins=model.n_features_in_, range=(0, model.n_features_in_)
+            m.feature.astype(cn.int64),
+            bins=model.n_features_in_,
+            range=(0, model.n_features_in_),
         )[0]
     return histogram / histogram.sum()
 
diff --git a/legateboost/test/test_examples.py b/legateboost/test/test_examples.py
@@ -4,6 +4,7 @@
 import subprocess
 import sys
 from pathlib import Path
+from urllib.error import HTTPError
 
 import pytest
 
@@ -29,7 +30,12 @@ def test_examples(path):
     os.environ["CI"] = "1"
     rel = path.relative_to(example_dir).with_suffix("")
     rel = str(rel).replace("/", ".")
-    importlib.import_module(rel)
+    # temporary openml issue, this try/catch could be removed once fixed
+    # https://github.com/openml/OpenML/issues/1232
+    try:
+        importlib.import_module(rel)
+    except HTTPError as e:
+        print(f"HTTP error in {path}: {e}")
 
 
 notebooks = list(noteboook_dir.glob("*.ipynb"))
@@ -55,8 +61,13 @@ def test_notebooks(path):
     except subprocess.CalledProcessError as e:
         print(e.stderr.decode())
         raise e
-    # import the script to run it in the existing python process
-    importlib.import_module(path.stem)
+
+    # temporary openml issue, this try/catch could be removed once fixed
+    # https://github.com/openml/OpenML/issues/1232
+    try:
+        importlib.import_module(path.stem)
+    except HTTPError as e:
+        print(f"HTTP error in {path}: {e}")
 
 
 def test_benchmark(benchmark_dir):
diff --git a/legateboost/utils.py b/legateboost/utils.py
@@ -105,7 +105,7 @@ def pick_col_by_idx(a: cn.ndarray, b: cn.ndarray) -> cn.ndarray:
     return result.sum(axis=1)
 
 
-def set_col_by_idx(a: cn.ndarray, b: cn.ndarray, delta: float) -> None:
+def set_col_by_idx(a: cn.ndarray, b: cn.ndarray, delta: float) -> cn.ndarray:
     """Alternative implementation for a[cn.arange(b.size), b] = delta."""
 
     assert a.ndim == 2
@@ -116,7 +116,7 @@ def set_col_by_idx(a: cn.ndarray, b: cn.ndarray, delta: float) -> None:
     bools = b[:, cn.newaxis] == range[cn.newaxis, :]
     a -= a * bools
     a += delta * bools
-    return
+    return a
 
 
 def mod_col_by_idx(a: cn.ndarray, b: cn.ndarray, delta: float) -> None:
diff --git a/pyproject.toml b/pyproject.toml
@@ -12,7 +12,7 @@ commit-files = ["legateboost/GIT_COMMIT"]
 matrix-entry = "cuda_suffixed=true"
 requires = [
     "cmake>=3.26.4,!=3.30.0",
-    "legate==25.07.*,>=0.0.0.dev0",
+    "legate==25.08.*,>=0.0.0.dev0",
     "ninja>=1.11.1.1",
 ] # This list was generated by `rapids-dependency-file-generator`. To make changes, edit dependencies.yaml and run `rapids-dependency-file-generator`.
 
@@ -31,8 +31,8 @@ classifiers = [
     "Programming Language :: Python :: 3.12",
 ]
 dependencies = [
-    "cupynumeric==25.07.*,>=0.0.0.dev0",
-    "legate==25.07.*,>=0.0.0.dev0",
+    "cupynumeric==25.08.*,>=0.0.0.dev0",
+    "legate==25.08.*,>=0.0.0.dev0",
     "numpy<2",
     "scikit-learn>=1.6",
     "typing-extensions>=4.0",
diff --git a/src/cpp_utils/cpp_utils.cuh b/src/cpp_utils/cpp_utils.cuh
@@ -16,17 +16,20 @@
 
 #pragma once
 
+#include "cpp_utils.h"
+
 #include <nccl.h>
 #include <thrust/execution_policy.h>
 #include <thrust/system_error.h>
-#include <cstddef>
+
+#include <cuda/std/span>
 #include <cstdio>
-#include <utility>
+#include <cstddef>
 #include <string>
-#include <tcb/span.hpp>
+#include <utility>
+
 #include "legate.h"
 #include "legate/cuda/cuda.h"
-#include "cpp_utils.h"
 
 namespace legateboost {
 
@@ -96,7 +99,10 @@ inline void LaunchN(size_t n, cudaStream_t stream, const L& lambda)
 }
 
 template <typename T>
-void AllReduce(legate::TaskContext context, tcb::span<T> x, ncclRedOp_t op, cudaStream_t stream)
+void AllReduce(legate::TaskContext context,
+               cuda::std::span<T> x,
+               ncclRedOp_t op,
+               cudaStream_t stream)
 {
   const auto& domain     = context.get_launch_domain();
   size_t const num_ranks = domain.get_volume();
@@ -123,7 +129,7 @@ void AllReduce(legate::TaskContext context, tcb::span<T> x, ncclRedOp_t op, cuda
 }
 
 template <typename T>
-void SumAllReduce(legate::TaskContext context, tcb::span<T> x, cudaStream_t stream)
+void SumAllReduce(legate::TaskContext context, cuda::std::span<T> x, cudaStream_t stream)
 {
   AllReduce(context, x, ncclSum, stream);
 }
diff --git a/src/cpp_utils/cpp_utils.h b/src/cpp_utils/cpp_utils.h
@@ -18,6 +18,7 @@
 #include <thrust/for_each.h>
 #include <thrust/iterator/counting_iterator.h>
 #include <thrust/execution_policy.h>
+#include <cuda/std/span>
 #include <assert.h>
 #include <cstdio>
 #include <string>
@@ -28,7 +29,6 @@
 #include <tuple>
 #include "legate_library.h"
 #include "legate/comm/coll.h"
-#include <tcb/span.hpp>
 
 namespace legateboost {
 
@@ -182,7 +182,7 @@ constexpr auto type_dispatch_float(legate::Type::Code code, Functor&& f, Fnargs&
 }
 
 template <typename T, typename OpT>
-void AllReduce(legate::TaskContext context, tcb::span<T> x, OpT op)
+void AllReduce(legate::TaskContext context, cuda::std::span<T> x, OpT op)
 {
   const auto& domain     = context.get_launch_domain();
   const size_t num_ranks = domain.get_volume();
@@ -222,7 +222,7 @@ void AllReduce(legate::TaskContext context, tcb::span<T> x, OpT op)
 }
 
 template <typename T>
-void SumAllReduce(legate::TaskContext context, tcb::span<T> x)
+void SumAllReduce(legate::TaskContext context, cuda::std::span<T> x)
 {
   AllReduce(context, x, std::plus<T>());
 }
diff --git a/src/encoder/target_encoder.cu b/src/encoder/target_encoder.cu
@@ -15,23 +15,23 @@
  */
 #include "target_encoder.h"
 #include <thrust/find.h>
+#include <cuda/std/span>
 #include <vector>
 #include <unordered_map>
 #include <random>
-#include <tcb/span.hpp>
 #include "../cpp_utils/cpp_utils.cuh"
 
 namespace legateboost {
 
 template <typename T>
 struct CategoriesMap {
-  tcb::span<const T> categories;
-  tcb::span<const int64_t> row_pointers;
+  cuda::std::span<const T> categories;
+  cuda::std::span<const int64_t> row_pointers;
   __device__ int64_t GetIndex(T x, int64_t feature_idx) const
   {
-    const auto* begin  = categories.begin() + row_pointers[feature_idx];
-    const auto* end    = categories.begin() + row_pointers[feature_idx + 1];
-    const auto* result = thrust::find(thrust::seq, begin, end, x);
+    const auto begin  = categories.begin() + row_pointers[feature_idx];
+    const auto end    = categories.begin() + row_pointers[feature_idx + 1];
+    const auto result = thrust::find(thrust::seq, begin, end, x);
     if (result == end) { return -1; }
     return result - categories.begin();
   }
diff --git a/src/models/nn/build_nn.cc b/src/models/nn/build_nn.cc
@@ -231,7 +231,7 @@ auto eval_cost(NNContext* context,
 
   sum /= total_rows * pred.extent[1];
 
-  SumAllReduce(context->legate_context, tcb::span<double>(&sum, 1));
+  SumAllReduce(context->legate_context, cuda::std::span<double>(&sum, 1));
 
   if (alpha > 0.0) {
     T L2 = 0.0;
diff --git a/src/models/nn/build_nn.cu b/src/models/nn/build_nn.cu
@@ -309,7 +309,7 @@ auto eval_cost(NNContext* context,
                          result.ptr(0),
                          cost_array.size(),
                          context->stream);
-  SumAllReduce(context->legate_context, tcb::span<T>(result.ptr(0), 1), context->stream);
+  SumAllReduce(context->legate_context, cuda::std::span<T>(result.ptr(0), 1), context->stream);
 
   T cost;
   cudaMemcpyAsync(&cost, result.ptr(0), sizeof(T), cudaMemcpyDeviceToHost, context->stream);
@@ -401,7 +401,7 @@ auto backward(NNContext* nn_context,
 
   // Scale and allreduce gradients
   SumAllReduce(nn_context->legate_context, grads.data, nn_context->stream);
-  // Cannot use tcb::span in kernel
+  // TODO(seberg): Could previously not use span in kernel
   LaunchN(grads.size(), nn_context->stream, [=] __device__(int64_t idx) {
     grads.data[idx] /= total_rows;
   });
diff --git a/src/models/nn/build_nn.h b/src/models/nn/build_nn.h
@@ -14,10 +14,10 @@
  *
  */
 #pragma once
+#include <cuda/std/span>
 #include <algorithm>
 #include <memory>
 #include <limits>
-#include <tcb/span.hpp>
 #include "../../cpp_utils/cpp_utils.h"
 #include "legate_library.h"
 #include "legateboost.h"
@@ -27,10 +27,10 @@ namespace legateboost {
 template <typename T>
 struct Matrix {
   std::array<int64_t, 2> extent{};
-  tcb::span<T> data;
+  cuda::std::span<T> data;
   std::shared_ptr<legate::Buffer<T, 2>> buffer;
 
-  Matrix(tcb::span<T> data, std::array<int64_t, 2> extent)
+  Matrix(cuda::std::span<T> data, std::array<int64_t, 2> extent)
     : extent({std::max(extent[0], 0L), std::max(extent[1], 0L)}), data(data)
   {
     if (extent[0] * extent[1] != data.size()) {
@@ -58,20 +58,22 @@ struct Matrix {
   static auto From1dOutputStore(const legate::PhysicalStore& store) -> Matrix<T>
   {
     auto shape = store.shape<1>();
-    tcb::span<T> const data(store.read_write_accessor<T, 1, true>().ptr(shape.lo), shape.volume());
+    cuda::std::span<T> const data(store.read_write_accessor<T, 1, true>().ptr(shape.lo),
+                                  shape.volume());
     return Matrix<T>(data, {shape.hi[0] - shape.lo[0] + 1, 1});
   }
 
   static auto From2dStore(legate::PhysicalStore store) -> Matrix<T>
   {
     auto shape = store.shape<2>();
-    tcb::span<T> data(store.read_accessor<T, 2, true>().ptr(shape.lo), shape.volume());
+    cuda::std::span<T> data(store.read_accessor<T, 2, true>().ptr(shape.lo), shape.volume());
     return Matrix<T>(data, {shape.hi[0] - shape.lo[0] + 1, shape.hi[1] - shape.lo[1] + 1});
   }
   static auto From2dOutputStore(const legate::PhysicalStore& store) -> Matrix<T>
   {
     auto shape = store.shape<2>();
-    tcb::span<T> const data(store.read_write_accessor<T, 2, true>().ptr(shape.lo), shape.volume());
+    cuda::std::span<T> const data(store.read_write_accessor<T, 2, true>().ptr(shape.lo),
+                                  shape.volume());
     return Matrix<T>(data, {shape.hi[0] - shape.lo[0] + 1, shape.hi[1] - shape.lo[1] + 1});
   }
 
diff --git a/src/models/tree/build_tree.cc b/src/models/tree/build_tree.cc
diff --git a/src/models/tree/build_tree.cu b/src/models/tree/build_tree.cu
diff --git a/src/models/tree/build_tree.h b/src/models/tree/build_tree.h
diff --git a/src/models/tree/update_tree.cc b/src/models/tree/update_tree.cc
diff --git a/src/utils/gather.cc b/src/utils/gather.cc
diff --git a/src/utils/gather.cu b/src/utils/gather.cu
