Optional context for the host device vector.

[trivialfis]

This is the first PR to add an optional parameter to the HostDeviceVector and the Tensor class.
ref: #12122

[Copilot]

Pull request overview

This PR introduces an optional Context parameter to HostDeviceVector and linalg::Tensor APIs to enable future stream-aware CUDA behavior (custom streams / device-ordinal handling) and begins updating call sites to pass context through.

Changes:

• Extend HostDeviceVector and linalg::Tensor interfaces to accept an optional Context const*.
• Update GPU transfer/sync paths in HostDeviceVector to use a stream from CUDAContext when provided.
• Adjust predictor and CUDA unit tests to pass Context into the updated APIs.

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
include/xgboost/host_device_vector.h	Adds optional Context parameters to most HostDeviceVector APIs.
src/common/host_device_vector.cu	Wires stream selection for copies/fill and updates method signatures to accept context.
src/common/host_device_vector.cc	Updates non-CUDA fallback implementation to match new signatures.
include/xgboost/linalg.h	Threads optional context through Tensor/views and MakeTensorView helpers.
src/predictor/predictor.cc	Passes predictor context into HostDeviceVector operations for predictions init/fill/copy.
tests/cpp/common/test_host_device_vector.cu	Updates tests to construct and pass Context through updated APIs.
tests/cpp/common/test_linalg.h	Updates linear algebra dispatch helper to pass Context into views/host views.
tests/cpp/common/test_linalg.cu	Updates CUDA linalg tests to pass context into Tensor and View calls.

Comments suppressed due to low confidence (1)

src/common/host_device_vector.cu:43

• HostDeviceVectorImpl(size_t, T, DeviceOrd, CUDAContext const* ctx) takes ctx but doesn't use it. This can trigger -Wunused-parameter warnings (often treated as errors) and is confusing given the new stream-aware API. Consider removing the parameter name (or marking it [[maybe_unused]] / casting to void), or actually using it where appropriate.

  HostDeviceVectorImpl(size_t size, T v, DeviceOrd device, CUDAContext const* ctx)
      : device_(device) {
    if (device.IsCUDA()) {
      gpu_access_ = GPUAccess::kWrite;
      SetDevice();
      data_d_->resize(size, v);

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 33 out of 33 changed files in this pull request and generated 6 comments.

Comments suppressed due to low confidence (2)

src/common/device_compression.cu:372

• In CompressSnappy, DeviceUVector now requires a stream for mutations. p_out->resize(...) (and the earlier p_out->clear() in the same function) should take the CUDA context stream to avoid compilation issues and to ensure allocations happen on the intended stream for cudaMallocAsync-backed allocators.

  SafeNvComp(nvcompBatchedSnappyCompressGetMaxOutputChunkSize(
      std::min(max_in_nbytes, chunk_size), nvcomp_batched_snappy_opts, &max_out_nbytes));
  p_out->resize(max_out_nbytes * n_chunks);
  std::vector<void*> h_out_ptrs(n_chunks);

src/data/ellpack_page_source.cu:235

• d_page is allocated using ctx.CUDACtx()->Stream(), but the subsequent cudaMemcpyAsync to/from that buffer omits the stream parameter (defaults to legacy stream). This can introduce cross-stream races when a custom stream is used; pass ctx.CUDACtx()->Stream() to the memcpy (and keep stream usage consistent for the earlier host-cache memcpy as well).

      auto d_page = common::MakeFixedVecWithCudaMalloc<common::CompressedByteT>(&ctx, remaining);
      if (remaining > 0) {
        dh::safe_cuda(cudaMemcpyAsync(d_page.data(), old_impl->gidx_buffer.data() + n_bytes,
                                      remaining, cudaMemcpyDefault));
      }

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.