Remove reintroduced GPU sketch memory estimation

[RAMitchell]

Summary

This removes the GPU sketch memory-estimation path that was accidentally reintroduced after #12118.

The GPU sketch batching policy goes back to the fixed 64M sliding-window size:

• no RequiredMemory(...)
• no SketchBatchNumElements(...)
• no batch-size override plumbing on DeviceSketch* / AdapterDeviceSketch

What Changed

• remove the estimator helpers and related API surface from src/common/hist_util.cuh and src/common/hist_util.cu
• restore fixed-window batching with detail::kSketchBatchNumElements
• remove the estimator-specific batch override test from tests/cpp/common/test_hist_util.cu
• make the fixed-window path a bit more robust:
  • DeviceSketchWithHessian(...) no longer depends on info.num_nonzero_ for loop progress
  • AdapterDeviceSketch(...) sizes the last window from the true remaining elements

Testing

• configured and rebuilt testxgboost with CUDA
• ran the focused GPU sketch / adapter / external-memory / hessian slice

Result:

• 66/66 passed

[Copilot]

Pull request overview

Removes the accidentally reintroduced GPU sketch memory-estimation code path and restores a fixed 64M-element sliding-window batching policy for GPU sketching.

Changes:

• Deleted memory-estimation helpers/APIs (RequiredMemory, SketchBatchNumElements, batch-size override plumbing) and reverted to a fixed detail::kSketchBatchNumElements window size.
• Updated GPU sketching implementations to use fixed-window batching (including more robust last-window sizing for adapter sketching).
• Removed the estimator-specific batch override test and updated hessian-sketch tests to match the simplified API.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated no comments.

File	Description
src/common/hist_util.cuh	Removes estimator-related declarations/API surface; introduces fixed batch constant; simplifies DeviceSketch*/AdapterDeviceSketch interfaces.
src/common/hist_util.cu	Deletes estimator implementations and uses fixed batch size in DeviceSketchWithHessian batching loop.
tests/cpp/common/test_hist_util.cu	Removes batch-override test coverage and updates call sites to the new fixed-batch API.

Comments suppressed due to low confidence (1)

src/common/hist_util.cuh:369

• The comment mentions an "estimation of sketch_batch_num_elements" assuming dense input, but sketch_batch_num_elements is now a fixed constant (clamped to remaining). Update the wording to refer to the dense assumption used when computing approx_n_samples, or remove the outdated phrasing to avoid confusion.

    // Re-estimate the needed number of cuts based on the size of the sub-batch.
    //
    // The estimation of `sketch_batch_num_elements` assumes dense input, so the
    // approximation here is reasonably accurate. It doesn't hurt accuracy since the
    // estimated n_samples must be greater or equal to the actual n_samples thanks to the

---

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