Use context for CUDA external memory DMatrix.

[trivialfis]

The change by #12121 is too significant, starting with a smaller place.

• Use a sync policy to let the caller handle stream sync. This avoids flaky lifetime management in the dtor.

[Copilot]

Pull request overview

This PR threads Context through CUDA external-memory ELLPACK/IterativeDMatrix paths so page creation, cache IO, and raw format read/write can use the caller-provided CUDA context (instead of constructing ad-hoc contexts / default-stream assumptions).

Changes:

• Add Context const* plumbing to IterativeDMatrix::Load, EllpackPage construction, EllpackPageRawFormat, and host-cache write paths.
• Introduce MakePage() hooks in format policies to centralize page allocation (supporting context-aware page types).
• Update CUDA tests to use the new context-aware constructors/signatures.

Reviewed changes

Copilot reviewed 16 out of 16 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
tests/cpp/data/test_iterative_dmatrix.cu	Update IterativeDMatrix::Load call to pass Context.
tests/cpp/data/test_ellpack_page_raw_format.cu	Update ELLPACK raw-format tests for context-aware SetCuts and EllpackPage construction.
src/data/sparse_page_source.h	Add MakePage() to format policy and use it during prefetch page allocation.
src/data/iterative_dmatrix.h	Change IterativeDMatrix::Load signature to accept Context const*.
src/data/iterative_dmatrix.cu	Pass context into EllpackPage construction and ELLPACK raw format read/write + Load.
src/data/iterative_dmatrix.cc	Update non-CUDA stub Load signature to match header.
src/data/gradient_index_page_source.h	Add MakePage() for gradient index format policy.
src/data/extmem_quantile_dmatrix.cc	Initialize and use fmat_ctx_ consistently for CPU/GPU init and category sync.
src/data/ellpack_page.h	Make EllpackPage require an explicit Context const* for empty-page construction.
src/data/ellpack_page.cuh	Make EllpackPageImpl require context and store a Context const*.
src/data/ellpack_page.cu	Wire context into EllpackPageImpl construction and use context stream in destructor sync.
src/data/ellpack_page.cc	Update non-CUDA dummy EllpackPage ctor signature.
src/data/ellpack_page_source.h	Store context in EllpackFormatPolicy, pass it into EllpackPageRawFormat, and create pages with it.
src/data/ellpack_page_source.cu	Require context for host-cache write and for constructing pages/impls during fetch/caching.
src/data/ellpack_page_raw_format.h	Add Context const* member and update ctor signature.
src/data/ellpack_page_raw_format.cu	Use provided context/stream for device copies, cache IO, and sync.

---

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

[Copilot]

EllpackPageImpl now stores a raw Context const* (used by its destructor for stream sync). In this code path, the else branch below constructs a temporary CUDA context via ctx->MakeCUDA() and passes its address into EllpackPageImpl, which will leave impl->ctx_ dangling once the branch exits and can trigger use-after-free later (e.g. in ~EllpackPageImpl()). Ensure the context pointer stored inside EllpackPageImpl always outlives the page (e.g. persist the CUDA context in a longer-lived owner like fmat_ctx_, or avoid storing a raw Context* in EllpackPageImpl and store a device/stream handle instead).

[Copilot]

ReadDeviceVec now takes a Context const* and immediately uses ctx->CUDACtx()->Stream() for cudaMemcpyAsync, but it never validates that ctx is non-null / CUDA, nor does it ensure the current CUDA device matches ctx's ordinal. This can lead to invalid stream/device usage when the caller's current device differs from ctx. Add a CHECK(ctx && ctx->IsCUDA()) and set the device (e.g. curt::SetDevice(ctx->Ordinal())) before allocating/copying.

[Copilot]

EllpackPageRawFormat::Read uses ctx_ (stored in the format object) for device copies and stream sync, but there’s no precondition check that ctx_ is non-null / CUDA or that the active device matches ctx_. Since the constructor now accepts an arbitrary context pointer, add validation (and potentially a device set) at the start of Read/Write to prevent null deref or cross-device stream usage.

[RAMitchell]

I find it a bit weird passing a raw pointer all the time without any guardrails for its state, but I guess we strongly guarantee that there is only ever one context and it always exists.

[trivialfis]

I find it a bit weird passing a raw pointer all the time without any guardrails for its state

That's indeed a significant issue. I would like to use a shared ptr instead at some point.