[jcrist]

By default cuml.accel enables a ManagedMemoryResource, which allocates managed memory. While benchmarking some workloads, I've noticed a measurable performance improvement when wrapping this mr with a PrefetchResourceAdaptor. This adaptor schedules a prefetch of memory to the device immediately after allocation. For cuml/cuml.accel workloads this makes sense - we typically allocate an array, then call a kernel to fill it in, before returning it. This also matches some performance advice noted by other teams (and the official CUDA programming guide) that prefetching may be useful when working with managed memory.

This also more closely matches the rmm configuration used by cudf.pandas by default (see here). Note that cudf.pandas also uses a PoolMemoryResource by default, but always enables prefetching of managed memory even without the pool. My intuition is that pooling makes sense when you have many smaller allocations (like might happen with pandas), since it can amortize the upstream allocator cost across calls. cuml workloads typically do fewer and larger allocations, so I don't think investigating pooling is necessary for now.

Fixes #7846.

[coderabbitai]

📝 Walkthrough

Summary by CodeRabbit

• Refactor

  • Improved GPU memory resource handling during acceleration initialization: more robust detection of existing memory setups to avoid redundant changes, broader recognition of pre-wrapped resources, and a more reliable path to enable pooled + prefetching managed memory for better stability and performance.

• Tests

  • Added an integration test ensuring SVR handles a zero-support scenario correctly and produces expected constant-target predictions.

Walkthrough

Refactors install() memory-resource handling to detect and treat a PrefetchResourceAdaptor wrapping PoolMemoryResource(ManagedMemoryResource()) as already enabled, skip enabling for non-CUDA resources, and when enabling use PrefetchResourceAdaptor(PoolMemoryResource(ManagedMemoryResource())). Adds a new SVR unit test for zero supports.

Changes

Cohort / File(s)	Summary
Memory Resource Handling python/cuml/cuml/accel/core.py	Adjusts install() checks to recognize PrefetchResourceAdaptor(PoolMemoryResource(ManagedMemoryResource())) as already enabled, skips enabling for non-CudaMemoryResource, and sets resource to PrefetchResourceAdaptor(PoolMemoryResource(ManagedMemoryResource())) when enabling.
SVR Tests python/cuml/cuml_accel_tests/integration/test_svr.py	Adds test_svr_fit_n_support_0 verifying SVR behavior with an identity input and zero supports (checks intercept_, n_support_, dual_coef_.shape, and predictions).

Estimated code review effort

🎯 3 (Moderate) | ⏱️ ~20 minutes

Possibly related issues

• Use PrefetchResourceAdaptor with managed memory in cuml.accel #7846 — Implements/detects PrefetchResourceAdaptor-wrapped managed memory (PrefetchResourceAdaptor(PoolMemoryResource(ManagedMemoryResource()))) for cuml.accel; aligns with the memory-resource detection changes in this PR.

🚥 Pre-merge checks | ✅ 2 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 66.67% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (2 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly summarizes the main change: introducing PrefetchResourceAdaptor in cuml.accel for managed memory resource wrapping.
Description check	✅ Passed	The description is directly related to the changeset, explaining the motivation for wrapping ManagedMemoryResource with PrefetchResourceAdaptor and its performance benefits.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Post copyable unit tests in a comment

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[jcrist]

This is blocked by #7842.

[coderabbitai]

Actionable comments posted: 1

🤖 Prompt for all review comments with AI agents

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@python/cuml/cuml_accel_tests/integration/test_svr.py`:
- Around line 82-89: In test_svr_fit_n_support_0 update the assertions to use
tolerance-based float comparisons (e.g., numpy allclose or pytest.approx) for
model.intercept_ and model.predict(X) instead of direct equality, avoid calling
.all() on model.n_support_ (treat it as an int or use np.array(...) then compare
to 0) and keep the shape check for model.dual_coef_ as-is; specifically, replace
exact equality checks on model.intercept_, model.predict(X) with approximate
checks and change the n_support_ assertion to handle scalar or array forms
reliably.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: 15b7cf76-3e75-4893-bf35-5fe94263a11c

📥 Commits

Reviewing files that changed from the base of the PR and between a9c6b31 and 86aeff3.

📒 Files selected for processing (2)

• python/cuml/cuml/accel/core.py
• python/cuml/cuml_accel_tests/integration/test_svr.py

[jcrist]

/merge