Enable C++ PPR Sampling

[mkolodner-sc]

Scope of work done

PPR-Based Neighbor Sampling for GiGL

Adds C++ based PPR (Personalized PageRank) based neighbor sampling as an alternative to k-hop sampling in GiGL's distributed training stack, switching from the previous pythonic implementation.

C++ Kernel (gigl-core/csrc/sampling/)

• Implements the Forward Push algorithm (Andersen et al., 2006) as PPRForwardPushState — a pybind11-wrapped C++ class that owns all hot-loop state (scores, residuals, queue, neighbor cache).
• Python drives the async RPC neighbor fetches; C++ handles the residual push loop, keeping the asyncio event loop unblocked via run_in_executor.
• gigl-core is a separately installable package (uv pip install -e gigl-core/) with a py.typed marker and .pyi stub for full mypy coverage.

Python Sampler (gigl/distributed/dist_ppr_sampler.py)

• DistPPRNeighborSampler extends BaseGiGLSampler and integrates with the existing DistNeighborLoader / Graph Store infrastructure.
• Output format: (seed_type, "ppr", neighbor_type) edge types with edge_index ([2, N] int64) and edge_attr ([N] float PPR scores).
• max_fetch_iterations: Optional[int] = None caps RPC calls per batch; the algorithm continues to convergence using cached neighbor lists after the budget is exhausted.
• num_neighbors_per_hop defaults to 1000 — high-degree hub nodes receive diminishing residual per neighbor, so capping the fetch has negligible effect on PPR accuracy.

Configuration

• PPRSamplerOptions added to sampler_options.py alongside existing KHopNeighborSamplerOptions; pass via sampler_options= on DistNeighborLoader.

C++ Tests (gigl-core/tests/ppr_forward_push_test.cpp)

• 6 GoogleTest cases covering: initial queue drain, convergence, score absorption, residual distribution, cross-seed deduplication, and top-k limiting.
• tests/CMakeLists.txt updated to auto-link torch and auto-discover kernel sources under csrc/ (excluding python_* entry points).

Where is the documentation for this feature?: N/A

Did you add automated tests or write a test plan?

Updated Changelog.md? NO

Ready for code review?: NO

[kmontemayor2-sc]

Thanks Matt! just nits now the only thing that's a bit odd is why we can't do std::optional<std::vector>

[kmontemayor2-sc]

Thanks Matt! just nits now the only thing that's a bit odd is why we can't do std::optional<std::vector>

[yliu2-sc]

Second stamp

[github-actions]

GiGL Automation

@ 22:20:44UTC : Starting to build base images for CUDA and CPU.
This may take a while, please be patient.
The images will be pushed to the GCP Artifact Registry.
Once done, the workflow will update the gigl/dep_vars.env file with the new image names.

[github-actions]

GiGL Automation

@ 22:48:36UTC : Built and pushed new images:

• CUDA base image: us-central1-docker.pkg.dev/external-snap-ci-github-gigl/public-gigl/gigl-cuda-base:7d3182eeb6446ce3e35910babba990c8e003879d.109.1
• CPU base image: us-central1-docker.pkg.dev/external-snap-ci-github-gigl/public-gigl/gigl-cpu-base:7d3182eeb6446ce3e35910babba990c8e003879d.109.1
• Dataflow base image: us-central1-docker.pkg.dev/external-snap-ci-github-gigl/public-gigl/gigl-dataflow-base:7d3182eeb6446ce3e35910babba990c8e003879d.109.1
• Builder image: us-central1-docker.pkg.dev/external-snap-ci-github-gigl/gigl-base-images/gigl-builder:7d3182eeb6446ce3e35910babba990c8e003879d.109.1

Updated gigl/dep_vars.env with new image names.
Updated .github/cloud_builder/run_command_on_pr_cloud_build.yaml with new builder image.