Optimize GFQL GPU hot paths and RAPIDS benchmarks

[lmeyerov]

Summary

This PR advances the GFQL/RAPIDS GPU regression investigation in three ways:

• optimizes the hot undirected single-hop GFQL traversal path in hop()
• adds a narrow safe to_cugraph() fast path for contiguous zero-based integer IDs
• adds official RAPIDS-image GPU benchmark tooling for direct to_cugraph(), GFQL pipeline, GFQL stage-breakdown, and timing-summary comparisons across RAPIDS/CUDA variants
• tracks the upstream-facing RAPIDS string-ID graph-build regression in #977

Kept code changes

• graphistry/compute/hop.py
  • narrow undirected single-hop fast path to avoid doubled pair materialization
• graphistry/plugins/cugraph.py
  • guarded auto-renumber=False when IDs are already signed integer, zero-based, and contiguous
  • cached fast-path decision reuse on repeated conversions
  • conservative rejection of null-bearing integer endpoint columns for the renumber=False fast path
• new/expanded tests
  • graphistry/tests/compute/test_hop.py
  • graphistry/tests/plugins/test_cugraph.py

Benchmark tooling

Committed in this PR:

• benchmarks/gfql/filter_pagerank/benchmark_to_cugraph_gpu.py
• benchmarks/gfql/filter_pagerank/filter_pagerank_stage_breakdown_gpu.py
• docker/benchmark-rapids-official-to-cugraph-gpu.sh
• docker/benchmark-rapids-official-gfql-gpu.sh
  • supports CUDA_VARIANT
  • supports RUN_STAGE_BREAKDOWN=1
• docker/test-rapids-official-local.sh
  • adds WITH_BENCHMARK=1 timing output as TIMING_JSON=...
• docker/benchmark-rapids-official-matrix.sh
  • timing-summary wrapper over the official RAPIDS image matrix

Follow-up issues

• #977 Investigate RAPIDS 26.02 string-ID graph-build regression with pure cuDF/cuGraph repro
• #978 Feature request: cache/reuse cuGraph graph projections across repeated GPU algorithm calls

Validation

Core validation on the kept optimization paths:

• PYTHONPATH=. pytest -q graphistry/tests/plugins/test_cugraph.py graphistry/tests/compute/test_hop.py graphistry/tests/test_compute_hops.py graphistry/tests/compute/test_chain.py graphistry/tests/test_compute_chain.py
  • 196 passed, 29 skipped
• ./bin/typecheck.sh
  • success
• added amplification coverage for:
  • to_cugraph() explicit kwarg forwarding and ineligible-ID no-op behavior
  • hop() fallback-equivalence on duplicate seeds and numeric-ID topologies

GPU validation on the kept optimization paths:

• hop() GPU stage-breakdown spot checks
  • 26.02-cuda13 / twitter
    • search1_median_s: 0.1121
    • pagerank_median_s: 0.0280
    • search2_median_s: 0.1043
    • pipeline_total_median_s: 0.2417
  • 25.02-cuda12 / twitter
    • search1_median_s: 0.1354
    • pagerank_median_s: 0.0322
    • search2_median_s: 0.1140
    • pipeline_total_median_s: 0.3199
• to_cugraph() GPU conversion spot checks
  • 26.02-cuda13
    • synthetic_contiguous: total 0.0200, build 0.0123, expected_vertex_match=True
    • synthetic_offset: total 0.0220, build 0.0140, expected_vertex_match=True
    • synthetic_string_gplus_shape: total 0.0504, build 0.0466, expected_vertex_match=True
    • twitter: total 0.0516, build 0.0358, expected_vertex_match=True
  • 25.02-cuda12
    • synthetic_contiguous: total 0.0255, build 0.0151, expected_vertex_match=True
    • synthetic_offset: total 0.0349, build 0.0227, expected_vertex_match=True
    • twitter: total 0.0584, build 0.0382, expected_vertex_match=True

Benchmark-shell validation for the newly committed timing helpers:

• bash -n docker/test-rapids-official-local.sh docker/benchmark-rapids-official-matrix.sh
  • success

CI is green on the last code-path validation head, and the benchmark-shell commit is now queued on top of it.

Findings

Accepted local wins:

• hop() undirected single-hop fast path
  • gplus GPU pipeline improved about -40% on both RAPIDS 25.02 and 26.02
• contiguous integer to_cugraph() fast path
  • real wins on contiguous integer graph families
• 26.02-cuda13 is consistently better than 26.02-cuda12

Remaining regression story:

• real GFQL gplus warm pipeline, 25.02-cuda12.8 -> 26.02-cuda13
  • pipeline_total: 1.7019s -> 2.1780s (+27.97%)
  • pagerank stage: 0.3240s -> 0.5108s (+57.65%)
• direct to_cugraph()+pagerank, same workload
  • total: 0.8171s -> 1.3695s (+67.60%)
  • build: 0.7630s -> 1.3215s (+73.20%)
  • pagerank: 0.0537s -> 0.0487s (-9.31%)

Pure RAPIDS follow-up:

• we now have a cuDF/cuGraph-only reproducer with no PyGraphistry imports
• strongest small-ish repro shape:
  • synthetic_string_gplus_shape
  • 10,000,000 edges
  • 107,614 unique vertices
• 25.02-cuda12.8 -> 26.02-cuda13
  • total: 0.1936s -> 0.3188s (+64.67%)
  • build: 0.1861s -> 0.3148s (+69.16%)
• sparse integer control remained slightly faster on 26.02

Known limits

• there is no safe local sparse-ID pre-factorization to keep in this PR
• the remaining large sparse string/object-ID renumber regression still appears upstream-facing in cuGraph graph build / renumbering
• the cache-reuse idea is tracked separately as a follow-up feature request and is not bundled into this PR

[lmeyerov]

move and cleanup