[None][feat] add Rubin topology and prototype locality-domain sharding to GVR V2 decode

[longcheng-nv]

Dependency

Built on #18446. Keep this PR in Draft and rebase after #18446 lands. Companion prefill work is in #18764.

Summary

• Extend GVR V2 self-sampling decode to SM107 using runtime SM topology instead of B200-specific constants.
• Add a default-off use_gvr_locality_domain prototype that splits eligible decode requests across two Rubin locality domains.
• Keep each request and its next_n rows together, with proportional partitioning based on the reported domain SM counts.
• Preserve the existing full-device path for BS=1, small workloads, unsupported configurations, and whenever the prototype is disabled.
• Fail closed on cold CUDA Graph capture, device/topology mismatch, silent partition-context failure, or invalid unsplit tensor geometry.

Current scope

This is topology enablement plus an experimental GVR-only compute-sharding path, not a completed maximum-performance Rubin implementation.

• The indexer logits producer still runs on the full device.
• The produced logits remain ordinary global allocations; they are not domain-local.
• TMA, PDL, domain-local logits/K-cache allocation, producer-side locality, and BS=1 column sharding/merge are not implemented here.
• Dispatch thresholds are provisional and must be derived from R200 measurements before any default-on or whitelist decision.
• The prototype supports one in-flight model execution per CUDA device. Concurrent GraphExec replay or multiple engines sharing a device are not supported yet because the execution-domain workspaces/events are shared.
• Each eligible graph shape must run eagerly once through the same TopK instance before capture.
• The new public prototype field requires the normal telemetry/privacy CODEOWNER review.

Validation

Passed locally:

• GVR host router tests: 11/11.
• Locality planner/orchestration tests: 17/17 through a dependency-stub harness.
• Planner property sweeps over BS=1-1024, compressed N=4K-256K, K={512,1024,2048}, supported next_n, and symmetric/asymmetric topologies; all eligible plans were request-aligned, disjoint, and exhaustive.
• Existing B200 route parity: 21,504 route cases and 43,008 streaming comparisons.
• Relevant pre-commit formatting/lint, Python compilation, DCO, and whitespace checks.

R200 kernel validation has not been completed. Full package/DSA pytest collection is blocked in the current environment by an external bindings.so mismatch (global_steady_clock_now), and the installed CuTe DSL does not provide the Rubin helper needed to compile the SM107 path. The stub tests are not a substitute for SM107 execution.

Draft exit criteria

Before marking Ready, run correctness, stability, and performance A/B on R200 for:

• Every integer BS from 1 through 1024 for route/correctness coverage, with performance focus around domain-SM and routing boundaries.
• Compressed N={4K,8K,16K,32K,64K,128K,256K} (V4 ISL/4).
• K={512,1024,2048}, compression ratios {1,4}, and supported next_n values.
• Full-device versus strict/balanced locality-domain execution.
• Eager execution, warmup/capture, and at least 1,000 alternating CUDA Graph replays.
• GVR-only latency and indexer-producer + GVR end-to-end latency with CUDA Graph/PDL settings recorded.
• Random and real indexer logits, variable/short rows, ties, denormals, and +/-inf; keep the existing NaN contract separate.

The results must justify or retune the provisional dispatch gate and confirm that SM100/SM103 code generation and all fallbacks remain unchanged.