{ "status": "rejected", "reason": "The exact temporal two-pair QK/PV schedule improved the isolated mainloop by only 0.10% and missed the absolute 220 ms gate.", "device": "NVIDIA GB10", "compute_capability": "SM121", "source_commit": "d1a57a546c3d395b1ffcbeecc66d81db76f3b4b5", "candidate": { "name": "temporal-pair", "description": "Two warp pairs alternate QK/online-softmax and prior-tile PV while preserving private per-warp scores, softmax state, and output accumulators.", "production_dispatch": "unchanged" }, "correctness": { "compute_sanitizer_errors": 0, "short_shape_cases": 13, "short_shape_result": "byte-exact", "real_block": 24, "real_tokens": 37810, "real_heads": 56, "reference_output_sha256": "4c666c20f5f8f651158a2ced33ccff08f3bada07665c595b99008d171db30574", "candidate_output_sha256": "4c666c20f5f8f651158a2ced33ccff08f3bada07665c595b99008d171db30574" }, "ptxas": { "baseline": { "registers_per_thread": 255, "static_spill_stores_bytes": 24, "static_spill_loads_bytes": 24 }, "candidate": { "registers_per_thread": 254, "static_spill_stores_bytes": 32, "static_spill_loads_bytes": 24 }, "dynamic_shared_memory_bytes": 32768 }, "timing": { "rounds": 5, "iterations_per_round": 10, "baseline_p50_ms": 245.437279, "candidate_p50_ms": 245.198769, "speedup_percent": 0.097272, "candidate_gate_ms": 220.0, "gate_result": "fail" }, "decision": { "ncu_followup": "skipped because the latency gate failed", "block_integration": "skipped", "trajectory_validation": "skipped", "runtime_integration": "not attempted", "conclusion": "Temporal separation alone does not unlock material INT-QK/FP8-PV overlap in the existing SM89 instruction stream on SM121. Further exact work requires a substantially different mainloop or native Blackwell numerical contract." }, "source_artifacts": [ "research/sage2_temporal_pair/patches/0001-sm89-temporal-two-pair-qk-pv.patch", "benchmarks/gb10-sage2-temporal-pair-short-20260825-p0.json", "benchmarks/gb10-sage2-temporal-pair-timing-20260825-p0.json" ] }