[CuTe] Tune B200 forward scheduling

[drisspg]

Stacked PRs:

• [CuTe] Add explicit forward config autotuning #2743
• ->[CuTe] Tune B200 forward scheduling #2741
• [CuTe] Tune GB300 forward scheduling #2738
• [CuTe] Add explicit forward config benchmark #2735
• [CuTe] Route forward through typed configs #2733
• [CuTe] Add typed forward configurations #2732
• [CuTe] Fix forward dynamic-shape correctness #2745

---

Human Note

Agent Note

Summary

This adds the same measured-policy treatment for B200. For long, sufficiently occupied BF16 output-only attention, D64 is better with direct output instead of TMA O, and D128 is better with 1CTA instead of 2CTA. The causal D64 control was basically flat/slightly worse, so causal stays on the old policy.

This is a Seaborn strip plot over the archived result rows: every circle is one actually timed workload cell, color is the campaign phase, the outlined diamond/error bar is the geomean and paired-round 95% timing interval, and X is the time-weighted result. The causal controls are real measurements, not an illustrative baseline.

New choice	Cells	Geomean [paired 95% CI]	Time-weighted	Min–max	W/N/L
Dense noncausal D64: TMA O → direct O	60	1.204× [1.203, 1.206]	1.191×	1.171–1.264×	60/0/0
Dense noncausal D128: 2CTA → 1CTA	72	1.205× [1.203, 1.208]	1.158×	1.105–1.560×	72/0/0
D64 causal control: rejected	12	0.993× [0.989, 0.997]	0.998×	0.994–1.007×	0/12/0

A fresh old-explicit-baseline versus config=None rerun then confirmed D64 at 1.208× and D128 at 1.211× geomean across 12 cells per rule, with 24/24 selector and direct-correctness checks passing.

The D128 campaign now spells out the resolved register allocation alongside CTA topology: the measured 2CTA baseline is (softmax=176, correction=88, other=72) and the measured 1CTA candidate is (192, 80, 48). These are not new measurements or a changed policy; they are exactly the values the old kernel-side table derived during the archived run, now made explicit so the checked campaign still reconstructs the same two binaries.

The rule is exact SM100 BF16, dense noncausal, D=DV in {64,128}, MHA or packed GQA ratios 4/8/16, packed Q > 256, K≥2048, at least 64 M-blocks, 128×128 tiles, no LSE, and no SplitKV or optional feature path. SM103, FP16/FP8, causal/local, training/Flex, paged, sparse, QV/MLA, gather, modifier, sink, packed-MHA, and underfilled cases keep their previous defaults.

Reproduction and validation

Produced on the B200 machine with:

gpu-run 0 -- python benchmarks/fwd_config_bench.py \
  --campaign benchmarks/configs/fwd_config_b200.yaml \
  --out-dir agent_space/b200_fwd_config/full

• Hardware: NVIDIA GB200/B200, exact SM100, 152 SMs.
• Software: PyTorch 2.14.0.dev20260708+cu132, CUDA 13.2, CUTLASS DSL 4.6.0.dev0, driver 580.126.20.
• Campaign: 144 cells total, 22 unique fake-tensor specializations, four independent correctness representatives, seven alternating rounds × 31 fixed-pointer CUDA-graph samples per arm.
• The plot generator reads the complete 144-cell result directly and recomputes the displayed aggregates.
• Holdouts use independent odd lengths and unseen batch/head combinations; every retained discovery, boundary, and holdout cell won.
• The 64-M-block floor and exact packed-layout predicate keep the production selector inside the measured region.
• Full raw samples, manifests, excluded controls, final confirmation, environment, validation logs, and checksums are archived at evidence commit 0a9f4dc.