perf(ops): pipeline Q4/Q5 Attention groups at T32

[MOVIBALE]

Decision

Keep the shared adjacent-group BK=128/GPB=2 grouped MMA mechanism and retain the exact T=32 route for Q4/Q5 A16 Attention only. Restore GDN T=32 selection to the upstream K64 route. No runtime cache detector, autotuner, device whitelist, or permanent experiment switch is added.

Why GDN is not enabled

At T=32, Attention improves in both Op cache cases, while GDN improves cold but regresses warm:

Op	B route cold/warm	AG route cold/warm	Change
Attention	101.536 / 85.840 us	72.928 / 68.176 us	-28.2% / -20.6%
GDN	124.928 / 92.304 us	99.568 / 102.288 us	-20.3% / +10.8%

Two same-protocol C8 blocks ran in reverse order, B/A/AG then AG/A/B. Every request completed 1024 tokens, batch stayed 8, and all six waves used the same request seeds. The table reports actual steady committed tokens and MTP work; no profiler was attached.

Block/order	Arm	Steady committed tokens / seconds / intervals	Batch	Completion / decode tokens	MTP rounds / drafted / accepted / fallback	tok/s
1/1	B1	6558 / 15.000 / 15	8.00	8192 / 8184	3319 / 9948 / 4863 / 2	437.2
1/2	A1	7190 / 17.081 / 17	8.00	8192 / 8184	3253 / 9746 / 4928 / 3	420.9
1/3	AG1	7220 / 18.022 / 18	8.00	8192 / 8184	3374 / 10113 / 4807 / 3	400.6
2/1	AG2	7348 / 18.000 / 18	8.00	8192 / 8184	3274 / 9811 / 4909 / 1	408.2
2/2	A2	7098 / 18.000 / 18	8.00	8192 / 8184	3309 / 9914 / 4873 / 2	394.3
2/3	B2	7068 / 19.000 / 19	8.00	8192 / 8184	3384 / 10139 / 4796 / 4	372.0

Mean throughput: B 404.6, A 407.6, AG 404.4 tok/s. AG versus A was -4.8% in block 1 and +3.5% in block 2. The sign flips. Rates also fall with wave position in both blocks (position means 422.7, 407.6, 386.3 tok/s). The small overall difference cannot isolate a GDN contribution, so GDN remains on K64. The earlier R8 B→AG +4.8% combined result is not used to justify retaining GDN.

C1/C4, prefill, and shared GPB=1 controls

Existing B/AG controls are reused: C4/T=16 stays on the old routes (baseline mean 327.1 tok/s; prior narrow AG mean 323.1), and the 2048-token prefill result is 3577.1 ± 53.9 vs 3580.3 ± 12.5 tok/s. The old B/AG C1 control was 154.0 vs 157.0 tok/s. A was not rerun at C1 because it does not select T=32. These are not new measurements of the A arm.

The existing full T=1..128 cold/warm curves and T=512/1024 points show no consistent GPB=1 regression. Isolated high warm medians at T=64/T=128 return near baseline in the other wave; the 512/1024 controls remain close. Raw CSVs are included in the review evidence.

Validation

• The existing Attention eager loop covers T=31/32/33; the T=32 graph test now compares all Q/G/K/V output rows against the FP64 reference in both replay phases: 917,504 output values total, with the existing tolerance.
• Existing targeted Attention, GDN, GDN snapshot, and GDN record tests pass 4/4. GDN T=32 eager and Record B=8/W=4 relation coverage are reused; no full Snapshot math result is claimed for B=8/W=4.
• Existing Attention T=32 Compute Sanitizer results are reused because the kernel mapping and arithmetic did not change.
• Existing C8 and prefill traces are reused for call shape/hotspot context. No profiled throughput is presented as official speed.
• The C1/C4/prefill controls and GPB=1 curve data are reused; no full candidate matrix was rerun.

PR #311 remains a draft. Do not merge automatically.