perf(qwen35): carry the GDN decode state slice in registers across the kv_mem reduction

[CAICAIIs]

Summary

The gated delta rule decode kernel did its two passes (decay, then rank-1 update) with a global round-trip between them: pass 1 stored the decayed state slice back to global memory, the kv_mem reduction consumed it, and pass 2 reloaded it to apply delta·k and store again. The key slice is only 32 floats per thread — it lives in registers across the reduction instead, so global memory sees one read and one write per state element.

Measured (1x A100-40GB, vLLM 0.27.0 baseline)

1024-token prompts, greedy, vllm bench serve, zero failed requests; single run per cell:

cell	before TPOT ms	after TPOT ms	delta	vLLM 0.27 TPOT ms
1024/256 c16	13.51	12.98	−3.9%	11.00
bs1 1024/256	9.77	9.36	−4.2%*	8.32

• c16 "before" = upstream/main + perf(qwen35): align the logits GEMM width so the output projection stops landing on an align-1 sm_75 kernel #1046 (single-variable A/B against this change).
• bs1 "before" = upstream/main at 139d925 without perf(qwen35): align the logits GEMM width so the output projection stops landing on an align-1 sm_75 kernel #1046: the cell bundles this change with perf(qwen35): align the logits GEMM width so the output projection stops landing on an align-1 sm_75 kernel #1046's alignment effect at bs1.
• c16 output throughput: 936 → 1000 tok/s.

The kernel's per-layer-step time was 97.0 µs vs vLLM's FLA fused_recurrent_gated_delta_rule_packed 43.6 µs; the register-carried state removes the extra global round-trip that was the difference from FLA's bandwidth-floor implementation. Found via nsys kernel attribution (a fuller doc covering the decode kernel gaps is being prepared separately).

Testing

• The rewrite is arithmetically identical — same f32 ops in the same order, state values carried in registers instead of a global round-trip.
• hf_golden_gate TP1 2/2; lib tests 102 passed.

Claim boundary

Single-run cells on one GPU; draft pending review. The kernel still has no sm_80 tuning pass; the remaining GDN gap to FLA and the c16 GEMM family deficit (+2.8 ms/step) stay open in the attribution doc's queue.

[FeathBow]

@codex review