perf(gdn): fuse alpha+beta into single GEMV for decode

[kekzl]

Summary

Two tiny [d_model, n_heads] matvecs (gdn_alpha and gdn_beta) that share the same input no are replaced with one matvec against an interleaved [d_model, 2*n_heads] packed weight built at load time. For n=1 decode the output is [α₀..α_{H-1}, β₀..β_{H-1}] contiguous in ssm_dt_buf_, so alpha_proj_out and beta_proj_out are trivial offset views — no deinterleave kernel needed.

Prefill (n>1) and any non-FP16/BF16 weight (Q*_K, MXFP4, NVFP4 raw-quant) keep the original 2-call dispatcher untouched.

Why this PR

PR #151's nsys findings flagged 5 FP16 internal projections per GDN layer × 60 layers as dominating ~21.7 % of Qwen3.6-NVFP4 decode kernel time. With CUDA Graphs ON (default), launch overhead is largely amortised so the practical headroom from fusing tiny launches is bounded. This is the smallest-possible first cut: alpha + beta share input, are tiny enough that interleave-into-one-buffer-and-take-views is the entire change, and don't touch the storage planner.

Bench

RTX 5090, identical seed/prompt, 256-token decode, graphs ON:

Model	clean main	+ fusion	Δ
Qwen3.6-35B-A3B-NVFP4	171.20	173.28	+1.2 %
Qwen3.5-4B-GDN Q8_0	193.45	193.83	unchanged (fusion inactive — Q8_0 weights)

make verify-fast: PASS (decode/prefill within thresholds, graphs gate 1.62×, smoke coherent).

Filter

Fusion only fires when:

• both gdn_alpha and gdn_beta are on-device
• both have qtype == F16 || qtype == BF16
• shapes match [d_model, n_heads]

Q8_0 / Q4_K / MXFP4 GDN models stay on the raw-quant 2-call path (which they have to use anyway for correct dequant dispatch). NVFP4-prequant Qwen3.6 / Qwen3.5 hybrid paths see the fusion fire because SSM-internal projections are kept FP16 (they're excluded from the NVFP4 cache for accuracy — see weight_map.cpp:842).

VRAM cost

Per layer: 2 × d_model × n_heads × 2 bytes for the packed buffer (originals stay live for the prefill path). Qwen3.6-35B: 2048 × 16 × 4 = 128 KB × 60 GDN layers = ~7.7 MB total. Negligible.

Out of scope (separate PR)

4-way fusion (ssm_in + gdn_gate + alpha + beta) — closes more of the 21.7 % path but needs storage planner + NVFP4 cache co-ordination because ssm_in and gdn_gate are bigger, have different output dims, and ssm_in's output is already sliced into z/xBC/dt downstream. Treating that as a follow-up keeps this PR small and reviewable.

Test plan

• make verify-fast (build + tests + perf + smoke + graphs gate)
• Qwen3.6-NVFP4 coherent decode ("capital of France" smoke)
• Qwen3.5-GDN Q8_0 coherent decode (Babbage essay)
• Bench A/B clean-main vs patched on Qwen3.6-NVFP4 + Qwen3.5-GDN

🤖 Generated with Claude Code