perf(qwen3.6): fuse decode kernel launches — shared-expert + GDN conv/L2-norm

[minion1227]

Summary

Reduce decode kernel-launch count on the Qwen3.6 path with two bit-identical kernel fusions (decode here is CUDA-graph / launch-bound, so per-layer launches cost real time):

1. Shared expert — fuse gate + up GEMVs and the SwiGLU into one one-warp-per-row kernel (hn staged once, both dot-products in one pass): 3 launches → 1 per layer (40 layers). down stays a separate GEMV.
2. GDN linear-attn layers (30 of 40) — fuse conv_split + the two per-head l2_norm_heads kernels into one block-per-head kernel (causal conv + SiLU, then the q/k L2-norm reduced in-block): 3 launches → 1 per GDN layer.

Both are bit-identical to main (same arithmetic and reduction order; projection/SiLU values rounded to bf16 exactly as the unfused paths do), so the accuracy gate is trivial. Both are no-ops for Qwen3-30B (no shared expert, no GDN layers) → the guard cannot regress. Only kernels/ + runtime/; touches no crowded/attempted file.

Proof of speedup

• Tested on RTX 5090 (sm_120)

Decode tok/s (end-to-end, Qwen3.6-35B-A3B-UD-Q4_K_M, n=256, ctx=0, bs=1):

	decode tok/s
before (main, cc356a5)	171.31
after (this PR)	176.37

→ +2.95% decode on the Qwen3.6 primary model.

# sparkinfer decode bench (qwen3_gguf_bench <model> 256 0), RTX 5090 sm_120, idle GPU, 3 runs each:
before (main):  171.27  171.32  171.33   -> 171.31 tok/s
after  (PR):    176.37  176.37  176.37   -> 176.37 tok/s
delta: +2.95%

# correctness gate (qwen3_gguf_score, this PR vs main, same input):
# per-token argmax + top-k logprobs bit-identical -> 100% top-1, KL = 0

[ai-hpc]

(wrong baseline — re-run)

[ai-hpc]

Auto-closed by sparkinfer eval: no single context cleared the 2% improvement gate and at least one context regressed.

[ai-hpc]

🔄 Evaluation was run against a wrong baseline — re-evaluating

The bot's same-box baseline measurement had two bugs during your PR's evaluation run:

1. The box was checked out to a stale PR branch (not origin/main), inflating the baseline to ~232-260 tok/s instead of the actual main speed (~172 tok/s without GDN_FAST, ~225 tok/s with it).
2. The Gated-DeltaNet fast kernel (merged in feat(qwen36): add optimized Gated-DeltaNet AR state update kernel wit… #229) was opt-in via SPARKINFER_GDN_FAST=1 env var — the eval bot never set it, so both main and PRs were measured on the slow path.

The correct Qwen3.6 main baseline (with GDN_FAST=1) is now:

context	correct main baseline
128-token	225.24 tok/s
512-context	222.86 tok/s
4k-context	212.42 tok/s

Your PR will be re-evaluated against this correct same-box main baseline once you rebase onto main. No code change needed on your side — the bot just needs a clean slate to measure against.

Apologies for the confusion — this was an eval pipeline error, not an issue with your submission.

[minion1227]

Rebased onto main (2a2ebde) as requested — ready for re-evaluation against the corrected same-box baseline (225.24 / 222.86 / 212.42, with GDN_FAST=1). No code changes; the two fusions are unchanged and bit-identical to the unfused paths (100% top-1, KL=0). Thanks for catching the eval-pipeline issue and re-running.

[ai-hpc]

✅ sparkinfer auto-eval — a711cd5

metric	value
label	eval:L
scored decode (128 ctx · 128-context · Qwen3.6)	259.09 tok/s
vs same-box main	225.24 tok/s → +15.0% (+33.9)
correctness (Qwen3.6 vs llama.cpp)	top-1 95.9% · KL 0.0207
Qwen3.6 128-token no-regression gate	259.09 tok/s vs main 225.24 tok/s · pass
Qwen3.6 512-context no-regression gate	256.08 tok/s vs main 222.86 tok/s · pass
Qwen3.6 4k-context no-regression gate	242.49 tok/s vs main 212.42 tok/s · pass
Qwen3-30B-A3B guard — accuracy	top-1 96.9% · KL 0.0177 · pass
Qwen3-30B-A3B guard — 128-token	498.15 tok/s · pass
Qwen3-30B-A3B guard — 512-context	473.08 tok/s · pass
Qwen3-30B-A3B guard — 4k-context	395.62 tok/s · pass
Qwen3-30B-A3B guard — 16k-context	331.83 tok/s · pass
Qwen3-30B-A3B guard — 32k-context	262.0 tok/s · pass

Verified speedup over same-box origin/main — 259.09 tok/s (main was 225.24 tok/s).

RTX 5090 (sm_120) · 128/512/4k/16k/32k guarded · scored vs same-box main · strongest context scores · built from source · correctness vs llama.cpp. Automated — not merged; merge manually after review.

[ai-hpc]

The round's merge-first PR was just merged. Please rebase this branch onto main — the bot re-evaluates it on push (crediting your marginal gain on top of what merged).

[ai-hpc]

✅ sparkinfer auto-eval — 8868b7c

metric	value
label	eval:M
scored decode (128 ctx · 128-context · Qwen3.6)	277.98 tok/s
vs same-box main	254.72 tok/s → +9.1% (+23.3)
correctness (Qwen3.6 vs llama.cpp)	top-1 98.2% · KL 0.0154
Qwen3.6 128-token no-regression gate	277.98 tok/s vs main 254.72 tok/s · pass
Qwen3.6 512-context no-regression gate	274.03 tok/s vs main 251.87 tok/s · pass
Qwen3.6 4k-context no-regression gate	258.85 tok/s vs main 239.0 tok/s · pass
Qwen3-30B-A3B guard — accuracy	top-1 96.5% · KL 0.015 · pass
Qwen3-30B-A3B guard — 128-token	481.38 tok/s · pass
Qwen3-30B-A3B guard — 512-context	458.09 tok/s · pass
Qwen3-30B-A3B guard — 4k-context	385.21 tok/s · pass
Qwen3-30B-A3B guard — 16k-context	326.97 tok/s · pass
Qwen3-30B-A3B guard — 32k-context	254.61 tok/s · pass

Verified speedup over same-box origin/main — 277.98 tok/s (main was 254.72 tok/s).

RTX 5090 (sm_120) · 128/512/4k/16k/32k guarded · scored vs same-box main · strongest context scores · built from source · correctness vs llama.cpp. Automated — not merged; merge manually after review.

[ai-hpc]

The round's merge-first PR was just merged. Please rebase this branch onto main — the bot re-evaluates it on push (crediting your marginal gain on top of what merged).

[ai-hpc]

The round's merge-first PR was just merged. Please rebase this branch onto main — once you push the rebase the bot re-evaluates it against the new frontier (crediting your marginal gain on top of what merged).