[qwen3_5] GDN/linear-attention performance pack: fused QKVZBA split, fused decode, PDL, MTP scatter (ports from sgl-project)

[TobyMint]

What

Performance pack for the GDN / linear-attention path of the Qwen3.5/3.8 hybrids, ported from sgl-project (sgl-project#34859 / sgl-project#35758). Stacked on #751 (its 2 correctness commits are included below; after that merges this PR shows only the 4 perf commits). All validated on H20 (SM90) with Qwen3.8-27B bf16; each port keeps its upstream tests, all green locally.

1. Ratio-3 fused QKVZBA split (triton_gdn_fused_proj.py, qwen3_5.py)

tl.arange only accepts power-of-two extents, so the fused split/reshape/cat contiguous kernel could not serve the dense 27B hybrids' v/k head ratio 3 — the model fell back to fix_query_key_value_ordering (two .contiguous() copies plus a torch.cat of q/k/v) in every one of the 48 GDN layers. The kernel now walks non-power-of-two groups one HEAD_V-sized head at a time (bitwise-exact; power-of-two groups keep the wide vector access). Upstream test test_gdn_fused_split_head_ratios.py ported (ratio 1/2/3/4, 4/4).

2. Fused GDN decode QKVZ/BA unpack + indexed Conv1D (gdn_backend.py, default on)

Decode keeps the projection pair packed; one Triton kernel unpacks it, applies the indexed causal Conv1D state update, and materializes mixed_qkv/z/b/a — replacing the split kernel plus a second conv launch. Eligibility-checked with explicit fallback, env kill switch SGLANG_ENABLE_GDN_DECODE_FUSED_PROJ_CONV (registered per the Envs convention, names kept upstream-verbatim for future syncs). Upstream test ported: parity vs split+causal_conv1d_update, widths 2-4, dtypes, CUDA-graph replay, OOB-slot masking, fp8 fallback (5/5).

Known caveat (documented in code): the debug-only live parity check (SGLANG_GDN_DECODE_FUSION_VERIFY_REAL_TENSORS=1) fires during CUDA graph capture because capture dummy batches alias mamba slots across rows while the reference uses a private state copy; real batches have unique slots.

3. PDL on GDN recurrent and gated-norm kernels (fla/*, SM90+)

Chain fused_sigmoid_gating_delta_rule_update and layer_norm_gated_fwd behind their producer conv1d_update via programmatic dependent launch. Scheduling-only and bit-exact.

4. Single-launch fused conv-window scatter for the MTP verify commit (mamba_state_scatter_triton.py)

The post-verify conv-state commit looped fused_conv_window_scatter_with_mask once per conv type and again for the track set. fused_conv_window_scatter_multi covers up to 8 (dst, src) pairs and both index sets in one launch; falls back when ineligible. Also relaxes the scatter kernels' dst contiguity assert to accept envelope-strided unified-pool views. Upstream test extended with the multi path (4/4).

Measured-and-excluded: SM90 FlashInfer GDN prefill default

The same upstream PRs default FlashInfer GDN prefill on for SM90 (max_chunk=32768, fp32 state). We measured a regression on H20 and did not port the default — it stays reachable via --linear-attn-prefill-backend flashinfer:

• chunk 8192, 16×8192: TTFT 15.1 s (Triton) vs 17.8 s (FlashInfer), input throughput 2896 vs 2643 tok/s
• chunk 32768, identical flags: Triton wins both workloads by ~37% (16×8192: 4934 vs 3609 tok/s; 16×16384: 2403 vs 1757 tok/s)

Validation summary (H20 TP1, dense 27B bf16)

Item	Result
Kernel tests	ratio 4/4, decode-fusion 5/5, scatter 4/4 — all bitwise/unit-green
Decode TPOT @bs8	18.86–19.04 ms across the individual ports (neutral-positive each)
Decode TPOT @Bs64	21.80–22.06 ms; PDL +0.8% output throughput; gsm8k 0.980
gsm8k 5-shot 200q	0.965–0.980 across configs (bf16 baseline band)
MTP NEXTN 3/1/4	gsm8k 0.975, accept len 3.55–3.60
Prefill deployment note	--chunked-prefill-size 32768 boosts prefill ~70% on H20 but needs manual --mem-fraction-static (~0.75) and a bounded --cuda-graph-bs-prefill list (auto planner goes negative on the mamba pool; full 106-size capture ≈35 min)

---

CI States

Latest PR Test (Base): ⚠️ Run #33582388625
Latest PR Test (Extra): ⚠️ Run #33582388460