perf(mtp): optimize Qwen3.5 GDN verification kernel

[sufubao]

Summary

• specialize the recurrent GDN kernel used by fixed-width MTP verification while preserving the variable-length path
• use LightLLM autotune v1 to select BV, num_warps, and num_stages for each CUDA Graph shape
• search 60 configurations: BV ∈ {4, 8, 16, 32, 64}, num_warps ∈ {1, 2, 4, 8}, and num_stages ∈ {1, 2, 3}
• key cached configurations by GDN dimensions, input/state dtype, fixed sequence width, sequence count, and token count
• protect in-place SSM state during tuning with mutates_args=["initial_state"]
• avoid request-index materialization, the cu_seqlens int64 cast, and the GDN gate copy in the MTP path
• trigger CUDA Graph tuning only from layer 0 / microbatch 0 using existing layer metadata
• restore committed regression tests for the optimized kernel, autotune behavior, gated RMSNorm, and Qwen3Next integration

Final H100 validation after rebase

Final head 0be08549f6fbd7fcf5764792d88e0f3b2184918c is based on the then-current upstream/main@aa7cab9ad12452fb08bfc316461242640b1b976c (PR #1509 merged).

The matched control is upstream/main@e1caab0b. Candidate and control used the same four H100 80GB GPUs, Qwen3.5-27B model files, TP4, MTP3, FP8 (fp8w8a8-pt-sgl), CUDA Graph shapes, launch environment, AIPerf corpus, and fixed streaming ISL=256/OSL=1024. Each point is the median of three recorded runs; all 42 final candidate runs completed with zero request errors.

Concurrency	control tok/s	final tok/s	gain	p99 TTFT change	p99 ITL change
1	323.10	333.86	+3.33%	+1.56%	-8.41%
2	632.21	655.13	+3.63%	-34.14%	-2.57%
4	1112.24	1185.66	+6.60%	-35.13%	-9.34%
8	2023.47	2200.98	+8.77%	-17.54%	-14.54%
16	3344.54	3533.89	+5.66%	-6.42%	-5.08%
32	4882.93	5380.04	+10.18%	-17.93%	-10.99%
64	6355.42	7021.90	+10.49%	-1.61%	-10.08%

Geometric-mean output-throughput gain across C1–C64 is +6.92%. The C1 p99 TTFT increase is below the 3% regression gate; all other reported p99 TTFT and ITL comparisons improve. First-run JIT/cache outliers at C1 and C4 occurred in both control and candidate sets and are retained in raw artifacts; medians are used as specified.

Final candidate run IDs on host 10-116-123-171:

• launch: 260829-005838-2654031-sudo-n-docker-run-d-name-pr1513-final-0be08549-p
• labeled residual-JIT warmup (excluded): 260829-010139-2664033-home-devsft-aiperf-venv-bin-aiperf-profile-model
• C1: 260829-010221-2666746-home-devsft-aiperf-venv-bin-aiperf-profile-model
• C2: 260829-010319-2668254-home-devsft-aiperf-venv-bin-aiperf-profile-model
• C4: 260829-010359-2669682-home-devsft-aiperf-venv-bin-aiperf-profile-model
• C8: 260829-010454-2671324-home-devsft-aiperf-venv-bin-aiperf-profile-model
• C16: 260829-010537-2673354-home-devsft-aiperf-venv-bin-aiperf-profile-model
• C32: 260829-010626-2676464-home-devsft-aiperf-venv-bin-aiperf-profile-model
• C64: 260829-010724-2681721-home-devsft-aiperf-venv-bin-aiperf-profile-model
• log archive / clean shutdown: 260829-010853-2687303-sh-c-sudo-n-docker-logs-pr1513-final-0be08549-p1

Raw results and server logs are retained under /home/devsft/qwen35_goal_20260829/pr1513_aa7_results.

Correctness and autotune evidence

• final local affected suite: 135 passed
• final H100 affected CUDA suite: 135 passed
• fixed-width and dynamic paths match the FLA reference, including output and written in-place SSM state
• production graph shapes load the keyed H100 cache and all 34 graph batches capture successfully
• fresh-cache TP4 startup on functional head 1c98ca56 tuned all 34 shapes once per rank and reached ready without traceback or OOM
• expanded 60-config H200 compile/equivalence sweep passed; the warp dimension is exercised by selected intermediate and large shapes
• worst recorded production-shape numerical differences versus FLA: output max abs 4.8828125e-4, output min cosine 1.0, state max abs 2.3841858e-7, state min cosine 1.0
• isolated service black-box comparison: 17 PASS / 2 baseline FAIL / 6 SKIP, with no PR-introduced failure
• Black, Flake8, repository pre-commit, and git diff --check pass

Established facts are the test, capture, and measured results above. The expected benefit on other models, GPUs, or workload distributions remains a hypothesis and is not claimed here. Every performance command was run through exp -m.