[richiejp]

Row

KERNEL-SSM-MAMBA — Qwen GDN post-conv and causal-conv kernel sublevers; one row per PR.

Before starting

• Issue/PR search and existing claim: this is a stacked follow-on to perf(gdn): dispatch exact causal-conv chunks #127, which supplies the exact causal-conv chunk descriptors measured here. No separate matching open PR was found for these two opt-in kernel specializations.
• Roadmap or matrix row, plus scripts/ready-for-helper.py result when applicable: .agents/kernel-matrix.md:157 (KERNEL-SSM-MAMBA, INVENTORIED), feeding ROAD-V1-C2-LOCAL-BF16. Spikes: .agents/specs/sm120-qwen35-postconv-token-tile-2026-08-08.md and .agents/specs/sm120-qwen35-conv-channel-tile-2026-08-08.md.
• Exact current-code and test/evidence anchors inspected: src/vt/cuda/cuda_gdn.cu, src/vt/cuda/gdn_prefill_conv.h, tests/vt/test_gdn_prefill_conv.cpp, tests/vt/test_ops_gdn.cpp, and pinned vLLM/FLA causal-conv plus fused post-conv kernels recorded with file:line anchors in the two spikes.

What changed

Add two independently selectable, default-off CUDA specializations. VT_GDN_POSTCONV_TOKEN_TILE=1 ports the upstream-shaped 16-token/four-warp post-conv schedule while preserving the existing 128-lane reduction tree byte-for-byte. VT_CONV_CHANNEL_TILE=1 specializes the causal-conv path for width four at the existing one-channel-per-thread grid; arm =2 is retained as the measured 256-channel/two-channels-per-thread falsification arm. Production dispatch and the portable mutation gate share one callback dispatcher, so the tests prove that each selected arm is actually invoked.

Evidence

• scripts/agent-preflight.sh passes after rebasing onto canonical upstream/main c05cee1d2; all record gates, mutation suites, live-row audit, and committed-range document checks are green.
• Tests covering this change: contained CUDA rebuild including an explicit vllm-bench relink; portable dispatch 9/9 cases, 88/88 assertions; CUDA GDN 67/67 cases, 4631/4631 assertions; Qwen3.5 paged-forward 4/4 cases, 8/8 assertions. The accepted production arms produce identical token hashes.
• Same-change doc obligations: docs/STATUS.md and docs/BENCHMARKS.md updated; the model/feature surface did not expand. Full profiler evidence and the void stale-link series are retained in the two specs and .agents/benchmark-record.md.

Speed claims

• The operator ran the numbers under ${GPU_LOCK} and they are recorded in docs/BENCHMARKS.md with the repro recipe.

On the local RTX 5070 Ti / Qwen3.5-4B c32 workload, the post-conv tile reduces graph-node time 227.887 to 122.587 ms (1.859x) and improves every observed enclosing axis, including total/output throughput by 0.532%. The causal-conv K4 arm reduces graph-node time 234.605 to 219.506 ms (6.44%) and improves whole-run total/output throughput by about 0.12%. The 256-channel arm is 4.05% slower than the K4 arm and is recorded as falsified.

Honest gaps

• This PR is stacked on perf(gdn): dispatch exact causal-conv chunks #127. Until perf(gdn): dispatch exact causal-conv chunks #127 merges, GitHub shows its exact-chunks commits in this PR too; after that merge, this PR reduces to the opt-in follow-on commits.
• Both specializations remain opt-in pending repeated A/B and hardware-unavailable 27B/35B gates. The 4B result is not extrapolated to the release models.
• Same-tool residuals remain: post-conv is 1.135x and causal conv K4 is 1.509x slower than pinned vLLM.
• The first production causal-conv profile series was void because vllm-bench had not relinked the rebuilt CUDA library. The accepted series explicitly relinked the benchmark and trace-proved the specialized kernels.
• Current main c05cee1d2 independently fails check-device-leakage.py: src/vllm/v1/worker/gpu/runner.cpp contains one new hard-coded DeviceType::kCUDA (DSR kcuda 1 > baseline 0). perf(gdn): dispatch exact causal-conv chunks #127 and this PR inherit that merged-tree failure; neither kernel diff introduced it.