[Kernel][SM70] Group QSA Page4 prefill

[yangzhuxinyzx]

Purpose

Follow up #378 with the exact grouped Page4 QSA prefill route for Qwen3.8
Flash-Next on SM70/V100.

• Group eight adjacent query rows so their overlapping sparse Page4 K/V blocks
  are loaded once, while preserving an exact 4-bit mask per query row.
• Use Volta WMMA for QK and PV and skip inactive row tiles.
• Add a deterministic GPU planner that maps logical QSA selections through the
  paged cache, groups blocks by active-row category, and validates request,
  visibility, logical-page, tail, and physical-page bounds.
• Keep the existing single-row Page4 route as a 1024-partition fallback.
• Add --cuda-profile-case so one case can be captured with Nsight in an
  already-loaded engine while endpoint gates remain outside the capture range.
• Record the design, correctness gates, and isolated measurements in the SM70
  migration control document.

Test Plan

1. Rebuild flash-attention-v100 for SM70.

2. Run the QSA operator/cache and V1 Qwen3.8 cache targeted suites:

   pytest -q \
     tests/benchmarks/test_benchmark_sm70_decode.py \
     tests/models/qwen4_exp/test_qsa_ops.py \
     tests/models/qwen4_exp/test_qsa_cache.py \
     tests/v1/core/test_qwen4_exp_kv_cache.py

3. Compare the grouped route against the Triton reference using production
   Qwen3.8 geometry, including early rows, Page16/Page4 boundaries, a partial
   tail, nonmonotonic physical pages, and all-invalid padding.

4. Run the no-MTP TP4 endpoint gate at 32K/64K/131K and capture only an 8192
   profiling case with Nsight Systems and NVML in the same loaded engine.

Test Result

• Targeted tests: 19 passed, 16 warnings in 8.72s.
• Four repeated planner/kernel replays are bitwise deterministic.
• 4096-row hybrid Page16/Page4 correctness gate: relative L2 2.434e-4,
  cosine 1.0; an all-invalid group is bitwise-equal zero.
• Production-layout isolated route, including Python dispatch, GPU planning,
  padding, and attention: 16.132 ms median (15.588 ms minimum) versus
  Triton 55.940 ms, or 3.468x; maximum absolute error 1.22e-4, relative
  L2 3.535e-4, cosine 0.99999988.
• Matching no-MTP TP4 pure-prefill throughput is 5998.65/5777.43/5450.92
  tok/s at 32K/64K/131K, improving the retained
  4532.07/4446.64/4108.16 baseline by 32.36%/29.93%/32.69%. An exact 8K
  case is 6394.74 tok/s. Arithmetic, Chinese, and every performance-case
  token hash are bitwise identical to the baseline. The long endpoint sweep
  overlapped host/disk activity from an unrelated GPU4-7 model load, so these
  are conservative values.
• The clean double-locked Nsight capture reduces grouped QSA attention to
  9.632 ms per QSA layer/rank plus 0.362 ms planning, versus the old
  55.151 ms path (5.518x, 81.88% lower). Per-GPU kernel duty is
  98.27-98.61%; the post-QSA first hotspot is now the routed NVFP4 MoE chain,
  not launch gaps or QSA.

---

Essential Elements of an Effective PR Description Checklist

• Purpose is described and linked to the merged precursor.
• Test plan includes focused commands and the required endpoint gate.
• Current correctness and isolated performance results are included.
• The SM70 migration control document is updated.

[yangzhuxinyzx]

TP4 endpoint acceptance — passed

Final source/workload contract:

• PR head: c334a77f8c2a0ab7b06970d40be49c3b755c3255.
• The loaded Flash-V100 extension SHA256 is
  108d740581aac41f9d72618b366622ef6ac2c294837a7107fe544aa149ac3636;
  it was built before the final clang-format-only commit.
• 4 x V100-SXM2-32GB, TP4, V2, Qwen3.8-Flash-Next-NVFP4, ModelOpt NVFP4,
  FP16 activations/KV, MTP off, 8192-token chunked prefill, PIECEWISE graph.
• Quality-safe route: online QPN8, FP8 QPN8, GDN QPN8 split, QPN2, and QPN2
  prefill are all explicitly disabled. QPN-M1 only repacks checkpoint-native
  FP4 expert weights for decode.

Case	Pure prefill	Prompt throughput	Repeats
8K	1.281053 s mean / 1.280154 s median	6394.74 tok/s mean / 6399.23 tok/s median	3
32K	5.462565 s	5998.65 tok/s	2
64K	11.343458 s	5777.43 tok/s	2
131K	24.032633 s	5450.92 tok/s	1

The 8K pure-prefill samples are
1.283281 / 1.280154 / 1.279723 s (0.124% population CV). Relative to the
latest QPN8-off pre-Page4 8K observation (1.842520 s, 4446.08 tok/s), this
is 1.4383x, a 43.83% throughput increase and 30.47% latency reduction.
Relative to the retained BN16 long-context baseline, 32K/64K/131K improve
1.3236x / 1.2993x / 1.3269x.

Quality and route gates:

• Arithmetic remains 42<|im_end|>, token hash
  93def17bbf14ec663ecd18061fa1ebaa5c7171deb4b1c64a16cb6c5e47213173.
• Chinese remains 在标准大气压下，水的沸点是100摄氏度。<|im_end|>, token hash
  11d98c01ccdae0ca8d2d19e36486b4a7e35c3042160f646a92019ab1e2a67334.
• All repeated performance-case token hashes are stable and exactly match the
  retained pre-Page4 baseline.
• Runtime log proves the grouped route at both full and tail chunks:
  rows=8192, groups=1024 and rows=8120, groups=1015.
• Peak device memory is 29493 MiB on every rank. Minimum host MemAvailable is
  34.441 GiB; minimum SwapFree is 244.539 GiB.

The focused 8K trace was recovered as both .nsys-rep and SQLite. Critical
rank NVTX wall is about 1260.5 ms; kernel span is 1288.5 ms. Grouped QSA
attention is 116.791 ms total across 12 layers and its planner is 4.371 ms.
The preceding selected-attention trace was 55.151 ms/layer, or 661.812 ms;
the new 121.162 ms attention+planner total saves 540.650 ms, closely
closing the measured 561.467 ms endpoint reduction.

The new first-order hotspots on rank 0 are NVFP4 MoE GEMM (308.187 ms,
24.22%), dense FP16 GEMM families, NCCL all-reduce (118.651 ms, 9.33%),
grouped QSA attention (116.791 ms, 9.18%), hyperconnection combine/norm
(82.453 ms, 6.48%), row expansion (71.380 ms, 5.61%), and FlashQLA GDN
(56.647 ms, 4.45%). QSA is no longer the dominant bottleneck.

Final local evidence:

• JSON: .artifacts/qwen38_grouped_page4_prefill_20260828/results/grouped-page4-endpoint-profile-tp4.json
• Log: .artifacts/qwen38_grouped_page4_prefill_20260828/results/grouped-page4-endpoint-profile-tp4.log
• Trace: .artifacts/qwen38_grouped_page4_prefill_20260828/profiles/grouped-page4-endpoint-profile-tp4.nsys-rep
• SQLite: .artifacts/qwen38_grouped_page4_prefill_20260828/profiles/grouped-page4-endpoint-profile-tp4.sqlite

The endpoint/quality gate passes. Keep the PR Draft while the post-Page4 trace
is used to select the next GEMM/MoE/hyperconnection optimization.