[Kernel][SM70] Scale NVFP4 E4M3 decode batches

[yangzhuxinyzx]

Purpose

Scale two exact SM70 inference-engine operator contracts without model/checkpoint/architecture identity checks:

• FP8-E4M3, GQA-6, D256 paged-XQA decode for B2-B16, including paired KV loads and long-context graph routes.
• Combined top-k/top-p masking for B8/B16 x 248,320 logits tensors using an eight-warp launch.

Qwen3.8-27B NVFP4 is the matched performance and quality workload only; it does not participate in runtime admission. Gates use SM70 capability, FP8 format, batch/tensor shape, paged-KV layout, graph context, and sampler settings.

Rollback switches:

• VLLM_FLASH_V100_E4M3_BATCH_XQA=0 restores scalar paged attention for B2-B16.
• VLLM_FLASH_V100_E4M3_BATCH_XQA_OPTIMIZED=0 keeps XQA but restores scalar KV loads, the original E4M3 converter, and baseline CTA routing. B1 always keeps the original converter.
• VLLM_SM70_TOPK_TOPP_B8_B16_8_WARPS=0 restores the prior Triton launch heuristic.

Base: onecat/main@f6a5b57b645867d87f83ada43231f2dd25b40a4a
Head: 5d05b74bb56f9d665b0f5c4788df3b17ef1fa6d3

Audit fixes

• Rebases cleanly over generic prefix-anchored SWA from [Attention][Core] Add generic prefix-anchored SWA on SM70 #281; both original [Kernel][SM70] Scale NVFP4 E4M3 decode batches #282 patches are git range-diff equivalent after rebase.
• Keeps the fast packed E4M3 converter inside batch optimized instances only, so B1 and rollback paths retain main SASS.
• Prevents Flash-V100-only E4M3 graph variants from being captured when FLASHINFER_SM70 is explicitly selected.
• Leaves all pre-existing XQA kernels unchanged: 115/115 common instances are SASS-identical, zero changed or missing; only six batch-optimized instances are added.

Performance contract

Four V100-SXM2-32GB, TP4, compressed-tensors NVFP4 weights, FP16 activations, FP8-E4M3 KV, Flash-V100, prefix cache, Mamba align, no speculation, FULL decode graphs. Official SPEED-Bench low_entropy, temperature 1.0, top-k 20, top-p 0.95, natural EOS, seed 20260822, 512 output cap, one warmup plus three measured repeats. Only steady full-batch pure decode is reported; TTFT, prefill, and endpoint throughput are excluded.

Context	B	Control tok/s	Candidate tok/s	Speedup	Candidate efficiency	FP8 reference efficiency
1K	1	72.65	72.62	1.000x	100.0%	100.0%
1K	8	402.98	465.24	1.154x	80.1%	87.4%
1K	16	641.95	772.44	1.203x	66.5%	72.3%
16K	1	63.10	63.00	0.998x	100.0%	100.0%
16K	8	186.97	356.52	1.907x	70.7%	67.7%
16K	16	222.61	527.00	2.367x	52.3%	49.1%

The geometric-mean scaling-efficiency gap to FP8 is 2.2% at B8 and 1.0% at B16. At 16K, NVFP4 exceeds the matched FP8 efficiency by 4.4%/6.5% at B8/B16.

Test plan and result

• Final-head focused CPU engine policy/dispatcher/sampler tests: 24 passed.
• [Attention][Core] Add generic prefix-anchored SWA on SM70 #281 prefix-anchored compatibility regressions on this head: 32 passed.
• Final-head SM70 source build with CUDA 12.0 / Torch CUDA 12.8 minor-version warning: passed and linked.
• Changed-file pre-commit including Ruff, clang-format, Markdown, mypy, SPDX, config and docs gates: passed.
• Same-toolchain SASS comparison to the [Attention][Core] Add generic prefix-anchored SWA on SM70 #281 main build: 115/115 common XQA instances identical, six new, zero changed/missing.
• V100 operator evidence on the same optimized math/routes: 8 attention comparisons (B2/B16 at p64/p128/p256 and B8/B16 long context), within one FP16 ULP and mean absolute error <= 1e-5; 2 sampler comparisons (B8/B16), bitwise equal to rollback.
• Matched GSM8K main/test, 32 prompts each at B1/B8/B16: 22/32 to 23/32 at B8, 23/32 unchanged at B16, zero invalid outputs.

No full-model end-to-end or throughput run was repeated after the final current-main rebase; the audited performance patches are equivalent, the measured optimized instances are unchanged, and the requested merge scope is source plus focused validation.

[yangzhuxinyzx]

CI note: the repository-wide pre-commit job failed on existing all-files debt, not on this PR's files. The job ran pre-commit ... --all-files, and Ruff format/clang-format/Markdown/SPDX plus legacy policy hooks changed or flagged unrelated base files. Intersecting the CI-generated diff headers with this PR's 12 changed paths gives zero matches. The same baseline failure is present on merged PRs #278 and #279.

For this head, the changed-file pre-commit suite passes locally in full (Ruff check/format, typos, clang-format, Markdown, mypy, SPDX, config/docs gates), along with 36 focused CPU tests and 10 focused SM70 GPU tests. The PR remains Draft pending the queued post-rebase TP4 full-model smoke.

[yangzhuxinyzx]

Final audit note: pre-run-check passed. The repository-wide manual pre-commit job failed because its all-files formatter rewrote 228 pre-existing main-branch files; that set has zero overlap with this PR's 12 changed files. The complete changed-file pre-commit suite passes on the final head. Additional current-main evidence: 24 focused engine-policy tests, 32 #281 compatibility regressions, successful SM70 compile/link, and 115/115 common XQA SASS instances identical with only six new optimized instances. The all-files baseline drift is not attributed to #282.