[Perf] Add opt-in DFlash2 schedules and numerical audits - #556

Purpose

DFlash2 TP4/B1/q8 verification spends substantial time in projection, state layout, attention and scheduling. This change adds independently selectable schedules and reproducible operator/model audits for the current approximately 16-ms complete-round path. The common installer selects GDN BV2, context/probe scheduling, grouped E4M3 attention and sparse selection through actual tensor/shape guards, independently of target quantization names. A separate hashed QPN2 manifest selects NVFP4 cap64 projections and TP4 publication. Unsupported shapes retain their existing operators.

The PR also contains the grouped-attention warp synchronization repair and reuses dependency #563's singleton-prefill metadata fix. It extends existing PR #556 rather than creating duplicate optimization or prefill-fix PRs. No experimental route is enabled by default; source integration does not certify unfinished runtime admission. Sub-15-ms complete rounds have not been achieved.

Test Plan

Freeze source, native libraries, model weights, sampling and physical GPUs 4–7. Preserve TP4/B1/q8, E4M3 target KV, FP32 logits/state, FP16 draft transport, T1/k20/p.95/xhigh and natural EOS. CUDA 12.8, Torch 2.10.0+cu128 and Python 3.12.13 are pinned. The server has --max-model-len 262144; each natural request uses max_tokens=262144-actual_prompt_tokens verified with the server tokenizer. Input is not truncated. One-token warmups and bounded operator diagnostics are excluded from natural-generation scoring.

The frozen corpus contains 148 prompts: 32 each GSM8K, MATH500, HumanEval and MBPP, 16 LiveCodeBench v6 and four JSON/tool cases, with seeds 0/1/2. Independent startups use five warmups and five measured requests per speed fixture/arm. Whole-stack all-off controls, all-layer repeatability, public-entry parity, FP8 and long-prefix boundaries are separate gates. These are dataset subsets, not full-dataset benchmark claims.

Test Result

• QPN2 publication/cap64, sparse gathering and grouped-attention operator/graph checks pass with byte-equal outputs. Individual fixed-prefix comparisons retain 144 records per arm with native-logit TV=0 and no support/top-1 changes. Their recurrent-state coverage is layers 0/1, not the queued all-layer final-combination audit.
• GDN BV2 passes all 48 layers on all four ranks with at least 2230 live shadow calls per layer and zero output/state bit differences. Eight selectors, strided pools, graph input changes, padding/retired slots, memcheck and racecheck pass. Context/probe scheduling retains its CPU numerical guard and passes at least 1280 live shadow comparisons per rank.
• The first new 256K-capacity speed startup measures request-average complete-round medians of 16.571456 → 16.339483 ms for release1k and 16.144814 → 15.907431 ms for MBPP28. Candidate pure decode medians are 182.259 / 306.098 token/s. Tokens, natural EOS and acceptance match. This A/B isolates BV8/BV2 inside the otherwise shared candidate stack; it is not the whole-stack comparison.
• The historical 16K-output-cap seed-zero campaign completes 148/148 exact incremental pairs. Both arms score GSM8K 30/32, MATH500 31/32, HumanEval base 31/32 and enhanced 30/32, MBPP original assertions 32/32 and eligible EvalPlus enhanced 29/31, LiveCodeBench 11/16, and structured fixtures 4/4. Six capped generations lack final code. These historical scores do not admit the new capacity protocol. Twelve bounded whole-stack controls also match exactly.
• The new capacity campaign has completed five formerly capped long-output pairs at 21162, 76955, 34520, 70725 and 52704 tokens, all natural stops with identical tokens/acceptance/finish/tool semantics between BV8/BV2. Executable scoring and the complete new campaign remain pending. LiveCodeBench/21 averages 26.916502 ms/round and 122.378 token/s; short-context 16-ms results are not a long-context speed claim.
• The LiveCodeBench scorer now rejects negative error sentinels using pinned official result > 0 semantics; eight boundary checks pass. Original assertions and EvalPlus eligible subsets have separate denominators.
• Dependency [Bugfix][Core] Preserve singleton prefill semantics in speculative GDN #563 passes 32 focused CPU metadata tests. Singleton/history-reuse GPU checks remain queued. Latest focused command: CUDA_VISIBLE_DEVICES='' .venv/bin/python -m pytest tests/v1/attention/test_gdn_metadata_builder.py --confcutdir=tests/v1/attention -q -k 'singleton_prefill or gdn_build_classification or common_gdn_metadata_matches or mixed_decode_stays_decode_fastpath or full_cuda_graph_decode_padding_uses_pad_slot' (32 passed, 30 deselected).
• Scoped route/document pre-commit checks pass. The final documentation update passes pre-commit run --files docs/design/sm70_dflash2_acceptance_20260909.md. Rejected arithmetic or slower scheduling experiments remain disabled and documented.

Integration and remaining validation

Main base b6d91d61ff030fe325dc69eb8372a5c72d0374bd is already included. The running evaluation checkout remains frozen at a7cc5ae305149d7a9ffdf42fb224dff34e5606aa; integration and documentation work do not change its source or native libraries. Reproduction commands, hashes, numerical evidence, long-generation observations and artifact locations are retained in docs/design/sm70_dflash2_acceptance_20260909.md, docs/design/sm70_quasar_dflash2_15ms.md and docs/design/sm70_quasar_dflash2_resource_audit_20260909.md.

This integration preserves opt-in selection. Complete 256K-capacity multi-seed scoring, independent starts, whole-stack quality/acceptance, all-layer repeatability, public-installer parity, real FP8 and long-prefix validation remain open before any default promotion. The previous 4.33% repeat-start TV discrepancy is not an allowed error margin. No serving API, weights, model format, sampling semantics or context capacity is changed by integration.

AI assistance was used for implementation, diagnostics and reporting.

Assisted-by: Codex