Skip to content

docs: add 2026 cache algorithm research - #1364

Merged
DefTruth merged 1 commit into
mainfrom
dev
Nov 19, 2026
Merged

DefTruth merged 1 commit into
mainfrom
dev

Conversation

@DefTruth

Copy link
Copy Markdown
Member

Why

Archive the 2026 DiT cache algorithm survey in the versioned cache-dit-cache-integration skill so future integration decisions can reuse the same evidence, layer boundaries, and acceptance criteria instead of relying on temporary research notes.

What changed

  • Added references/2026-cache-algorithms.md: 24 full-paper reviews and 6 abstract/official-page exclusions, with arXiv v1 dates, equations, tables, and source links.
  • Updated references/landscape.md with algorithm skeletons, candidate verdicts, paper-reported evidence, and integration constraints.
  • Added the report to SKILL.md ROUTE FIRST and recorded the BRACE / BeziCast / SeaCache / RACER PoC recommendations.
  • Synchronized references/paper-triage.md with the versioned report workflow.

Candidate priorities and paper-reported evidence

All entries below are PoC recommendations only. None is implemented or benchmarked in cache-dit by this PR. Numbers remain within their original papers and must not be compared across rows as a leaderboard.

Priority Candidate Integration seam Paper-reported evidence Proposed status
1 BRACE Pure calibrator for Bn_residual FLUX includes an approximately 5.55x FLOPs setting; HunyuanVideo includes N=6 First experimental calibrator PoC
2 BeziCast Pure calibrator with an internal forecast guardrail Up to 4.79x latency on FLUX and 4.11x on HunyuanVideo Second experimental calibrator PoC
3 SeaCache Replace the relative-L1 probe metric only Evaluated on FLUX / HunyuanVideo / Wan2.1; authors report approximately 0.2% filtering overhead Probe PoC; geometry, scheduler, and FFT costs require validation
4 RACER Split forecast shrinkage from disagreement-based gating Multiple paired high-acceleration settings across image and video backbones Validate shrink-only first, then the guard; no single-component attribution

The full report also covers GP-Refiner, HybridFF Adaptive, Spectrum, BeCARE, and the rejected or out-of-scope candidates.

Verification

  • pre-commit run --files ... on all four changed documents: passed.
  • pre-commit run --from-ref origin/main --to-ref HEAD: passed.
  • git diff --check: passed.
  • Rechecked arXiv v1 dates, current source references, Markdown tables, and cross-document links after review.
  • No cache-dit runtime tests, model inference, GPU benchmarks, or performance claims were added; this is a documentation-only research PR.

Compatibility

No runtime code, public API, configuration, CI, or model behavior changes. The temporary .tmp/research/ draft remains untracked; only the versioned skill documentation is included.

Closes #1363.

@DefTruth
DefTruth merged commit cea1c67 into main Nov 19, 2026
2 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

[RFC] 2026 DiT Cache Algorithm Integration Research

1 participant