Skip to content

Remove DBPrune and unify calibrator CLI - #1366

Merged
DefTruth merged 3 commits into
mainfrom
dev
Nov 19, 2026
Merged

DefTruth merged 3 commits into
mainfrom
dev

Conversation

@DefTruth

Copy link
Copy Markdown
Member

Summary

  • Completely remove DBPrune: configs, context/manager, PrunedBlocks execution, public exports, cache-type aliases, YAML loading, stats/summary, dedicated tests, benchmark config, docs, and 14 dedicated image assets.
  • Unify calibrator selection and options on --calibrator and --calibrator-params.
  • Remove legacy selector flags (--taylorseer, --dmd, --foca) and algorithm-specific CLI flags/aliases (--taylorseer-order, -order/--order, --dmd-history, --dmd-svd-precision/--dmd-svd).
  • Update benchmark scripts, examples, user docs, and agent integration guidance. Python/YAML dataclass fields remain supported.
  • Preserve the historical perf.py TaylorSeer default of order=2; reject ambiguous --cache-config + CLI calibrator combinations.

Breaking changes

  • DBPrune symbols and configs are no longer supported; no compatibility stubs are retained.
  • Removed CLI flags now fail in argparse instead of being silently ignored.

Scope notes

  • docs/papers/ is intentionally unchanged in this task, so docs/assets/dbprune.png remains for its existing reference.
  • This dev branch also contains the previously committed 2026 cache-algorithm research documentation (1b30e998, related to [RFC] 2026 DiT Cache Algorithm Integration Research #1363).

FLUX E2E performance and accuracy

Environment: FLUX.1-dev, 1024x1024, 28 steps, bf16, seed 0, one RTX PRO 5000 Blackwell. Timing is the single-run formal Inference Time, excluding warmup/compile. PSNR/SSIM use the no-cache baseline as reference.

Configuration Cache Steps Inference Time Speedup vs. baseline PSNR SSIM
Baseline — 16.91s 1.00x Ref. Ref.
Compile — 14.80s 1.14x 27.81842 0.97303
Compile + DBCache 12/28 8.66s 1.95x 28.09625 0.96119
Compile + DBCache + TaylorSeer 12/28 8.66s 1.95x 28.98003 0.97323

The unified CLI was also exercised with --calibrator taylorseer --calibrator-params order=2: E2E succeeded, logged T1O2, hit 12/28 cache steps, and completed inference in 8.68s.

Validation

  • Main calibrator + DBCache refresh regression: 131 passed.
  • Config loading + calibrator reset + multi-transformer regression: 13 passed.
  • Rejection coverage for all removed selector/value flags and aliases.
  • Three benchmark parsers verified to resolve --calibrator-params order=2; perf.py default order=2 and conflict rejection verified.
  • compileall, shell syntax checks, git diff --check, IDE diagnostics, and all pre-commit hooks passed.
  • Code review completed; no blocker/major findings remained after fixes.
  • git merge-tree against origin/main: clean.

Downstream compatibility (static audit only)

Text and AST scans found no DBPrune usage in sglang, vllm-omni, xDiT, or TensorRT-LLM. vllm-omni and TensorRT-LLM summary call sites do not consume removed prune stats. No downstream repositories were modified and no downstream test suites were run.

Closes #1365.

@DefTruth
DefTruth merged commit a98ba73 into main Nov 19, 2026
4 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.

Remove DBPrune and unify calibrator CLI configuration

1 participant