Stack additional acceleration techniques on top of LocalLeap (SlowFast, dInfer, ES-dLLM)

[iOptimizeThings]

Background

v0.1 ships LocalLeap (arXiv:2510.07081) as the primary diffusion acceleration. The 2025–2026 literature has produced several follow-up techniques that target different inefficiencies in the denoising loop. Most compose orthogonally with LocalLeap and Fast-dLLM KV caching (#4).

Candidates in priority order:

1. SlowFast Sampling - adaptive step scheduling based on confidence variance. Skips low-information "slow" steps when the canvas is locally settled, takes more "fast" steps in uncertain regions. Reported up to 15.63× speedup on LLaDA standalone, 34.22× when combined with dLLM-Cache on GPQA.

2. dInfer - full inference framework with several optimizations bundled. Achieves 1,100 TPS at batch=1 on 8× H800. Worth porting individual components.

3. ES-dLLM (Early Skipping) - skips later denoising steps once the model's confidence on the active block exceeds a threshold. Reports 226 TPS on H200 for LLaDA-8B.

Acceptance criteria

• One PR per technique (do not bundle)
• License audit before implementation (per CLAUDE.md rule)
• Inline citation comment + CREDITS.md entry before code
• All 4 LocalLeap quality gates pass with each new technique enabled (quality re-verification required - speedup gate alone is not sufficient)
• Per-technique opt-in flag in SamplingParams
• Per-model calibration if the technique has thresholds (analogous to _LOCAL_LEAP_MODEL_DEFAULTS)

Out of scope for this issue

This is a meta-issue tracking the research direction. Each technique gets a dedicated implementation PR.

[johndpope]

Nice nice.

I’m working on open-dllm
But am doing upgrades - found the d3llm
And worked on trajectory distillation of 27B model
That’s fine
https://github.com/scrya-com/dllm-castlehill

Need some help with compute

[iOptimizeThings]

Hey @johndpope! checked out dllm-castlehill, looks cool. hitting 400 tok/s on a 27B using a 5090 is exactly the kind of question I want answered too. Following it, will tinker when I get some breathing room.

On compute, I'm on a single RTX 5070 12GB at the moment so nothing useful to share for 27B training runs, sorry :/ Lambda/CoreWeave is probably your best bet there.

If your distilled checkpoints come out as masked-discrete models, dlmserve should be able to serve them :D

cheers