lingbot: add 1-step speed presets, 2.2x speed up inspired by PDD - #434
Open
wenqingw-nv wants to merge 1 commit into
Open
lingbot: add 1-step speed presets, 2.2x speed up inspired by PDD#434wenqingw-nv wants to merge 1 commit into
wenqingw-nv wants to merge 1 commit into
Conversation
…riant) The distilled schedule is front-loaded (all four solver steps sit at sigma >= 0.83; at LoRA-identity the first step's flow matches the 4-step integrated mean velocity to ~97%), so a single step reaches quality parity. Measured (GB300, 18 held-out cells, 40 chunks): MUSIQ 58.9 vs 58.8 (4-step), drift Delta +8.01 vs +8.27, per-chunk 1.80 s -> 0.81 s (2.2x); interactive taehv preset ~1.2 s -> ~0.45 s (~2.6x, ~30 fps effective at 464x832). Known issue: dynamic degree -18% vs 4-step (53.8 vs 65.6 RAFT) -- disclosed in the README row. ci_cpu test asserts the 1-step schedule equals the 4-step schedule's first step for both presets. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Signed-off-by: Wenqing Wang <wenqingw@nvidia.com>
Contributor
Greptile SummaryThe PR adds two optional one-step LingBot-World runner presets while retaining the existing four-step defaults.
Confidence Score: 5/5The PR appears safe to merge. No blocking failure remains. Important Files Changed
Flowchart%%{init: {'theme': 'neutral'}}%%
flowchart LR
User[Runner slug] --> EntryPoint[Package entry point]
EntryPoint --> Runner[RunnerConfig]
Runner --> Pipeline[Derived LingBot pipeline]
Pipeline --> Scheduler[FlowMatch scheduler]
Scheduler --> Step[Single denoising step at first four-step schedule point]
Reviews (2): Last reviewed commit: "lingbot: add 1-step speed presets (lingb..." | Re-trigger Greptile |
Collaborator
Author
|
/ok to test 589dedb |
wenqingw-nv
force-pushed
the
lingbot-1step-presets
branch
from
August 11, 2026 19:11
589dedb to
b992d3d
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Adds single-solver-step speed presets for the LingBot-World distilled checkpoints. The distilled schedule is front-loaded (all four solver steps sit at sigma >= 0.83; at identity the first step's flow matches the 4-step integrated mean velocity to ~97%), so a single step reaches quality parity at ~2.2-2.6x lower chunk latency. Found while evaluating PDD-style mean-velocity distillation (Parallel Decoding Distillation, arXiv 2607.26004) on this host: the first step's flow already captures ~97% of the 4-step trajectory's mean velocity, so the speedup needs no distillation — a config-only preset suffices.
Contents
lingbot-world-fast-1step— 1-step preset on the Wan-VAE decoder slug.lingbot-world-fast-taehv-window15-sink3-1step— 1-step mode on the LightTAE window=15 + sink=3 interactive preset.derive_configscheduler override:num_inference_steps=1,denoising_timesteps=[1000]); registered entry points; README slug-table rows.(t, sigma)for both presets.Measured (464x832, 40-chunk ≈30 s held-out rollouts, GB300; 18 cells)
lingbot-world-fast)Interactive preset (taehv window15/sink3, 21-block runs): ~1.2 s -> ~0.45 s per chunk (~2.6x, ~30 fps effective at 464x832).
Deploy notes