[diffusion] feat: opt-in uint8 rollout video via rollout_video_dtype - #3
Draft
Rockdu wants to merge 1 commit into
Draft
[diffusion] feat: opt-in uint8 rollout video via rollout_video_dtype#3Rockdu wants to merge 1 commit into
Rockdu wants to merge 1 commit into
Conversation
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
What
rollout_video_dtypeonRolloutRequest:"keep"(default, unchanged) returns the decoded video as-is;"uint8"quantises it engine-side before serialisation, with the same formula the training-side reward path applies, so scores are unaffected.Draft: stacked on
feat/rollout-timing. Counterpart of radixark/miles_diffusion#198's--rollout-video-dtype; an older client that never sends the field gets today's behaviour.Why
H3's rollout response is 1384 MiB of raw float video per sample, of which the consumer keeps 8 of 107 frames after a uint8 cast — 177 GiB moved per 128-sample rollout. Quantising at the source cuts the body to 436 MiB (3.2x). Opt-in because some consumers legitimately need the unquantised tensor.
Validation
resp_bytes1384 -> 436 MiB, 0 non-200 over the run,log_prob_mean_abs_diff3.0e-05 (training numerics untouched — they ride the latents, not the video).Checklist
CI States
Latest PR Test (Base): ❌ Run #32523050643
Latest PR Test (Extra): ❌ Run #32523050373