Skip to content

fix(sa3): use the monotonic audio-to-audio schedule - #320

Draft
ryanontheinside wants to merge 1 commit into
ryanontheinside/chore/sa3-current-upstream-pinfrom
ryanontheinside/fix/sa3-monotonic-a2a-schedule
Draft

fix(sa3): use the monotonic audio-to-audio schedule#320
ryanontheinside wants to merge 1 commit into
ryanontheinside/chore/sa3-current-upstream-pinfrom
ryanontheinside/fix/sa3-monotonic-a2a-schedule

Conversation

@ryanontheinside

Copy link
Copy Markdown
Collaborator

Outcome

Advance DEMON's managed SA3 pin to the monotonic audio-to-audio schedule branch being reviewed upstream in Stability-AI/stable-audio-3#94.

This is intentionally a one-line pin change. The sampler correction, regression tests, and CJ Carr attribution live in the upstream branch rather than in a DEMON-local shim.

Pinned combined test commit: 960da1f8cbe205ab3b702edbfabd91113ab22473.

That fork commit contains:

  1. current upstream main
  2. the pending TensorRT plugin correction from Fix diff_attn_swa plugin int attribute under TRT 10.16 / NumPy 2.x Stability-AI/stable-audio-3#49
  3. the schedule correction from Fix shifted audio-to-audio schedules across backends Stability-AI/stable-audio-3#94

Why it is stacked

Depends on DEMON #319, which establishes the current-upstream pin and versioned SAME-L engine boundary. Keeping this separate makes the source refresh independently reviewable and prevents the sampler fix from being confused with engine migration work.

Engine impact

No engine rebuild is required after #319. The schedule is Python/host-side and this pin retains the same explicit SAME-L plugin revision, so it discovers and reuses:

same_l_decode_window_aot_mma_vc07698548567_t32_56_96

Existing DiT engines are also unchanged.

Verification

Against the combined fork branch:

52 DEMON SA3 tests passed
5 upstream schedule tests passed

Representative schedules at sigma_max=0.1, 0.25, 0.5, 0.75, 1.0 were checked for:

  • exact requested first point
  • monotonic decrease
  • no value above the entry sigma
  • zero endpoint

At sigma_max=0.5, the current schedule begins [0.50000, 0.70562, ...]; the corrected schedule begins [0.50000, 0.48849, ...].

Merge instructions

  1. Merge DEMON chore(sa3): repin current upstream and version SAME-L engines #319 first and complete its SAME-L engine rollout.
  2. Keep this PR draft while upstream rtmg/web: four UX fixes — handle, ref upload, midi reconnect, network pill #94 is under review.
  3. If upstream rtmg/web: four UX fixes — handle, ref upload, midi reconnect, network pill #94 merges, refresh the fork pin to the upstream merge commit before merging this PR.
  4. If DEMON needs extended production testing first, this fork pin is reproducible and can be deployed without another decoder build.
  5. After the parent merges, retarget/rebase this PR onto main; its final diff should remain only the SA3 SHA change.

Not included

The nonlinear product remapping of the denoise knob is deliberately excluded and lives in a separate dependent PR.

Advance the managed fork pin to the current-main schedule correction,
stacked on the TensorRT plugin compatibility commit. The SAME-L plugin
revision is unchanged, so deployments reuse the decoder built by the
parent pin PR rather than rebuilding it for this Python-only change.

The schedule implementation is submitted upstream separately with
credit to CJ Carr's original sa3-a2a-monotonic-schedule branch.
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.

1 participant