daniel/h3-blackwell-runtime
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Record seed-exact initialization result

Browse source
This commit is contained in:
Daniel Maddern 2026-08-13 16:18:33 +07:00
parent 09f5d2ef65
commit 58791dbd61
1 changed files with 5 additions and 4 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

9
PARITY.md
View file

@ -189,7 +189,7 @@ gate and is exact.
| Beta/RES sampler | Yes | H3 replay is bit-exact for all 12 captured sampler calls and RES update is bit-exact from captured denoised; direct raw-output to model-wrapper denoised conversion retains small BF16/state boundary deltas |
| Video VAE decoder | Yes | Same final sampled latent decoded through direct VAE versus Comfy PNG frames: mean abs pixel delta `0.0833898`, max `6` |
| Audio VAE/decode/mux | No | Preview intentionally produces video only |
| End-to-end prompt-only FL2VA preview | Yes | Direct 320x192x22 dragon smoke video generated with Sage2 at `/home/daniel/StoryStudioAssets/H3-output/h3-blackwell-runtime/direct-fl2va-dragon-e2e-55e2073.mp4`; seed-exact Comfy initial AV state parity is not implemented |
| End-to-end prompt-only FL2VA preview | Yes | Direct 320x192x22 dragon smoke videos generated with Sage2. Comfy-equivalent CPU joint AV seed initialization now matches captured `initial.pt` exactly; seed-exact direct MP4 versus Comfy PNG frames still has mean abs pixel delta `14.0714`, max `193`, consistent with the remaining model-wrapper denoised conversion boundary |
| Full Ref2VA | No | References, vision conditioning, VAE encode, audio, and muxing are unimplemented |
## Remaining Gates, In Dependency Order
@ -199,9 +199,10 @@ unless the checkpoint, Comfy version, prompt, or backend changes.
1. **Final video latent layout.** Compare direct `unpatchify_video` output
with Comfy before entering the sampler.
2. **Seed-exact prompt-only initialization.** Replace direct preview's local
`torch.manual_seed` video-only noise setup with Comfy-equivalent joint AV
initial latent generation before claiming identical seed output.
2. **Model-wrapper denoised conversion.** Direct H3 and RES update gates pass,
and direct initial AV noise now matches Comfy exactly. Close the remaining
small `h3_output -> denoised` conversion delta before claiming identical
seed output.
3. **Feature/performance work.** Only then add audio, Ref2VA/reference paths,
Sage3, CUDA graphs, and multi-GPU execution.