Download lbcache/__init__.py from Cccccz/comparison: direct link, hf CLI and curl.
- Browser
- Download file 906 Bytes
-
https://huggingface.co/Cccccz/comparison/resolve/main/lbcache/__init__.py
- Command line
-
hf download hf://Cccccz/comparison/lbcache/__init__.py
-
curl -L -o __init__.py https://huggingface.co/Cccccz/comparison/resolve/main/lbcache/__init__.py
906 Bytes
| """LingBot-World-V2 1.3B causal-fast base: cache-strategy plumbing for the | |
| comparison study (the analogue of ``hycache`` for HY-WorldPlay). | |
| The DiT is ``wan.modules.model_fast.WanModelFast``: a block-causal Wan2.1-1.3B | |
| (30 blocks, 12 heads) with a rolling self-attention KV cache. Each chunk of 4 | |
| latents runs 4 denoising forwards (t = 1000/750/500/250 on the shift-5 grid, | |
| renoised between steps) followed by one untimed context pass at t = 0 that | |
| rewrites the chunk's KV slots from the clean latent -- exactly Self-Forcing's | |
| KV-refresh pass. Denoising forwards write the chunk's own KV slots, which the | |
| context pass overwrites before the next chunk reads them, so a skipped step | |
| leaves nothing stale behind (same argument as ``cachelib``). | |
| The reuse baseline here reuses the *velocity* (the DiT's flow prediction) rather | |
| than the block-stack residual: an ``R`` step does not call the DiT at all. | |
| """ | |