n0p/mlx-video
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

More poodles

Browse Source
This commit is contained in:
Daniel
2026-03-11 08:14:12 +01:00
parent d207275fea
commit 1cf878f5e0
9 changed files with 347 additions and 23 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
mlx_video/models/wan/docs/DIAGNOSTICS.md
View File

@@ -315,11 +315,6 @@ Applied alongside bug fixes to improve inference speed:
- **Redundant type cast removal**: MLX type promotion handles `bfloat16 * float32 → float32` automatically — removed 240 unnecessary graph nodes per step (6 casts × 40 blocks)
- **Euler scheduler sync fix**: Pre-store sigmas as Python floats to avoid `.item()` evaluation sync
### TeaCache Integration
- Polynomial rescaling stays in MLX lazy graph (Horner's method)
- Single `.item()` call on the accumulated distance for the skip/compute decision
- Configurable threshold, retention steps, and cutoff steps
---
## Resolved: CFG Effectiveness (was Open Investigation)