[Perf]: CFG parallel abstraction (#851)

Signed-off-by: Didan Deng <33117903+wtomin@users.noreply.github.com>
Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>
This commit is contained in:
Didan DengandHongsheng Liu authored and GitHub committed 2026-01-30 16:23:44 +00:00
1 parent 8ad39570b8
commit 23cf48d7bb
31 files changed
+1506 -814

No files matched your search

@@ -52,6 +52,7 @@ Key arguments:
- `--num_frames`: Number of frames (default 81).
- `--guidance_scale` and `--guidance_scale_high`: CFG scale (applied to low/high-noise stages for MoE).
- `--negative_prompt`: Optional list of artifacts to suppress.
- `--cfg_parallel_size`: the number of devices to run CFG Parallel. CFG Parallel is valid only if classifier-free guidance is enabled and `cfg_parallel_size` is set to 2.
- `--boundary_ratio`: Boundary split ratio for two-stage MoE models.
- `--flow_shift`: Scheduler flow shift (5.0 for 720p, 12.0 for 480p).
- `--num_inference_steps`: Number of denoising steps (default 50).