Single-rank demo script that loads
nvidia/AnyFlow-Wan2.1-T2V-14B-Diffusers into FastVideo's
WanTransformer3DModel via param_names_mapping, then samples 81 frames
at 480x832 with the new FlowMapEulerDiscreteScheduler at both NFE=4
(~30s) and NFE=50 (~5min) and decodes each via the Wan VAE (tiling
enabled). Uses guidance_scale=1.0 since the on-policy distilled
checkpoint has fuse_guidance_scale=3.0 baked into the weights, matching
the AnyFlow paper's official demo.py default.
Memory tactics for single H200 (141 GB HBM):
- Encode prompts with UMT5 first, free the text encoder.
- Build 14B transformer (~28 GB bf16), load AnyFlow shards.
- Sample at both NFEs, free transformer, then VAE decode with tiling.
Peak GPU usage ~57 GB on a single H200.
Enderfga
bbb1f9781b