README floors
| Model | Official single GPU | Flags called out | Multi-GPU |
|---|---|---|---|
| T2V-A14B / I2V-A14B / S2V-14B | ≥ 80 GB | --offload_model True --convert_model_dtype [--t5_cpu] | FSDP + Ulysses, FlashAttention3 on Hopper |
| TI2V-5B | ≥ 24 GB (4090) | offload + convert + --t5_cpu | Same FSDP path; drop offload on 80 GB to go faster |
Efficiency-table notes: prompt extend off; 14B multi-GPU uses --ulysses_size 4/8 --dit_fsdp --t5_fsdp; averages after warmup. Read the PNG for the per-SKU numbers rather than quoting a blog’s “it runs on 8 GB.”
Knobs that actually move the needle
- --offload_model True — page DiT blocks to CPU between steps.
- --t5_cpu / --t5_fsdp — UMT5-XXL is huge; do not leave it on a 24 GB card next to the 5B DiT.
- --convert_model_dtype — param_dtype cast, not 4-bit.
- Resolution and frame count — sequence length is the activation killer.
- CFG — a second forward unless the graph fuses it.
- FP8 / GGUF / distill — see the quantization pages when 80 GB is fantasy.
FAQ
- Will 24 GB run T2V-A14B?
- Not with the official generate.py recipe. Use multi-GPU, or a community FP8/GGUF/distill pack, and expect quality/speed tradeoffs.
- Why 1280×704 not 1280×720 on 5B?
- The 5B 720P grid is 1280×704 / 704×1280 so it aligns with the high-compression VAE patches.
Primary sources
Checked against Wan-Video/Wan2.2 README · Aug 28, 2026
