---
license: apache-2.0
pipeline_tag: text-to-video
tags: [video-generation, diffusion, quantization, w4a8, w4a4, svdquant, mxfp4, wan]
base_model:
  - Wan-AI/Wan2.2-TI2V-5B-Diffusers
  - Wan-AI/Wan2.2-T2V-A14B-Diffusers
---

# Wan2.2 W4A8 SVDQuant — NVIDIA (CUDA) artifacts

Calibrated **W4A8 SVDQuant** checkpoints (MXFP4 weights, dynamic MXFP8
activations, rank-32 low-rank correction) for Wan2.2, exported with
[vLLM-Omni](https://github.com/vllm-project/vllm-omni) (PR #6527 head + CUDA
torch provider) and validated end-to-end on NVIDIA B300 (sm_103, CUDA 13.3).

The offline export is produced without Quark/ROCm by
`examples/quantization/export_quark_svdquant_w4a8_cuda.py` — a pure-torch
replication of Quark's SVDQuant recipe (SmoothQuant, alpha=0.5, on real
pipeline activations; exact `torch.linalg.svd` on device; the `1/s` factor
folded into the residual and `proj_down`, so the stored checkpoint is
self-contained).

## Contents

| Path | What it is | Size |
| --- | --- | --- |
| `wan5b-calibrated/` | Wan2.2-**TI2V-5B** quantized `transformer` + `model_index.json` + README | ~9.5 GB |
| `wana14b-calibrated/` | Wan2.2-**T2V-A14B** quantized `transformer` + `transformer_2` (dual expert) + `model_index.json` + README | ~54 GB |
| `wan5b_calibrated_w4a8_svd.mp4` | 5B calibrated generation (480x832, 41f, 20 steps) | 84 KB |
| `wana14b_calibrated_w4a8_svd.mp4` | A14B calibrated generation (720x1280, 81f, 40 steps) | 817 KB |
| `wan5b_svdquant_calibrated_similarity.json` | 5B vs BF16 trajectory-similarity metrics | — |
| `wan_a14b_svdquant_calibrated_similarity.json` | A14B vs BF16 trajectory-similarity metrics | — |

Each model's README covers the one-line setup (stock `vae` / `text_encoder` /
`tokenizer` / `scheduler` are the base-model components and are not duplicated
here) and the exact vLLM-Omni commands.

## Quick start (5B example)

```bash
hf download Yi30/wan2.2-ti2v-w4a8-svd-nvidia \
  --include "wan5b-calibrated/*" --local-dir ./wan5b-calibrated
hf download Wan-AI/Wan2.2-TI2V-5B-Diffusers \
  --include "vae/*" "text_encoder/*" "tokenizer/*" "scheduler/*" \
  --local-dir ./wan5b-calibrated

# vLLM-Omni (vllm 0.26.0 pairing; torch provider auto-selected on CUDA)
python examples/offline_inference/text_to_video/text_to_video.py \
  --model ./wan5b-calibrated --quantization quark_svdquant \
  --prompt "A serene lakeside sunrise with mist over the water." \
  --height 480 --width 832 --num-frames 41 --num-inference-steps 20 --output out.mp4
```

The embedded `quantization_config` in each transformer's `config.json`
(`is_checkpoint_w4a8_serialized: true`) makes the loader take the offline
serialized path automatically. Experimental W4A4 (MXFP4 activations):
`VLLM_OMNI_W4A8_CUDA_ACT_FMT=mxfp4`.

## Accuracy (seed 42, vs BF16 pipeline, same settings as online runs)

| Model | Setting | Tier | PSNR (dB) | Cosine | Peak memory |
| --- | --- | --- | --- | --- | --- |
| TI2V-5B | 480x832, 41f, 20 steps | online SVD (baseline) | 27.10 | 0.9979 | 25.1 GB |
| TI2V-5B | 480x832, 41f, 20 steps | **calibrated (this repo)** | 26.52 | 0.9975 | 26.3 GB |
| T2V-A14B | 720x1280, 81f, 40 steps, gs 4.0/3.0 | online SVD (baseline) | 12.59 | 0.794 | 55.6 GB |
| T2V-A14B | 720x1280, 81f, 40 steps, gs 4.0/3.0 | **calibrated (this repo)** | 12.71 | 0.804 | 56.4 GB |

On the MXFP4/MXFP8 torch backend the calibrated tier is accuracy-*equivalent*
to the online tier (SmoothQuant's benefit targets per-channel quantizers;
MXFP4's per-32-element group scale is insensitive to the column rescaling —
measured residual Q4 error: 11.63% online vs 11.64% calibrated across all 300
5B target layers). The calibrated checkpoints are the portable, deterministic
artifact: no `svd_lowrank` at load, factors pre-computed. On-disk format is
BF16 residual + BF16 factors (packed to MXFP4 at load); the exporter also
supports `--pack-format packed|unshuffled` for ~4x smaller MXFP4-on-disk
variants.