Run Wan2.2-A14B on One RTX 5090: 720p Video in 3.8s

AI Guide

Generating a 5-second 720p video with a 14B-class video DiT used to mean waiting tens of minutes. LightX2V just changed that math: LightWan2.2-A14B compresses Wan2.2-A14B from 40 denoising steps to 4, folds NVFP4 quantization-aware training into the distillation, and adds dynamic sparse attention — so the full 14B MoE model runs on a single RTX 5090. Push it to 8 cards and a 5-second 720p clip renders in 3.8s (T2V), a 705.8× speedup over the 40-step baseline.

Why this matters

Wan2.2-A14B has two hard problems on consumer hardware: it's slow (40 denoising steps, and at ~120K-token sequences self-attention alone can eat 80%+ of a single DiT pass) and it doesn't fit (weights + long video tokens + attention buffers blow past the ~30GB ceiling of consumer GPUs). LightX2V attacks both: fewer steps, cheaper steps, and lower peak memory.

The 4 moves that make it fast

1. Step distillation: 40 → 4

Phased DMD splits the SNR interval into phases and trains different experts to denoise within each noise subinterval, keeping gradients for only one sampling step per stage. SGMD (Score Gradient Matching Distillation) aligns the fake score to the teacher score with a stop-gradient Fisher objective, tracking the continuously-updated generator through outer/inner loop corrections. The result is a "two high-noise experts + two low-noise experts" 4-step pipeline — the DiT backbone runs 36 fewer times while keeping motion dynamics, visual detail, and output diversity.

2. NVFP4 quantization-aware training

NVFP4 uses E2M1 4-bit values with FP8 E4M3 block scaling per 16 elements plus FP32 global scaling — near-FP4 storage cost with better dynamic range. The student model is trained to tolerate 4-bit error during distillation, then mapped to Blackwell Tensor Cores via CUTLASS kernels: dual-NVFP4 block-scaled GEMM with FP32 accumulation, BF16 output, and fused column bias in the epilogue. Linear-layer memory traffic and GEMM latency both drop.

3. Dynamic sparse attention

Instead of computing full attention over ~120K tokens, the runtime estimates block importance from Q/K block-level means and keeps only the top 10–20% of blocks (80–90% sparsity), computing densely inside the survivors to preserve GPU throughput. It plugs into SageAttention2 for an FP8 sparse attention path, cutting both dense matmul and attention costs.

4. Fused kernels, async offload, and multi-GPU parallelism

3D RoPE and RMSNorm get fused/parallel kernels (fewer Python dispatches, fewer temp tensors). Block-level offload moves weights per Transformer block and overlaps transfer with compute, so a 14B model plus long video sequences fits in one card's VRAM. For 8 cards, Light-Ulysses sequence parallelism turns Ulysses into a real-time pipeline: FP8 communication, QKV tensors fused into a single All-to-All payload, Triton fused pre/post-processing kernels, and head-parallel async All-to-All that transmits the next heads while current heads compute.

Quick start

# Official image (pinned deps for 5090)
docker pull lightx2v/lightx2v:26052801-cu130-5090

# Single GPU: T2V / I2V
bash scripts/wan22/extreme/run_wan22_moe_t2v_extreme.sh
bash scripts/wan22/extreme/run_wan22_moe_i2v_extreme.sh

# Multi-GPU sequence parallel (set seq_p_size = GPU count)
bash scripts/wan22/extreme/run_wan22_moe_t2v_extreme_sp_parallel.sh
bash scripts/wan22/extreme/run_wan22_moe_i2v_extreme_sp_parallel.sh

Manual install: build from source and compile the NVFP4 kernels against NVIDIA CUTLASS.

What you can expect

  • 8×RTX 5090: T2V 720p → 3.8s (705.8×), I2V 720p → 4.5s (599.3×) — both under the 5s real-time bar.
  • 1×RTX 5090: T2V 720p → 22.5s (118.7×), I2V 720p → 26.7s (100.5×), I2V 480p → 10.7s.
  • vs TurboWan2.2 (same 4-step, single card): ~3.5× faster at 480p and ~2.4× at 720p for I2V.

Practice tips

  • Start from the official Docker image — the pinned environment is the difference between "runs" and "compiles for three hours".
  • Single card: block-level async offload is what makes the 14B model fit; don't disable it to chase a few percent of speed.
  • Multi-GPU: set seq_p_size to your actual GPU count; the scripts ship with the SP config enabled.
  • Want the real-time bar? 8 cards get 5s of video in under 5s. One card is interactive-grade for drafts and iteration.

Resources

Scroll to top