JuicyTalk
Nunchux AI has released VC-Attention, a training-free low-bit attention kernel built for video Diffusion Transformers (DiTs). It targets 2 problems at once: value quantization error and a slow softmax stage.
Why Attention is the Video Bottleneck
Video DiTs flatten a clip into 1 sequence of spatiotemporal tokens and run full self-attention at every layer. A 5-second 720p Wan2.2-14B clip spans about 70K tokens. On the RTX 5090, attention takes more than 64% of generation time. The research team states that attention is about two thirds of every MiniMax-H3 denoising step on a single B200.
Low-bit Tensor Cores speed up the 2 matrix products, QK and PV. 2 obstacles remain. First, prior methods like SageAttention2 smooth queries and keys. After QK smoothing and rotation, the value term accounts for 82% of output error on Wan2.2. Second, the softmax between the products still runs in FP32. On B200 and H200, that exponential and its FP8 cast become the longest pipeline stage.
V-Smooth: Fixing Value Outliers
Value outliers sit in a few tokens, and their channels shift across heads, layers, and steps. A Hadamard rotation preserves token norms, so it does not remove them. Rotating V changes value error by just 0.2%.
V-Smooth takes a different route:
- Group: An online k-means clusters value tokens per batch and head. Keys and values are permuted together, so non-causal attention output is unchanged.
- Demean: Each 128-token hardware block subtracts its mean. Only the residual is quantized, using per-channel E4M3 at 8 bits or NVFP4 at 4 bits.
- Restore: The mean is added back using the row sum online softmax already keeps. No second pass or extra buffer is needed.
Averaged over 100 Wan2.2 heads, the block mean removes 8% of block energy in sequence order. It removes 12% under DeltaQuant’s static cube and 36% after sorting. Each mean costs 0.125 bit per value element.
Grouping runs only on the first 25% of denoising steps. The permutation is reused across 4 adjacent steps. Averaged over the full schedule, grouping costs 3 to 4% of attention time.
ExpCast-FP8: Removing the Softmax Bottleneck
An E4M3 byte is already close to a logarithm of the value it stores. Read as an integer, it equals roughly 8 log2(v) + 56. So ExpCast-FP8 writes the byte directly from the log-domain score with 1 fused multiply-add. The constant β = -0.35 centers the leftover error, and no constant is fitted per model.
The direct path writes the same byte as the FP32 exponent-then-cast path on 79.6% of each doubling. Elsewhere it lands 1 code away. The paper proves a per-row total variation bound under 3.64%, plus any underflow tail. Across 204.8K Wan2.2 attention rows, the measured average is 1.6%. ExpCast-FP8 applies only to the 8-bit kernel, since NVFP4 has no single affine log-to-code map.
Hand-written CuTe/CUDA fusion of the preprocessing chain cuts 1 V-Smooth call from 42.2 ms to 4.8 ms on B200.
