VDN-H3 ported to Vpipe hits ~2.6x faster H3 generation on M5 Pro 24GB
TgoAI · reddit · 2026-09-07
The author ported VDN-H3 (Video DeltaNet for MiniMax H3) to Vpipe, a native C++20/Metal inference runtime for Apple Silicon (no PyTorch/MPS/MLX), in about a day.
Results:
- Speedup scales with length and resolution: at 832×480 it pulls ahead around 6s reaching 1.7× at 15s; at 1344×768 the crossover comes earlier with 2.6× at 14s (6 DiT steps + Turbo LoRA)
- Reference points: FastH3 needs 465–504s on M4 Max 36GB for the same spec; Vpipe's baseline H3 without VDN cold-starts at 469s on a 16GB M5 MacBook Air
Approach & costs: VDN exploits inherent redundancy in video diffusion rather than applying compute uniformly, using a hybrid frame-wise linear-attention + softmax-attention architecture. Tradeoff: 4GB of non-fusable extra weights plus 1GB potentially fusable LoRA-style (not yet implemented); on 24GB machines this can outweigh compute savings for small workloads. 8-bit quantization of incremental weights is a future optimization.
Quality: compositions closely match baseline with the same seed, but some sequences show long-range temporal inconsistencies (e.g., water flow reversing), likely tied to sliding-window/linear attention; mixing in global attention may help. Reference conditioning (first/last-frame) remains open.
More from Infra
- Anthropic generated $2.08 revenue per $1 of compute spend in Q2 2026, up from $0.63 — cedric_chee · 2026-10-26
- CircuitHub unveils The Grid, a factory-scale robot making PCBs in the US twice as fast as Shenzhen — garrytan · 2026-10-26
- Ubicloud publishes "Fixing Tool-Call Accuracy in vLLM" — petrusenko_max · 2026-10-26
- Meta approves nonvoting Class C shares as 2026 capex guidance hits $135B–$145B — thione · 2026-10-26
- Unsloth Ships MiniMax-M3.1-Flash Quants: 2-bit Retains 91.5% Accuracy, Runs in 150GB RAM — danielhanchen · 2026-10-26
- Engineer with thousands of fabricated devices pours cold water on photonic crystal modulator yields — jwt0625 · 2026-10-26