SOL Attention + SageAttention Speeds Up H3 Video Gen Up to 2.5x on Apple Silicon
TgoAI · reddit · 2026-09-10
A developer added SOL Attention (dynamic block-sparse attention) and a SageAttention-style INT8 QK path to Vpipe on Apple Silicon. On an M5 Pro 24GB, 6-step DiT: at 832×480, vanilla H3 takes 17 min vs SOL's 9.3 min; at 1344×768 both accelerated paths hit 30 min vs 74 min vanilla (2.5x). Unlike VDN's projection+linear-attention replacement, SOL keeps exact attention for selected blocks and produced no artifacts (e.g. reversed water flow) seen with VDN. The author reimplemented SOL kernels natively in Metal rather than reusing MPS, and the shared backend benefits other image/video models. Open source: github.com/tgo-app-dev/vpipe.
Related event: SOL Attention plus SageAttention speeds Apple Silicon inference 2.5x(2 posts)→
More from Infra
- E2E Networks to add ~5.4MW for NVIDIA HGX B300, lifting capacity to 14.6MW — cneuralnetwork · 2026-10-26
- INT21's AI agent swarm rewrote Drama 3's inference runtime in Rust within one hour, beating SGLang-Omni — bingxu_ · 2026-10-26
- Anthropic's S-1 reveals $434.5B in contracted compute, with SpaceX the largest counterparty — _AustinCalvert_ · 2026-10-26
- Everyone ships AI sandboxes, yet Modal and Daytona still top growth charts — FinanceYF5 · 2026-10-26
- Ant S-1 validates 'compute is fungible' thesis: single pool, dynamic allocation — BenBajarin · 2026-10-26
- SLC NAND for AI may need 2% more wafer capacity per 1% share shift, echoing HBM — toptickcrypto · 2026-10-26