On: Bonsai 2 27B: Near-Lossless Compression in a 9x Smaller Footprint

[Dwedit]

Even with a small context size (4096 at 64KB per token), that's like 256MB for the context. It's more than just the context that's eating up VRAM.