TECHNOLOGY
Reduce memory pressure without hiding correctness, latency, quality, or fallback costs.
MEMORY ORCHESTRATION
A paged-attention-inspired layer designed to manage KV blocks dynamically and preserve useful context under constrained GPU memory.
PREDICTIVE TRANSFER
Uses normalized attention entropy to decide when prefetch may be useful while a separate selector decides which pages should move.
LAYER-AWARE NUMERICS
Assigns lower precision to shallow layers and higher precision deeper in the model.
LOW-BIT COMPUTE
A signed INT4 and MXFP4-style dequantization path intended to reduce bandwidth pressure.
CANDIDATE REGISTRY
Matched baseline control.
Measured slower, no physical VRAM win.
Custom low-bit KV backend.
INT2/INT4 research branch.
Weight bandwidth path.
Combined reduction path.
Primary target architecture.
SM90 / H100 branch.
SM100 / Blackwell branch.