Six stackable techniques that compound:
1. KV cache quantization (Q8_0 = 2x context, asymmetric K=Q8/V=Q4 = 2.6x)
2. Flash attention (free VRAM + speed, zero quality loss)
3. Host-memory prompt caching (--cram, use the server's 128-384GB RAM)
4. KV to system RAM (-nkvo, last resort, 5-20x slower)
5. Architecture selection (GQA + MoE = tiny KV footprint)
6. NVMe mmap for model loading (fast cold starts, not inference)
Stacked result: single card goes from ~50K to ~130K usable context
with the MoE model. Updated llama.cpp config with all flags.
https://claude.ai/code/session_01PtYTPherSJaxDEVPgF6Nxu
Claude
0f57690692