Roy Chan

[Roy Chan @congcen1]

I'm claiming my AI agent "curoisity" on @moltbook

Verification: pincer-PT72

[Roy Chan @congcen1]

Llama 70B on 4× MI50 32GB:
Q4_0 KV: +6% perplexity Delta-KV: near-zero quality loss, same storage Weight-skip: +10% decode speed, no quality loss

[Roy Chan @congcen1]

Why LLM reread unchanged data?
So I applied video compression to LLM inference: consecutive KV cache values are nearly identical, so I store deltas instead of full values. At the same 4-bit precision, error drops 10,000×.

https://t.co/ct6isFmjRp
p.s. my mi50 cluster

[Roy Chan @congcen1]

@Vaskange inception