Baidu's Virtual Head Attention tames the KV cache, FLOPs, quality triangle in long-context LLMs
baidu · hf · 2026-10-16
A Baidu and Peking University paper proposes Virtual Head Attention (VHA), addressing the "impossible triangle" of KV cache size, training/prefill FLOPs, and quality in long-context attention. VHA uses two lightweight linear ops — Q Premix, a near-identity query transform that recovers virtual head diversity from halved physical heads, and Linear Postmix, a low-rank I+AB^T residual fusion — improving all three at once with negligible inference overhead since the ops fold into existing projections.
More from Infra
- GitHub rewrote the 800k-line Copilot runtime in Rust using AI agents, cutting CPU and memory by over 90% — bibryam · 2026-10-16
- NVIDIA quietly delists the entry-level RTX 5050 from its marketplace, following the RTX 5090 — chemist_slime · 2026-10-16
- Grok claims 816B MoE model runs locally on dual M5 Ultra Mac Studios via MLX — grok · 2026-10-16
- Same 1,000 Cloudflare Durable Objects: $4,050/month cut to $22 with one trick — threepointone · 2026-10-16
- NVIDIA ships FP4-quantized Qwen3.8-Flash for local use, trending on Hugging Face — nvidia · 2026-10-16
- DuoDiscover agents auto-build fast eval proxies, scoring 2.68x TPU kernel speedup — berkeley_ai · 2026-10-16