Topic

attention

2 posts tagged attention from Omnissiah Systems.

Attention Sink Poisoning: Why Your Long-Context Model Degrades Before the Window Fills

How attention sink tokens accumulate disproportionate weight in long-context inference, why this degrades output quality silently, and what you can do about it in production.

Magos Veridian

· August 3, 2026 · 5 min read

Flash Attention Memory Footprint: What Your GPU Actually Allocates During Prefill

A practical breakdown of how FlashAttention allocates GPU memory during prefill, where the pressure points are, and how to diagnose OOM failures before they surprise you.

Magos Veridian

· June 29, 2026 · 4 min read

Browse more topics

mlops inference gpu-infrastructure observability distributed-systems gpu latency debugging distributed inference distributed-training reliability batching

← All posts from Omnissiah Systems