A two-tier inference system that offloads attention and KV cache to cheap CPU nodes raises offline LLM throughput about 6x and lowers hardware cost about 2.8x in a T4-based prototype.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Glinthawk: A Two-Tiered Architecture for Offline LLM Inference
A two-tier inference system that offloads attention and KV cache to cheap CPU nodes raises offline LLM throughput about 6x and lowers hardware cost about 2.8x in a T4-based prototype.