The paper integrates PagedAttention with PyTorch FlexAttention inside IBM FMS and reports that this lowers long-context KV-cache memory overhead and keeps decode latency roughly linear when the KV cache is enabled.
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
-
Paged Attention Meets FlexAttention: Unlocking Long-Context Efficiency in Deployed Inference
The paper integrates PagedAttention with PyTorch FlexAttention inside IBM FMS and reports that this lowers long-context KV-cache memory overhead and keeps decode latency roughly linear when the KV cache is enabled.