BALANCE jointly schedules users between autoregressive and speculative decoding on a single edge GPU, with a 1/2-approximation algorithm, and reports simulated throughput gains of 27 to 39 percent over either mode alone.
Attention is all you need,
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.NI 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
BALANCE: Hybrid Autoregressive-Speculative LLM Inference in Wireless Edge Networks
BALANCE jointly schedules users between autoregressive and speculative decoding on a single edge GPU, with a 1/2-approximation algorithm, and reports simulated throughput gains of 27 to 39 percent over either mode alone.