mid-to-long-context

BudgetDraft proposes a multi-view training method for speculative decoding that aligns a sparse-KV drafter with a full-KV verifier, achieving significant speedups for mid-to-long context inference.