eagle3: extend tree to top-3 siblings — speedup_e2e = 1.20×
Widen the tree from 2 siblings to 3 at slot 0 (+ chain from top-1):
[pending_prev, d0_top1, d0_top2, d0_top3, d1_chain]
positions: [P, P+1, P+1, P+1, P+2]
5×5 tree mask enforcing sibling isolation.
50 prompts × 64 tokens on dash5:
acceptance_rate = 12.1% (4 candidates/round)
target_steps = 2101 (vs 2231 top-2, 2432 non-tree)
spec_tpot_ms = 10.43 ms
baseline_tpot_ms = 12.54 ms
speedup_e2e = 1.20× (vs 1.17× top-2, 1.10× non-tree)
Verify cost at batch=5: ~1.12× single decode (nearly free). The extra
sibling adds ~3% additional rounds where EAGLE's top-3 matches target.
Gahow Wang
2fe903ecea