Adaptive Speculative Decoding

Adaptive speculative decoding lets SGLang adjust speculative_num_steps/speculative_num_draft_tokens at runtime instead of keeping a single fixed value for the whole server lifetime. It is designed for workloads whose accept length changes over time, where one static step count is rarely optimal.

Current support

• Only --speculative-algorithm EAGLE
• Only --speculative-eagle-topk 1
• If either condition is not met, SGLang falls back to static speculative settings

Why adaptive steps help

speculative_num_steps controls how many draft-model autoregressive steps run in each speculative round. In practice, the best value depends on the current workload.

• If num_steps is too small, the draft model could have produced more accepted tokens, but the round stops too early.
• If num_steps is too large, the draft model produces many candidate tokens that the target model rejects, so extra draft work is wasted.

Real traffic often moves between high-acceptance and low-acceptance phases, so one fixed step count is usually a compromise. Adaptive mode tries to follow the workload instead of hard-coding a single global num_steps.

Design overview

The adaptive mechanism has three pieces:

• AdaptiveSpeculativeParams: the EMA-based policy
• SpecRuntimeState: the per-tier runtime state bundle
• AdaptiveController: the coordinator that chooses a tier and activates the matching runtime state

At startup, SGLang pre-builds one runtime state per candidate tier. By default, the candidate tiers are candidate_steps = [1, 3, 7].

┌──────────────────────────────────────────────────────────┐
│                      SpecRuntimeState                    │
│                                                          │
│  speculative_num_steps / speculative_num_draft_tokens   │
│                                                          │
│  ┌────────────────┐ ┌────────────────┐ ┌──────────────┐  │
│  │  Draft stage   │ │  Verify stage  │ │ Extend stage │  │
│  │                │ │                │ │              │  │
│  │  attn_backend  │ │  attn_backend  │ │ attn_backend │  │
│  │  cuda_graph    │ │  cuda_graph    │ │ cuda_graph   │  │
│  └────────────────┘ └────────────────┘ └──────────────┘  │
└──────────────────────────────────────────────────────────┘

This matters because CudaGraphRunner is shape-dependent. Each candidate tier owns its own graph and backend state, so runtime switching is a reference swap, not an online graph recapture.

Runtime flow

The adaptive update happens after verify and affects the next round, not the current one:

┌─────────────────────────────────────────────────────────────────────┐
│           EAGLEWorker.forward_batch_generation() — decode path      │
│                                                                     │
│   ① draft(batch)                                                    │
│       │  draft model multi-step generation with current tier        │
│       v                                                             │
│   ② verify(batch, spec_info)                                        │
│       │  target model tree verification                             │
│       │  → produces accept_length_per_req                           │
│       v                                                             │
│   ③ forward_draft_extend_after_decode(batch)                        │
│       │  draft model KV-cache catch-up                              │
│       v                                                             │
│   ④ adaptive_controller.on_verify_complete(accept_lengths)          │
│       │                                                             │
│       │  update EMA, apply warmup / interval / hysteresis gates     │
│       │  if tier changed, select a pre-built state from pool        │
│       v                                                             │
│     worker.apply_runtime_state(state)                               │
│                                                                     │
│   Tier switch happens after the current round completes.            │
│   Backends and CUDA graphs are never swapped mid-round.             │
└─────────────────────────────────────────────────────────────────────┘

How the policy decides

After each verify pass, SGLang reads the accepted draft length per request, computes the batch average, smooths it with an exponential moving average (EMA), and switches among the pre-built candidate tiers [1, 3, 7] by default.

The decision logic is intentionally conservative:

• warmup_batches skips the first few batches
• update_interval avoids switching every batch
• down_hysteresis and up_hysteresis reduce oscillation

Conceptually, the policy probes one step beyond the observed acceptance:

target_steps ≈ clamp(round(ema_accept_len) + 1, min(candidate_steps), max(candidate_steps))

So if recent requests consistently accept more drafted tokens, the policy tends to move up. If they start rejecting earlier, it tends to move down.

Usage

--speculative-adaptive-config is optional, but the speculative setup still needs to be valid for adaptive mode.

python3 -m sglang.launch_server \
    --model meta-llama/Llama-2-7b-chat-hf \
    --speculative-algorithm EAGLE \
    --speculative-draft-model-path lmsys/sglang-EAGLE-llama2-chat-7B \
    --speculative-eagle-topk 1 \
    --speculative-num-steps 3 \
    --speculative-num-draft-tokens 4 \
    --speculative-adaptive

If you want to override the defaults, add --speculative-adaptive-config /path/to/adaptive_spec.json.

Example config:

{
  "candidate_steps": [1, 3, 7],
  "ema_alpha": 0.2,
  "warmup_batches": 10,
  "update_interval": 5
}

Config file reference

The config file is optional. Any omitted keys use defaults.

The initial --speculative-num-steps is snapped to the nearest value in candidate_steps.

Monitoring

You can inspect the active tier and acceptance metric via /server_info:

curl -s http://127.0.0.1:30000/server_info | jq '.internal_states[0] | {speculative_num_steps, avg_spec_accept_length}'

• speculative_num_steps is the current active tier
• avg_spec_accept_length helps explain whether the server is likely to move up or down

Tuning tips

• Start with the default candidate tiers [1, 3, 7]
• Use fewer tiers if you want lower startup and graph-memory overhead
• Increase ema_alpha to react faster, or lower it for more stability
• Increase warmup_batches or update_interval if tier switching is too noisy
• If your workload is already stable and one static setting is well tuned, adaptive mode may not help much