vllm:spec_decode_* counters stay 0 on the stepped bridge (DFlash/DSpark)

[xiaguan]

Symptom

vllm:spec_decode_num_drafts_total / num_draft_tokens_total / num_accepted_tokens_total stay at 0 in production even when DFlash/DSpark speculative decoding is actively running (visible per step in the debug log as Qwen3 DFlash request=... accepted_draft=N).

Root cause

#787 wires SpecDecodeCounters from Qwen3Executor::execute_speculative_verify_impl into LoadSnapshot.spec_decode, and the legacy handle bridge forwards per-interval deltas via publish_scheduler_stats (pegainfer-frontend/src/vllm/bridge.rs). But every served model line launches as LaunchedEngine::Stepped (e.g. qwen3 at pegainfer-qwen3/src/model_line.rs), and SteppedEngineBridge:

• passes load_watch = None, so publish_scheduler_stats never spawns, and
• stamps stats with a bare scheduler_stats_from(load) per output batch (bridge/stepped.rs, two sites), where ..SchedulerStats::default() leaves spec_decoding_stats = None forever.

The sim frontend e2e covers only the handle path, so the gap has no test net.

Why it matters

Upstream vllm-bench (now maintained in vllm-project/vllm at rust/src/bench) scrapes the server's /metrics before and after a run and auto-reports a Speculative Decoding section (acceptance rate/length, per-position rates) plus spec_decode_* JSON fields — no flags needed. With the counters stuck at 0 that whole surface is unusable for spec-decode A/B work against pegainfer.

Suggested fix

Give SteppedEngineBridge its own last_spec: SpecDecodeCounters state and diff the cumulative snapshot with the same spec_decode_delta the legacy publisher uses, attaching spec_decoding_stats only on intervals that drafted (delta.num_drafts > 0). Minimal sketch:

// in SteppedEngineBridge, one fn used by both stats-stamping sites
let snapshot = self.scheduler.load();
let mut stats = scheduler_stats_from(&snapshot);
if let Some(cur) = &snapshot.spec_decode {
    let delta = spec_decode_delta(&self.last_spec, cur);
    self.last_spec = *cur;
    stats.spec_decoding_stats = (delta.num_drafts > 0).then_some(delta);
}

Local verification of the patch

Qwen3-4B + dspark_qwen3_4b_block7, vllm-bench on openai-chat, c4, 20 prompts each:

dataset	acceptance rate	acceptance length
random	14.2%	1.97
sonnet	28.9%	2.99
sharegpt	20.5%	2.40

All in the documented DFlash/DSpark ranges (see docs/models/qwen3/dspark-integration.md); result JSON carries all spec_decode_* fields.

Follow-up edge case (from review)

With the state held in the bridge, a step whose request updates are all dropped (e.g. an abort arriving while a speculative step is in flight — dispatch_step returns at outputs.is_empty()) never stamps stats, so the last cumulative increment could stay unpublished while the engine idles. The final version should also emit/refresh stats for steps with no surviving request output, or snapshot on every received step.

[scatyf3]

can you assign this to me, I'll fix this issue with in this week