server: Adding spec-decode counters to /metrics endpoint

[BLSharda]

Overview

Adds speculative decoding counters to the /metrics (Prometheus) endpoint, matching the vLLM counter schema:

• llamacpp:spec_decode_num_draft_tokens
• llamacpp:spec_decode_num_accepted_tokens
• llamacpp:spec_decode_num_drafts
• llamacpp:spec_decode_num_accepted_tokens_per_pos{position="N"} — per-draft-position labeled counter; absent before
  the first completed speculative request

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES, human code polished and reviewed by AI.

[BLSharda]

@gaugarg-nv , can you help review?

[gaugarg-nv]

Does this change work fine with tools like AIPerf those use \metrics endpoint?

[gaugarg-nv]

vLLM adds a _total suffix, I guess we should do the same: https://github.com/vllm-project/vllm/blob/62195e9784ebec1ece42b88a861734e0702cc2d5/vllm/v1/spec_decode/metrics.py#L177-L196

[BLSharda]

@ggerganov, I have filed an issue #26516 for this request, can you please take a look?

[ggerganov]

@ruixiang63 @ngxson PTAL too

[ruixiang63]

Should we add mean_accept_length into Prometheus metrics as well?

[BLSharda]

Should we add mean_accept_length into Prometheus metrics as well?

It can be derived from existing counters as 1 + num_accepted_tokens_total / num_drafts_total so there isn’t much value in tracking it separately. AIPerf already uses the same formula to compute it internally.