vLLM spec-decode accepted-token counter
vllm · counter
Fields
id
metric-exposures/w3m-vllm-vllm-spec-decode-num-accepted-tokensMetric name
vllm:spec_decode_num_accepted_tokens
Metric type
counter
Unit
tokens
Exposure
prometheus
Endpoint
:8000/metrics
Labels
model_nameengine
Enabled by default
yes
notes
vllm/v1/spec_decode/metrics.py, SpecDecodingProm, main, read 2026-10-04. Only constructed when speculative_config is not None (or the model is a diffusion model), so absent otherwise by design. rate(accepted)/rate(drafts) is the documented acceptance derivation.
Notes
vllm/v1/spec_decode/metrics.py, SpecDecodingProm, main, read 2026-10-04. Only constructed when speculative_config is not None (or the model is a diffusion model), so absent otherwise by design. rate(accepted)/rate(drafts) is the documented acceptance derivation.