Field DecodePathTag
Tag naming which decoder served a request. Two decoders exist and they own their KV cache differently: the classic controller keeps one context per conversation, the parallel engine leases a slot from a shared pool. Which one ran decides how a workload scales with concurrency, so it has to be readable from the metric stream rather than inferred by the caller from request shape.
public const string DecodePathTag = "lmkit.decode.path"Returns
- string
- Tag naming which decoder served a request. Two decoders exist and they own their KV cache differently: the classic controller keeps one context per conversation, the parallel engine leases a slot from a shared pool. Which one ran decides how a workload scales with concurrency, so it has to be readable from the metric stream rather than inferred by the caller from request shape.