inference/research
verifiedhigh · 0.95updated 2026-10-04

obs-sglang-request-stage-enum

repo

Fields

id
sources/obs-sglang-request-stage-enum
Publisher
sgl-project/sglang
Kind
repo
Accessed
2026-10-04
notes
The RequestStage enum: the canonical stage_name strings used both as trace slice names and as the stage label on sglang:per_stage_req_latency_seconds. Verbatim: tokenize, api_server_dispatch, dpc_dispatch, request_process, prefill_waiting, decode_forward, decode_loop, prefill_forward, chunked_prefill, mm_encode, prefill_prepare, prefill_bootstrap, prefill_transfer_kv_cache, decode_prepare, decode_bootstrap, decode_waiting, decode_transferred, fake_output, quick_finish, spec_draft, spec_verify, run_batch_cpu. level 1/2/3 is trace nesting depth; metrics_is_observed marks which stages also feed the per-stage latency histogram.

Notes

The RequestStage enum: the canonical stage_name strings used both as trace slice names and as the stage label on sglang:per_stage_req_latency_seconds. Verbatim: tokenize, api_server_dispatch, dpc_dispatch, request_process, prefill_waiting, decode_forward, decode_loop, prefill_forward, chunked_prefill, mm_encode, prefill_prepare, prefill_bootstrap, prefill_transfer_kv_cache, decode_prepare, decode_bootstrap, decode_waiting, decode_transferred, fake_output, quick_finish, spec_draft, spec_verify, run_batch_cpu. level 1/2/3 is trace nesting depth; metrics_is_observed marks which stages also feed the per-stage latency histogram.

Cross-references

This record references no other records.