chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

[Bug] Fix missing TTFT histogram for single-batch requests (#20122)

Browse Source 
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
Liangsheng Yin
2026-03-08 15:18:51 -07:00
committed by GitHub
parent 7662b8b919
commit 7105bf3782
2 changed files with 9 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
python/sglang/srt/managers/tokenizer_manager.py
View File

@@ -134,6 +134,7 @@ class ReqState:
obj: Union[GenerateReqInput, EmbeddingReqInput]
time_stats: APIServerReqTimeStats
last_completion_tokens: int = 1
ttft_observed: bool = False
# For streaming output
last_output_offset: int = 0
@@ -1611,12 +1612,13 @@ class TokenizerManager(TokenizerCommunicatorMixin, TokenizerManagerMultiItemMixi
}
state.finished = recv_obj.finished_reasons[i] is not None
# Set first_token_time on the first output batch.
# This is the single write point for first_token_time.
if state.time_stats.first_token_time == 0.0:
state.time_stats.set_first_token_time()
if state.finished:
# Ensure first_token_time is set before computing decode_throughput.
# Without this, requests that finish on their first output batch
# would have first_token_time=0.0, producing bogus throughput.
if state.time_stats.first_token_time == 0.0:
state.time_stats.set_first_token_time()
state.time_stats.trace_ctx.trace_set_root_attrs(
self.convert_to_span_attrs(state, recv_obj, i)
)
@@ -1927,10 +1929,10 @@ class TokenizerManager(TokenizerCommunicatorMixin, TokenizerManagerMultiItemMixi
if priority is not None:
labels["priority"] = str(priority)
if (
state.time_stats.first_token_time == 0.0
not state.ttft_observed
and self.disaggregation_mode != DisaggregationMode.PREFILL
):
state.time_stats.set_first_token_time()
state.ttft_observed = True
state.last_completion_tokens = completion_tokens
self.metrics_collector.observe_time_to_first_token(
labels, state.time_stats.get_first_token_latency()