Add metrics to RequestOutput (#2876)

This commit is contained in:
Antoni Baum
2024-02-20 21:55:57 -08:00
committed by GitHub
parent 181b27d881
commit 017d9f1515
6 changed files with 61 additions and 9 deletions
+5 -2
View File
@@ -728,6 +728,7 @@ class LLMEngine:
def _process_model_outputs(
self, output: SamplerOutput,
scheduler_outputs: SchedulerOutputs) -> List[RequestOutput]:
now = time.time()
# Update the scheduled sequence groups with the model outputs.
scheduled_seq_groups = scheduler_outputs.scheduled_seq_groups
for seq_group, outputs in zip(scheduled_seq_groups, output):
@@ -739,6 +740,7 @@ class LLMEngine:
# Create the outputs.
request_outputs: List[RequestOutput] = []
for seq_group in scheduled_seq_groups:
seq_group.maybe_set_first_token_time(now)
request_output = RequestOutput.from_seq_group(seq_group)
request_outputs.append(request_output)
for seq_group in scheduler_outputs.ignored_seq_groups:
@@ -876,11 +878,12 @@ class LLMEngine:
# Latency Timings.
time_last_iters = []
for seq_group in scheduler_outputs.scheduled_seq_groups:
# Time since last token. (n.b. updates seq_group.last_token_time)
# Time since last token. (n.b. updates seq_group.metrics.last_token_time)
time_last_iters.append(seq_group.get_last_latency(now))
# Time since arrival for all finished requests.
if seq_group.is_finished():
time_e2e_requests.append(now - seq_group.arrival_time)
time_e2e_requests.append(now -
seq_group.metrics.arrival_time)
time_to_first_tokens = time_last_iters if prompt_run else []
time_per_output_tokens = [] if prompt_run else time_last_iters