Drain inflight requests after early stop
This commit is contained in:
@@ -135,6 +135,7 @@ def _replay_requests(
|
||||
max_lag_s: float | None,
|
||||
max_elapsed_s: float | None,
|
||||
evaluate_outcome: Callable[[RequestOutcome], Any],
|
||||
drain_inflight_on_early_stop: bool = True,
|
||||
) -> tuple[list[RequestOutcome], bool, str]:
|
||||
outcomes_by_id: dict[str, RequestOutcome] = {}
|
||||
lock = threading.Lock()
|
||||
@@ -209,18 +210,47 @@ def _replay_requests(
|
||||
if sleep_for > 0:
|
||||
time.sleep(min(sleep_for, 0.1))
|
||||
if early_stopped:
|
||||
for future in list(futures_by_request):
|
||||
future.cancel()
|
||||
for request in futures_by_request.values():
|
||||
outcomes_by_id[request.row_id] = RequestOutcome(
|
||||
request_id=request.row_id,
|
||||
success=False,
|
||||
ttft_ms=None,
|
||||
tpot_ms=None,
|
||||
prompt_tokens=request.prompt_tokens_hint,
|
||||
completion_tokens=request.completion_tokens_hint,
|
||||
error=early_stop_reason or "probe_early_stopped",
|
||||
)
|
||||
if drain_inflight_on_early_stop and futures_by_request:
|
||||
done, not_done = wait(list(futures_by_request), timeout=timeout_s)
|
||||
for future in done:
|
||||
request = futures_by_request[future]
|
||||
try:
|
||||
outcomes_by_id[request.row_id] = future.result(timeout=0)
|
||||
except Exception: # noqa: BLE001
|
||||
outcomes_by_id[request.row_id] = RequestOutcome(
|
||||
request_id=request.row_id,
|
||||
success=False,
|
||||
ttft_ms=None,
|
||||
tpot_ms=None,
|
||||
prompt_tokens=request.prompt_tokens_hint,
|
||||
completion_tokens=request.completion_tokens_hint,
|
||||
error=early_stop_reason or "probe_early_stopped",
|
||||
)
|
||||
for future in not_done:
|
||||
future.cancel()
|
||||
request = futures_by_request[future]
|
||||
outcomes_by_id[request.row_id] = RequestOutcome(
|
||||
request_id=request.row_id,
|
||||
success=False,
|
||||
ttft_ms=None,
|
||||
tpot_ms=None,
|
||||
prompt_tokens=request.prompt_tokens_hint,
|
||||
completion_tokens=request.completion_tokens_hint,
|
||||
error=early_stop_reason or "probe_early_stopped",
|
||||
)
|
||||
else:
|
||||
for future in list(futures_by_request):
|
||||
future.cancel()
|
||||
for request in futures_by_request.values():
|
||||
outcomes_by_id[request.row_id] = RequestOutcome(
|
||||
request_id=request.row_id,
|
||||
success=False,
|
||||
ttft_ms=None,
|
||||
tpot_ms=None,
|
||||
prompt_tokens=request.prompt_tokens_hint,
|
||||
completion_tokens=request.completion_tokens_hint,
|
||||
error=early_stop_reason or "probe_early_stopped",
|
||||
)
|
||||
for request in requests:
|
||||
if request.row_id in submitted_ids:
|
||||
continue
|
||||
|
||||
Reference in New Issue
Block a user