(self)
| 13516 | return max_batch_size is None or draft_batch_size <= max_batch_size |
| 13517 | |
| 13518 | def admit_waiting(self) -> None: |
| 13519 | while self.pending_requests: |
| 13520 | request = self.pending_requests[0] |
| 13521 | if request.cancelled: |
| 13522 | self.fail_request(request, CompletionRequestCancelledError("request cancelled")) |
| 13523 | continue |
| 13524 | if not self.can_admit(request): |
| 13525 | break |
| 13526 | self.pending_requests.popleft() |
| 13527 | self.admit_request(request) |
| 13528 | |
| 13529 | def can_admit(self, request: CompletionRequest) -> bool: |
| 13530 | return self.memory_policy.can_admit(request) |
no test coverage detected