fix: guard against KeyError with _req_ids2idx

Signed-off-by: Travis Johnson <[email protected]>

Author: tjohnson31415

vllm_spyre/worker/spyre_model_runner.py

@@ -284,7 +284,10 @@ def prepare_model_input(
             # updating indices: set indices of newly finished sequences False
             if finished_requests_ids:
                 for seq_id in finished_requests_ids:
-                    self.model.indices[self._req_ids2idx[seq_id]] = False
+                    # ignore requests that are not in the batch, eg. requests
+                    # cancelled while waiting
+                    if idx := self._req_ids2idx.get(seq_id):
+                        self.model.indices[idx] = False
             (input_tokens, input_positions,
              input_masks) = self._prepare_decode(seq_group_metadata_list)
             seq_lens = []