[FIXBUG] Correctly Apply Grammar Bitmask in Mixed Batches (#22896)

Signed-off-by: JartX <sagformas@epdcenter.es>

vllm/v1/worker/gpu_model_runner.py

@@ -1337,9 +1337,10 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
         out_indices = []
 
         # Reorder the bitmask to match the order of the requests in the batch.
-        sorted_bitmask = np.zeros_like(grammar_bitmask,
-                                       shape=(logits.shape[0],
-                                              grammar_bitmask.shape[1]))
+        sorted_bitmask = np.full(shape=(logits.shape[0],
+                                        grammar_bitmask.shape[1]),
+                                 fill_value=-1,
+                                 dtype=grammar_bitmask.dtype)
         cumulative_index = 0
         seq = sorted(scheduler_output.structured_output_request_ids.items(),
                      key=lambda x: x[1])