[BugFix] Fix seeded random sampling with encoder-decoder models (vllm-project#8870)

Co-authored-by: Roger Wang <[email protected]>
Signed-off-by: Sumit Dubey <[email protected]>

Author: 2 people

vllm/worker/enc_dec_model_runner.py

@@ -268,11 +268,13 @@ def prepare_model_input(
             encoder_input_positions=encoder_input_positions_tensor,
         )
 
+        generators = self.get_generators(finished_requests_ids)
         sampling_metadata = SamplingMetadata.prepare(seq_group_metadata_list,
                                                      model_input.seq_lens,
                                                      model_input.query_lens,
                                                      self.device,
-                                                     self.pin_memory)
+                                                     self.pin_memory,
+                                                     generators=generators)
         is_prompt = (seq_group_metadata_list[0].is_prompt
                      if seq_group_metadata_list else None)
         return dataclasses.replace(model_input,