[Core][5/N] Fully working chunked prefill e2e (#3884)

vllm/engine/llm_engine.py

@@ -633,7 +633,10 @@ class LLMEngine:
             seq_group = scheduled_seq_group.seq_group
             seq_group.update_num_computed_tokens(
                 scheduled_seq_group.token_chunk_size)
-            self._process_sequence_group_outputs(seq_group, outputs)
+            # If uncomputed tokens > 0, it means prefill is chunked.
+            # We don't need to process outputs in that case.
+            if seq_group.get_num_uncomputed_tokens() == 0:
+                self._process_sequence_group_outputs(seq_group, outputs)
 
         # Free the finished sequence groups.
         self.scheduler.free_finished_seq_groups()