fix(v1)：修复抢占恢复时 BlockTable 溢出

  - 缺失请求统一延后重新加入，避免同一轮重复写入
  - 抢占恢复用覆盖写入（add_row），正常运行用增量追加（append_row）
  - 保持正常请求的追加语义不变

vllm/v1/worker/gpu_model_runner.py

@@ -539,19 +539,16 @@ class GPUModelRunnerBase(LoRAModelRunnerMixin):
                 # The request is not in the persistent batch.
                 # The request was either preempted and resumed later, or was not
                 # scheduled in the previous step and needs to be added again.
-                if not is_last_rank:
-                    req_state = self.requests[req_id]
-                    self.input_batch.add_request(req_state)
-                    req_index = self.input_batch.req_id_to_index.get(req_id)
-                else:
-                    req_ids_to_add.append(req_id)
-                    continue
+                req_ids_to_add.append(req_id)
+                continue
 
             # Update the persistent batch.
             self.input_batch.num_computed_tokens_cpu[req_index] = (
                 num_computed_tokens)
-            self.input_batch.block_table.append_row(new_block_ids, req_index)
-
+            if resumed_from_preemption:
+                self.input_batch.block_table.add_row(new_block_ids, req_index)
+            else:
+                self.input_batch.block_table.append_row(new_block_ids, req_index)
             # For the last rank, we don't need to update the token_ids_cpu
             # because the sampled tokens are already cached.
             if not is_last_rank:
@@ -3699,4 +3696,4 @@ class GPUModelRunnerMTP(GPUModelRunnerBase):
 if envs.VLLM_USE_ZERO_MTP:
     GPUModelRunner=GPUModelRunnerMTP
 else:
-    GPUModelRunner=GPUModelRunnerBase
+    GPUModelRunner=GPUModelRunnerBase
\ No newline at end of file